I've just finished major update to installer/launcher as well as some environment auto-tunes.
Hopefully now those steps should not be needed and should work out-of-the-box

Check README for updated notes and updated thread #92 with details.

Update for 04/14/2023:

• handle duplicate extensions
• redo exception handler
• fix generate forever
• enable cmdflags compatibility
• change default css font
• fix ti previews on initial start
• enhance tracebacks
• pin transformers version to last known good version
• fix extension loader

Update for 04/15/2023:

• update default values
• remove ui-config.json from repo, its not fully under user control
• updated extensions mangager
• updated locon/lycoris plugin
• enable quick launch by default
• add multidiffusion upscaler extensions
• add model keyword extension
• enable strong linting
• fix circular imports
• fix extensions updated
• fix git update issues
• update github templates

Update for 04/16/2023:

• support for ui themes! to to settings -> user interface -> "ui theme*
  includes 12 predefined themes
• ability to restart server from ui
• updated requirements
• removed styles.csv from repo, its now fully under user control
• removed model-keyword extension as overly aggresive
• rewrite of the fastapi middleware handlers
• install bugfixes, hopefully new installer is now ok
  i really want to focus on features and not troubleshooting installer

Update for 04/17/2023:

• themes are now dynamic and discovered from list of available gradio themes on huggingface
  its quite a list of 30+ supported themes so far
• added option to see theme preview without the need to apply it or restart server
• integrated image info functionality into process image tab and removed separate image info tab
• more installer improvements
• fix urls
• updated github integration
• make model download as optional if no models found

Update for 04/18/2023:

• reconnect ui to active session on browser restart
  this is one of most frequently asked for items, finally figured it out
  works for text and image generation, but not for process as there is no progress bar reported there to start with
• force unload xformers when not used
  improves compatibility with AMD/M1 platforms
• add styles.csv to UI settings to allow customizing path
• add --skip-git to cmd flags for power users that want
  to skip all git checks and operations and perform manual updates
• add --disable-queue to cmd flags that disables Gradio queues (experimental)
  this forces it to use HTTP instead of WebSockets and can help on unreliable network connections
• set scripts & extensions loading priority and allow custom priorities
  fixes random extension issues:
  ScuNet upscaler dissapearing, Additional Networks not showing up on XYZ axis, etc.
• improve html loading order
• remove some asserts causing runtime errors and replace with user-friendly messages
• update README.md
• update TODO.md

Update for 04/19/2023:

• fix live preview
• fix model merge
• fix handling of user-defined temp folders
• fix submit benchmark
• option to override torch and xformers installer
• separate benchmark data for system-info extension
• minor css fixes
• created initial merge backlog from pending prs on a1111 repo
  see Pending PRs #258 for details

Update for 04/20/2023:

• full CUDA tuning section in UI Settings
• improve exif/pnginfo metadata parsing
  it can now handle 3rd party images or images edited in external software
• optimized setup performance and logging
• improve compatibility with some 3rd party extensions
  for example handle extensions that install packages directly from github urls
• fix initial model download if no models found
• fix vae not found issues
• fix multiple git issues

note: if you previously had command line optimizations such as --no-half, those are now ignored and moved to ui settings

Update for 04/22/2023:

• autodetect which system libs should be installed
  this is a first pass of autoconfig for nVidia vs AMD environments
• fix parse cmd line args from extensions
• only install xformers if actually selected as desired cross-attention method
• do not attempt to use xformers or sdp if running on cpu
• merge tomesd token merging
• merge 23 PRs pending from a1111 backlog (!!)

expect shorter updates for the next few days as i'll be partically ooo

Update for 04/23/2023:

• fix VAE dtype
  should fix most issues with NaN or black images
• add built-in Gradio themes
• reduce requirements
• more AMD specific work
• initial work on Apple platform support
• additional PR merges
• handle torch cuda crashing in setup
• fix setup race conditions
• fix ui lightbox
• mark tensorflow as optional
• add additional image name templates

Update for 04/24/2023:

• identify race condition where generate locks up while fetching preview
• add pulldowns to x/y/z script
• add VAE rollback feature in case of NaNs
• use samples format for live preview
• add token merging
• use Approx NN for live preview
• create default styles.csv
• fix setup not installing tensorflow dependencies
• update default git flags to reduce number of warnings

Update for 04/25/2023:

• update process image -> info
• add VAE info to metadata
• update GPU utility search paths for better GPU type detection
• update git flags for wider compatibility
• update environment tuning
• update ti training defaults
• update VAE search paths
• add compatibility opts for some old extensions
• validate script args for always-on scripts
  fixes: deforum with controlnet

Update for 04/27/2023:

a bit shorter list as:

1. i've been busy with buxfixing
   there are a lot of them, not going to list each here.
   but seems like critical issues backlog is quieting down and soon i can focus on new features development.
2. i've started collaboration with couple of major projects,
   hopefully this will accelerate future development.

• ability to view/add/edit model description shown in extra networks cards
• add option to specify fallback sampler if primary sampler is not compatible with desired operation
• make clip skip a local parameter
• remove obsolete items from UI settings
• set defaults for AMD ROCm
  if you have issues, you may want to start with a fresh install so configuration can be created from scratch
• set defaults for Apple M1/M2
  if you have issues, you may want to start with a fresh install so configuration can be created from scratch

Update for 05/01/2023

Been a bit quieter for last few days as changes were quite significant, but finally here we are...

• Updated core libraries: Gradio, Diffusers, Transformers
• Added support for Intel ARC GPUs via Intel OneAPI IPEX (auto-detected)
• Added support for TorchML (set by default when running on non-compatible GPU or on CPU)
• Enhanced support for AMD GPUs with ROCm
• Enhanced support for Apple M1/M2
• Redesigned command params: run webui --help for details
• Redesigned API and script processing
• Experimental support for multiple Torch compile options
• Improved sampler support
• Google Colab: https://colab.research.google.com/drive/126cDNwHfifCyUpCCQF9IHpEdiXRfHrLN
  Maintained by https://github.com/Linaqruf/sd-notebook-collection
• Fixes, fixes, fixes...

To take advantage of new out-of-the-box tunings, its recommended to delete your config.json so new defaults are applied. Its not necessary, but otherwise you may need to play with UI Settings to get the best of Intel ARC, TorchML, ROCm or Apple M1/M2.

Update for 2025-02-20

Highlight for 2025-02-20

We're back with another update with nearly 100 commits!

• Starting with massive UI update with full localization for 8 languages
  and 100+ new hints
• Big update to Docker containers
  with support for all major compute platforms
• A lot of outpainting goodies
• Support for new models: AlphaVLLM Lumina 2 and Ostris Flex.1-Alpha
• And new Mixture-of-Diffusers regional prompting & tiling pipeline
• Follow-up to last weeks interrogate/captioning rewrite
  now with redesigned captioning UI, batch support, and much more
  plus JoyTag, JoyCaption, PaliGemma, ToriiGate, Ovis2 added to list of supported models
• Some changes to prompt parsing to allow more control as well as
  more flexibility when mouting SDNext server to custom URL
• Of course, cumulative fixes...

...and more - see changelog for full details!

Details for 2025-02-20

• User Interface
  • Hints
    • added/updated 100+ ui hints!
    • hints documentation and contribution guide
  • Localization
    • full ui localization!
      english, croatian, spanish, french, italian, portuguese, chinese, japanese, korean, russian
    • set in settings -> user interface -> language
    • localization documentation
  • UI
    • force browser cache-invalidate on page load
    • configurable request timeout
    • modernui improve gallery styling
    • modernui improve networks styling
    • modernui support variable card size
• Docs
  • New Outpaint step-by-step guide
  • Updated Docker guide
    includes build and publish and both local and cloud examples
• Models
  • AlphaVLLM Lumina 2
    new foundation model for image generation based o Gemma-2-2B text encoder and a flow-based diffusion transformer
    fully supports offloading and on-the-fly quantization
    simply select from networks -> models -> reference
  • Ostris Flex.1-Alpha
    originally based on Flux.1-Schnell, but retrained and with different architecture
    result is model smaller than Flux.1-Dev, but with similar capabilities
    fully supports offloading and on-the-fly quantization
    simply select from networks -> models -> reference
• Functions
  • Mixture-of-Diffusers
    Regional tiling type of a solution for SDXL models
    select from scripts -> mixture of diffusers
  • [Automatic Color Inpaint]
    Automatically creates mask based on selected color and triggers inpaint
    simply select in scripts -> automatic color inpaint when in img2img mode
  • RAS: Region-Adaptive Sampling experimental
    Speeds up SD3.5 models by sampling only regions of interest
    Enable in settings -> pipeline modifiers -> ras
• Interrogate/Captioning
  • Redesigned captioning UI
    split from Process tab into separate tab
    split clip vs vlm models processing
    direct send-to buttons on all tabs: txt/img/ctrl->process/caption, process/caption->txt/img/ctrl
  • Advanced params:
    VLM: max-tokens, num-beams, temperature, top-k, top-p, do-sample
    CLiP: min-length, max-length, chunk-size, min-flavors, max-flavors, flavor-count, num-beams
    params are auto-saved in config.json and used when using quick interrogate
    params that are set to 0 mean use model defaults
  • Batch processing: VLM and CLiP
    for example, can be used to caption your training dataset in one go
    add option to append to captions file, can be used to run multiple captioning models in sequence
    add option to run recursively on all subfolders
    add progress bar
  • Add additional VLM models:
    JoyTag
    JoyCaption 2
    Google PaliGemma 2 3B
    ToriiGate 0.4 7B
    AIDC Ovis2 1B/2B/4B
  • Note some models require flash-attn to be installed
    due to binary/build dependencies, it should not be done automatically,
    see flash-attn for installation instructions
• Docker
  • updated CUDA receipe to torch==2.6.0 with cuda==12.6 and add prebuilt image
  • added ROCm receipe and prebuilt image
  • added IPEX receipe and add prebuilt image
  • added OpenVINO receipe and prebuilt image
• System
  • improve python==3.12 compatibility
  • Torch
    • for zluda set default to torch==2.6.0+cu118
    • for openvino set default to torch==2.6.0+cpu
  • OpenVINO
    • update to openvino==2025.0.0
    • improve upscaler compatibility
    • enable upscaler compile by default
    • fix shape mismatch errors on too many resolution changes
  • ZLUDA
    • update to zluda==3.8.8
• Other
  • Asymmetric tiling
    allows for configurable image tiling for x/y axis separately
    enable in scripts -> asymmetric tiling
    note: traditional symmetric tiling is achieved by setting circular mode for both x and y
  • Styles
    ability to save and/or restore prompts before or after parsing of wildcards
    set in settings -> networks -> styles
  • Access tokens
    persist models -> hugginface -> token
    persist models -> civitai -> token
  • global switch to lancosz method for all interal resize ops and bicubic for interpolation ops
  • Text encoder
    add advanced per-model options for text encoder
    set in settings -> text encoder -> Optional
  • Subpath
    allow setting additional mount subpath over which server url will be accessible
    set in settings -> user interface
  • Prompt parsing
    better handling of prompt parsing when using masking char \
• Fixes
  • update torch nightly urls
  • docs/wiki always use relative links
  • ui use correct timezone for log display
  • ui improve settings search behavior
  • ui log scroll to bottom
  • ui fix send to inpaint/sketch
  • modernui add control init image toggle
  • modernui fix sampler advanced options
  • outpaint fixes
  • validate output before hires/refine
  • scheduler fix sigma index out of bounds
  • force pydantic version reinstall/reload
  • multi-unit when using controlnet-union
  • pulid with hidiffusion
  • api: stricter access control
  • api: universal handle mount subpaths

SD.Next Release 2025-02-24

Primarily a hotfix/service release plus few UI improvements and one exciting new feature: Remote-VAE!

• Remote Decode
  • final step of image generate, VAE decode, is by far the most memory intensive operation and can easily result in out-of-memory errors
    what can be done? Well, Huggingface is now providing free-of-charge remote-VAE-decode service!
  • how to use? previous Full quality option in UI is replaced with VAE type selector: Full, Tiny, Remote
    currently supports SD15, SDXL and FLUX.1 with more models expected in the near future
    depending on your bandwidth select mode in settings -> vae -> raw/png/jpg
    if remote processing fails SD.Next will fallback to using normal VAE decode process
    privacy note: only passed item is final latent itself without any user or generate information and latent is not stored in the cloud
• UI
  • modern ui reorg main tab
    improve styling, improve scripts/extensions interface and separate ipadapters
  • additional ui hints
• Other
  • add --extensions-dir cli arg and SD_EXTENSIONSDIR env variable to specify extensions directory
  • update zluda==3.9.0
• Fixes
  • skip trying to register legacy/incompatibile extensions in control ui
  • add additional scripts/extensions callbacks
  • remove ui splash screen on auth fail
  • log full config path, full log path, system name, extensions path
  • zluda hotfixes
  • zluda force sync
  • fix torch import on compile
  • infotext parser force delimiter before params
  • handle pipeline class switch errors

SD.Next Release 2025-04-03

Time for another major release with ~120 commits and ChangeLog that spans several pages!

Highlights?
Video...Brand new Video processing module with support for all latest models: WAN21, Hunyuan, LTX, Cog, Allegro, Mochi1, Latte1 in both T2V and I2V workflows
And combined with on-the-fly quantization, support for Local/Tiny/Remote VAE, acceleration modules such as FasterCache or PAB, and more!
Models...And support for new models: CogView-4, SANA 1.5,

Plus...

• New Prompt Enhance using LLM,
• New pipelines such as InfiniteYou
• New CLiP models, improvements to remote VAE, additional wiki/docs/guides
• More quantization options and granular control
• Pretty big performance updates to a) Any model using DiT based architecture due to new caching methods, b) ZLUDA with new attention methods, c) LoRA with much lower memory usage

ReadMe | ChangeLog | Docs | WiKi | Discord

Details

• Video tab
  • see Video Wiki for details!
  • new top-level tab, replaces previous video script in text/image tabs
    old scripts are still present, but will be removed in the future
  • support for all latest models:
    • Hunyuan: HunyuanVideo, FastHunyuan, SkyReels | T2V, I2V
    • WAN21: 1.3B, 14B | T2V, I2V
    • LTXVideo: 0.9.0, 0.9.1, 0.9.5 | T2V, I2V
    • CogVideoX: 2B, 5B | T2V, I2V
    • Allegro: T2V
    • Mochi1: T2V
    • Latte1: *T2V
  • decoding:
    • Default: use vae from model
    • Tiny VAE: support for Hunyuan, WAN, Mochi
    • Remote VAE: support for Hunyuan
  • LoRA
    • support for Hunyuan, LTX, WAN, Mochi, Cog
    • add option to apply LoRA directly on GPU or use CPU first in low-memory scenarios
    • improve metadata and preview parallel fetch
    • support for mp4 so first frame is extracted as used as lora preview
  • additional key points:
    • all models are auto-downloaded upon first use
      uses system paths -> huggingface folder
    • support for many video types
    • optional video interpolation while creating video files
    • optional video preview in ui
      present if video output is selected
    • support for balanced offloading and model offloading
      uses system settings
    • on-the-fly quantization: BnB, Quanto, TorchAO
      uses system settings, granular for transformer and text-encoder separately
    • different video models support different video resolutions, frame counts, etc.
      and may require specific settings - see model links for details
    • see ToDo/Limitations section for additional notes
• Models & Pipelines
  • THUDM CogView 4 6B variant
    new foundation model for image generation based o GLM-4 text encoder and a flow-based diffusion transformer
    fully supports offloading and on-the-fly quantization
    simply select from networks -> models -> reference
    note cogview4 is compatible with flowmatching samplers
  • NVLabs SANA 1.5 in 1.6B, 4.8B and Sprint variations
    big update to previous SANA model
    fully supports offloading and on-the-fly quantization
    simply select from networks -> models -> reference
  • ByteDance InfiniteYou: Flexible Photo Recrafting While Preserving Your Identity
    face-transfer model for FLUX.1
    select from Scripts -> InfiniteYou
    its large, ~12GB on top of FLUX.1 base model so make sure you have offloading and quantization setup
    note model will be auto-downloaded on first use
  • New zer0int CLiP-L models:
    download text encoders into folder set in settings -> system paths -> text encoders (default is models/Text-encoder)
    load using settings -> text encoder
    tip: add sd_text_encoder to your settings -> user interface -> quicksettings list to have it appear at the top of the ui
• Prompt Enhance
  • see Prompt Enhance Wiki for details!
  • new built-in extension available in text/image/control tabs
  • can be used to manually or automatically enhance prompts using LLM
  • built-in presets for Gemma-3, Qwen-2.5, Phi-4, Llama-3.2, SmolLM2, Dolphin-3
  • support for custom models
    load any models hosted on huggingface
    load either model in huggingface format or gguf format
    note: any hf model in transformers.AutoModelForCausalLM standard should work
    note: not all model architecture are supported for gguf format
  • models are auto-downloaded on first use
  • support quantization and offloading
  • auto-detect censored output
  • debug using SD_LLM_DEBUG=true env variable
• Acceleration
  • Support for most DiT-based models, for example: FLUX.1, SD35, Hunyuan, Mochi, Latte, Allegro, Cog
  • Enable and configure in Settings -> Pipeline modifiers
  • FasterCache
  • PyramidAttentionBroadcast
• Remote VAE
  • add support for remote vae encode in addition to remote vae decode
  • used by img2img, inpaint, hires, detailer
  • remote vae encode is disabled by default, you can enable it in settings -> variable auto-encoder
  • add remote vae info to metadata, thanks @iDeNoh
  • remote vae use scaling_factor and shift_factor
• Caption/VLM
  • Google Gemma 3 4B
    simply select from list of available models in caption tab
  • ByteDance/Sa2VA 1B, 4B
    simply select from list of available models in caption tab
  • add option to set system prompt for vlm models that support it: Gemma, Smol, Qwen
• NudeNet extension updates
  • add detection of prompt language and alphabet and filter based on those values
  • add image policy checks using LlavaGuard VLM to detect policy violations (and reasons)
    against top-10 standard harmful content categories
  • add banned words/expressions check against prompt variations
• LoRA
  • enable memory cache by default
  • significantly reduce memory usage
  • improve performance
  • improve detection of lora changes
  • unload lora only when changes are detected
  • refactor code for modularity
• IPEX
  • add --upgrade to torch_command when using --use-nightly
  • add xpu to profiler
  • fix untyped_storage, torch.eye and torch.cuda.device ops
  • fix torch 2.7 compatibility
  • fix performance with balanced offload
  • fix triton and torch.compile
• ROCm
  • add --upgrade to torch_command when using --use-nightly
  • disable fp16 for gfx1102 (rx 7600 and rx 7500 series) gpus
• ZLUDA
  • triton for ZLUDA v3.9.2
    • torch.compile is now available
    • Flash Attention 2 is now available
• Other
  • new command line option --monitor PERIOD to monitor CPU and GPU memory ever n seconds
  • upscale: new asymmetric vae v2 upscaling method
  • upscale: new experimental support for libvips upscaling
  • quantization: add support for optimum-quanto on-the-fly quantization during load for all models
    note: previous method for quanto is still valid and is noted in settings as post-load quantization
  • add quantization support to CogView-3Plus
  • update diffusers and other requirements
  • rename vae, unet and text-encoder settings None to Default to avoid confusion
  • CLI: add cli/api-grid.py which can generate grids using params-from-file for x/y axis
  • Samplers add ability to set sigma adjustment for each sampler
  • ModernUI updates
  • CSS updates
  • settings vertiocal/dirty indicator restores to default setting instead to previous value
  • video interpolate do not skip duplicate frames
  • settings UI full refactor
• Wiki/Docs
  • updated Models info
  • new Video guide
  • new Caption guide
  • new VAE guide
  • updated SD3 guide
  • updated ZLUDA guide
  • updated OpenVINO guide
  • updated AMD-ROCm guide
  • updated Intel-ARC guide
• Fixes
  • fix installer not starting when older version of rich is installed
  • fix circular imports when debug flags are enabled
  • fix cuda errors with directml
  • fix memory stats not displaying the ram usage
  • fix RunPod memory limit reporting
  • fix flux ipadapter with start/stop values
  • fix progress api eta_relative
  • fix insightface loader
  • fix remove vae for flux.1
  • guard against git returining invalid timestamp
  • fix hires with latent upscale
  • fix legacy diffusion latent upscalers
  • fix upscaler selection in postprocessing
  • fix sd35 with batch processing
  • fix extra networks cover and inline views
  • fix token counter error style with modernui
  • fix sampler metadata when using default sampler
  • fix paste incorrect float to int cast
  • fix server restart from ui
  • fix style apply params
  • do not allow edit of built-in styles
  • improve lora compatibility with balanced offload

SD.Next Release 2025-04-12

Highlights for 2025-04-12

Last release was just over a week ago and here we are again with another update as a new high-end image model, HiDream-I1 jumped out and generated a lot of buzz!
There are quite a few other performance and quality-of-life improvements in this release and 40 commits, so please take a look at the full ChangeLog

ReadMe | ChangeLog | Docs | WiKi | Discord

Details for 2025-04-12

• Models
  • HiDream-I1 in fast, dev and full variants!
    new absolutely massive image generative foundation model with 17B parameters and 4 text-encoders with additional 8.3B parameters
    simply select from networks -> models -> reference
    due to size (over 25B params in 58GB), offloading and on-the-fly quantization are pretty much a necessity
    see HiDream Wiki page for details
• Features
  • Custom model loader
    can be used to load any known diffusion model with default or custom model components
    in models -> custom tab
    see docs for details: https://vladmandic.github.io/sdnext-docs/Loader/
    • Pipe: SoftFill
• Caching
  • add TeaCache support to Flux, CogVideoX, Mochi, LTX
  • add FasterCache support to WanAI, LTX (other video models already supported)
  • add PyramidAttentionBroadcast support to WanAI, LTX (other video models already supported)
• UI
  • client polling speeds up and slows down depending if client page is visible or not
    client polling does not ask for live preview if page is not visible
    significantly reduces server load if you hide or minimize the page
  • progress: use batch-count for progress
  • grid: add of max-rows and max-columns in settings to control grid format
  • gallery: add max-columns in settings for gradio gallery components
• Other
  • ZLUDA: add more GPUs to recognized list
    select in scripts, available for sdxl in inpaint model
  • LoRA: add option to force-reload LoRA on every generate
  • settings: add Model options sections as placeholder for per-model settings
  • video: update LTXVideo-0.9.5 pipeline
  • te loader: allow free-form input in which case sdnext will attempt to load it as hf repo
  • diag: add get-server-status to UI generate context menu
  • diag: memory monitor detect gpu swapping
  • use hf-xet for huggingface downloads where possible
  • quant: update & fix optimum-quanto for transformers
  • quant: update & fix torchao
  • model load: new setting for model load initial device map
    can be used to force gpu vs cpu when loading model to avoid oom before model offloading is even activated after load
• Changes
  • params: Reset default guidance-rescale from 0.7 to 0.0
  • progress: add additional fields to progress API
• Fixes
  • styles: resize and bring quick-ui to forward on hover
  • LoRA: obey configured device when performing calculations
  • ZLUDA: startup issues
  • offload: balanced offload remove non-blocking move op
  • logging: debug causes invalid import
  • logging: cleanup
  • ROCm: flash attention repo with navi rotary fix
  • prompt: prompt scheduling with te caching
  • ui: progress allow for longer timeouts
  • internal: cleanup defined pipelines

SD.Next Release 2025-04-28

Another major release with over 120 commits!
Highlights include new Nunchaku Wiki inference engine that allows running FLUX.1 with 3-5x higher performance!
And a new FramePack extension for high-quality I2V and FLF2V video generation with unlimited duration!

What else?

• New UI History tab
• New models: Flex.2, LTXVideo-0.9.6, WAN-2.1-14B-FLF2V, schedulers: UniPC and LCM FlowMatch, features: CFGZero
• Major updates to: NNCF, OpenVINO, ROCm, ZLUDA
• Cumulative fixes since last release

ReadMe | ChangeLog | Docs | WiKi | Discord

Details for 2025-04-28

• Features
  • Nunchaku inference engine with custom SVDQuant 4-bit execution
    highly experimental and with limited support, but when it works, its magic: Flux.1 at 6.0 it/s (not sec/it)!
    basically, it can speed up supported models by 2-5x by using custom quantization and execution engine
    see Nunchaku Wiki for installation guide and list of supported models & features
  • FramePack based on HunyuanVideo-I2V
    full support and much more for Lllyasviel FramePack
    implemented as an extension for SD.Next (for the moment while dev is ongoing)
    generate high-quality videos with pretty much unlimited duration and with limited VRAM!
    install as any other extension and for details see extension README
    • I2V & FLF2V support with explicit strength controls
    • complex actions: modify prompts for each section of the video
    • LoRA support: use normal HunyuanVideo LoRAs
    • decode: use local, tiny or remote VAE
    • custom models: e.g. replace llama with one of your choice
    • video: multiple codecs and with hw acceleration, raw export, frame export, frame interpolation
    • compute: quantization support, new offloading, more configuration options, cross-platform, etc.
  • Ostris Flex.2 Preview
    more than a FLUX.1 finetune, FLEX.2 is created from Flux.1 Schnell -> OpenFlux.1 -> Flex.1-alpha -> Flex.2-preview
    and it has universal control and inpainting support built in!
    supported for text and control workflows
    when using in control mode, simply choose preprocessor and do not load actual controlnet
    supported control modes are: line, pose and depth
    available via networks -> models -> reference
  • LTXVideo 0.9.6 T2V and I2V
    in both Standard and Distilled variants
    available in video tab
  • WAN 2.1 14B 720P FLF2V
    new first-to-last image video model from WAN-AI
    available in video tab
  • CFG-Zero new guidance method optimized for flow-matching models
    implemented for FLUX.1, HiDream-I1, SD3.x, CogView4, HunyuanVideo, WanAI
    enable and configure in settings -> pipeline modifiers -> cfg zero
    experiment with CFGZero support in XYZ-grid
• Optimizations
  • NNCF update to 2.16.0
    major refactoring of NNCF quantization code
    new quant types: INT8_SYM (new default), INT4 and INT4_SYM
    quantization support for the convolutional layers on unet models with sym methods
    pre-load quantization support
    LoRA support
    if you're low on VRAM, NNCF is as close as a catch-all solution
  • OpenVINO update to 2025.1.0 and Torch to 2.7
  • ROCm update to Torch 2.7
  • HiDream-I1 optimized offloading and prompt-encode caching
    it now works in 12GB VRAM / 26GB RAM!
  • CogView3 and CogView4 model loader optimizations
  • Sana model loader optimizations
  • add explicit offload after encode prompt
    configure in settings -> text encoder -> offload
• UI
  • new History tab where you can see all jobs since the server startup
    and optionally download any of the previously generated images/videos
    access via system -> history
  • server restart from ui now replaces currently running process
    instead of trying to reload python modules in-place
  • add option to enable/disable clip skip
    disabled by default to avoid issues with frequent incorrect recommendations
    in settings -> pipeline modifiers
  • configurable restore metadata from image to settings and to params
    in settings -> image metadata
• API
  • new API Wiki
  • server will now maintain job history which can be queried via API
    so you can check previous jobs as well as request any previously generated images/videos
  • history endpoint: /sdapi/v1/history?id={id}
  • download endpoint: /file={filename}
  • progress api /sdapi/v1/progress now also include task id in the response
• Other
  • OMI support for sd15/sdxl omi-standard LoRAs
  • text/image/control/video pipeline vs task compatibility check
  • HiDream-I1, FLUX.1, SD3.x add HF gated access auth check
  • HiDream-I1 LoRA support
    currently limited to diffusers-only LoRAs, CivitAI LoRA support is TBD
  • HiDream-I1 add LLM info to image metadata
  • add model_type as option for image filename pattern
  • add UniPC FlowMatch scheduler
  • add LCM FlowMatch scheduler
  • networks: set which networks to skip when scanning civitai
    in settings -> networks -> network scan
    comma-separate list of regex patterns to skip
  • ui display reference models with subdued color
  • xyz grid support bool
  • do not force gc at end of processing
  • add SD_LORA_DUMP env variable for dev/diag to dump lora/model keys
• Wiki
  • new Nunchaku, API pages
  • updated HiDream, Quantization, NNCF, Video, Docker, WSL, ZLUDA pages
• Fixes
  • HunyuanVideo-I2V with latest transformers
  • NNCF with TE-only quant
  • ONNX init fix
  • Quanto with TE/LLM quant
  • HiDream live preview
  • FLUX.1 controlnet i2i
  • SD35 InstantX IP-adapter
  • OpenVINO device selection
  • xyz grid restore settings
  • config save unnecessary keys
  • recursive wildcards
  • extension installer handling of PYTHONPATH
  • trace logging
  • api logging
  • sd/sdxl-inpaint model loader
  • settings list display only visible items
  • checkpoint match when searching for model to load
  • video vae selection load correct vae

SD.Next Release 2025-05-06

Minor refesh with several bugfixes and updates to core libraries. Plus new features with FramePack and HiDream-E1

• Features
  • FramePack
    add T2V mode in addition to I2V and FLF2V
    support for new F1: forward-only model variant in addition to regular bi-directional
    add prompt enhance using VLM: it will analyze input image and then create enhanced prompt based on user prompt and image
    add prompt interpolation, section prompts do not need to match exact video section count
    and improved performance
    Docs rewrite!
  • Prompt-Enhhance
    add Qwen3 0.6B/1.7B/4B models
    add thinking mode support (for models that have it)
  • HiDream-E1 natural language image-editing model built on HiDream-I1
    available via networks -> models -> reference
    note: right now HiDream-E1 is limited to 768x768 images, so you must force resize image before running it
• Other
  • CUDA: set default to torch==2.7.0 with cuda==12.8
  • ZLUDA: update to zluda==3.9.4 and flash-attn-2
  • Docker: pre-install ffmpeg
  • Wiki: updated pages: FramePack, Video, ROCm, ZLUDA, Quantization
  • Gallery: support JXL image format
  • Scheduler: add sigmoid beta scheduler
  • GitHub: updated issue template
• Fixes
  • FramePack: correct dtype
  • NNCF: check dependencies and register quant type
  • API: refresh checkpoint list
  • API: vlm-api endpoint
  • Styles: save style with prompt
  • Texture tiling: fix apply when switching models
  • Diffusers: slow initial startup
  • Gated access: obfuscate and log token used for access
  • SDXL refiner workflow
  • Control: t2i-adapter workflow
  • Control: xs-controlnet workflow
  • Control: lllite-workflow
  • Control: refiner workflow with multiple control elements

SD.Next Release 2025-05-12

First of all NNCF quantization engine has gone through some major enhancements and its now much faster, both in quantization as well as actual inference!
And its a only truly cross-platform solution for quantization as all other methods are platform specific.

Note if you're a ZLUDA user, see notes on GPU driver compatibility as recent Andrenaline drivers do cause problems!
And if you're a ROCm user, this release brings much faster compile times on Linux as well as first (experimental) builds for Windows!

ReadMe | ChangeLog | Docs | WiKi | Discord

Details for 2025-05-12

• Compute
  • NNCF
    • Faster quantization
    • Faster inference with support for torch.triton
      up to 3.5x faster with INT4 and 2x faster with INT8
    • New settings: NNCF -> Group size
      default is a balance between performance (higher size) and quality (lower size)
      0 is default at 64, -1 disables grouping
  • ZLUDA:
    • warning: AMD Adrenaline 25.5.1 drivers are NOT COMPATIBLE with ZLUDA
      see issue for details
  • ROCm
    • first working builds of Torch with ROCm on Windows
      highly experimental
      reach out on Discord if you want to test it
• Features
  • Prompt Enhancer: support for img2img workflows
    in img2img prompt enhancer will first analyze input image and then incorporate user prompt to create enhanced prompt
  • FramePack
    • improve LoRA compatibility
    • add metadata to video
  • UI
    • ModernUI: support for History tab
    • ModernUI: support for FramePack tab
  • API
    • add /sdapi/v1/framepack endpoint with full support for FramePack including all optional settings
      see example: sd-extension-framepack/create-video.py
    • add /sdapi/v1/checkpoint endpoint to get info on currently loaded model/checkpoint
      see example: cli/api-checkpoint.py
    • add /sdapi/v1/prompt-enhance endpoint to enhance prompt using LLM
      see example: cli/api-enhance.py
      supports text, image and video prompts with or without input image
      note: if input image is provided, model should be left at default gemma-3-4b-it as most other LLMs do not support hybrid workflows
• Fixes
  • Latent Diffusion Upscale
  • Model load: support SDXL safetensors packaged without VAE
  • ROCm: disable cuDNN benchmark, fixes slow MIOpen tuning with torch==2.7
  • Extensions: use in-process installer for extensions-builtin, improves startup performance
  • FramePack: monkey-patch for dynamically installed av
  • Logging: reduce spam while progress is active
  • LoRA: legacy handler enable/disable
  • LoRA: force clear-cache on model unload
  • ADetailer: fix enable/disable
  • ZLUDA: improve compatibility with older GPUs

SD.Next Release 2025-05-15

Curious how your system is performing?
Run a built-in benchmark and compare to over 15k unique results world-wide: Benchmark data!
From slowest 0.02 it/s running on 6th gen CPU without acceleration up to 275+ it/s running on tuned GH100 system!

Also, since quantization is becoming a necessity for almost all new models, see comparison of different quantization methods available in SD.Next: Quantization
Hint: Even if you may not need quantization for your current model, it may be worth trying it out as it can significantly improve performance!

For ZLUDA users, this update adds compatibility with with latest AMD Adrenaline drivers

Btw, last few releases have been smaller, but more regular so do check posts about previous releases as features do quickly add up!

• Wiki
  • Updates for: Quantization, NNCF, WSL, ZLUDA, ROCm
• Compute
  • ZLUDA: update to zluda==3.9.5 with torch==2.7.0
    Note: delete .zluda folder so that newest zluda will be installed if you are using the latest AMD Adrenaline driver
  • NNCF: added experimental support for direct INT8 MatMul
• Feature
  • Prompt Enhance: option to allow/disallow NSFW content
• Fixes
  • OpenVINO: force cpu device
  • Gradio: major cleanup and fixing defaults and ranges
  • Pydantic: update to api types
  • UI defaults: match correct prompt components

ReadMe | ChangeLog | Docs | WiKi | Discord

SD.Next Release 2025-06-02

Highlights

This release is all about quantization: with new SD.Next own quantization method: SDNQ
SDNQ is based on NNCF, but has been re-implemented, optimized and evolved enough to become its own quantization method!
It's fully cross-platform, supports all GPUs and includes tons of quantization methods:

• 8-bit, 6-bit, 4-bit, 2-bit and 1-bit int and uint
• 8-bit e5, e4 and fnuz float

Also unlike most traditional methods, its also applicable to nearly all model types

Hint: Even if you may not need quantization for your current model, it may be worth trying it out as it can significantly improve performance or capabilities of your existing workflow! For example, you may not have issues with SD15 or SDXL, but you may have been limited running at high resolutions or with multiple ControlNet due to VRAM requirements - this will significantly reduce memory requirements. And on-the-fly quantization takes just few seconds during model load, there is no need to have multiple quant models permanently saved.

On a different topic, SD.Next Wiki & Docs and its UI Hints and UI Localization system are community efforts and any contributions are welcome!
You dont need any coding experience, but if you learned something and you find documentation either wrong or insufficient, please do suggest edits!
Take a look at Docs, Hints and Localization contribution guides

ReadMe | ChangeLog | Docs | WiKi | Discord

Details

• SDNQ Quantization
  • Renamed NNCF to SDNQ
  • Renamed quantization scheme names to the underlying dtype names instead of NNCF names
    • INT8_SYM -> int8
    • INT8 -> uint8
    • INT4_SYM -> int4
    • INT4 -> uint4
  • Add float8_e4m3fn, float8_e5m2, float8_e4m3fnuz, float8_e5m2fnuz, int6, uint6, int2, uint2 and uint1 support
  • Add quantized matmul support for float8_e4m3fn and float8_e5m2
  • Set the default quant mode to pre
  • Use per token input quant with int8 and fp8 quantized matmul
  • Implement better layer hijacks
  • Add an option to toggle quantize with GPU
  • Fix conv quant and add support for conv quant with asym modes
  • Fix lora weight change
  • Fix high RAM usage with pre mode
  • Fix scale and zero_point not being offloaded
• IPEX
  • Disabe Dynamic Attention by default on PyTorch 2.7
  • Remove GradScaler hijack and use torch.amp.GradScaler instead
• Feature
  • TeaCache support for HiDream I1
• Changes
  • Set the default attention optimizer to Scaled-Dot-Product on all backends
  • Enable Dynamic attention for Scaled-Dot-Product with ROCm, DirectML, MPS and CPU backends
• Fixes
  • Gallery duplicate entries
  • Prompt enhancement args mismatch

SD.Next Release 2025-06-30

New release with ~100 commits...So what's new? Well, its been a busy few weeks with new models coming out quite frequently:

• New T2I/I2I models: OmniGen-2, Cosmos-Predict2, FLUX.1-Kontext, Chroma
• Additional VLM models: JoyCaption Beta, MoonDream 2
• Additional upscalers: UltraSharp v2

And (as always) many bugfixes and improvements to existing features!

ReadMe | ChangeLog | Docs | WiKi | Discord

Details for 2025-06-30

• Models
  • Models Wiki page is updated will all new models
    note all new image models larger than 30GB, so offloading and quantization are necessary!
  • OmniGen2
    • OmniGen2 is a powerful unified multimodal model that supports t2i and i2i workflows and uses 4B transformer with Qwen-VL-2.5 4B VLM
    • available via networks -> models -> reference
  • nVidia Cosmos-Predict2 T2I 2B and 14B
    • Cosmos-Predict2 T2I is a new foundational model from Nvidia in two variants: small 2B and large 14B
    • available via networks -> models -> reference
    • note: 14B variant is a very large model at 36GB
    • note: this is a gated model, you need to accept terms and set your huggingface token
  • Black Forest Labs FLUX.1 Kontext I2I Dev variant
    • FLUX.1-Kontext is a 12B model billion parameter capable of editing images based on text instructions
    • model is primarily designed for image editing workflows, but also works for text-to-image workflows
    • requirements are similar to regular FLUX.1 although 2x slower
    • available via networks -> models -> reference
    • note: this is a gated model, you need to accept terms and set your huggingface token
  • lodestones Chroma
    • Chroma is a 8.9B parameter model based on FLUX.1-schnell and fully Apache 2.0 licensed
    • available via networks -> models -> reference
    • note: model is still in training so future updates will trigger re-download
    • large credits to @Trojaner for work on bringing Chroma support to SD.Next and all the optimizations around it!
  • JoyCaption Beta support (in addition to existing JoyCaption Alpha)
    • new version of highly popular captioning model
    • available via caption -> vlm caption
  • MoonDream 2 support (updated)
    • really good 2B captioning model that can work on different levels of detail
    • available via caption -> vlm caption
  • UltraSharp v2 support
    • one of the best upscalers (traditional, non-diffusion) available today!
    • available via process -> upscale -> chainner
• Changes
  • Update all core requirements
  • Support Remote VAE with Omnigen, Lumina 2 and PixArt
  • Enable quantization for captioning: Gemma, Qwen, SMOL, Florence, JoyCaption
  • Add --trace command line param that enables trace logging
  • Use Diffusers version of OmniGen
  • Control move global settings to control elements -> control settings tab
  • Control add setting to run hires with or without control
  • Update OpenVINO to 2025.2.0
  • Simplified and unified quantization enabled for options
• SDNQ Quantization
  • Add auto quantization mode
  • Add modules_to_not_convert support for post mode
  • Improve offload compatibility
  • Fix Qwen 2.5 with int8 matmul
  • Fix Dora loading
  • Remove per layer GC
  • Add support for XYZ grid to test quantization modes
    note: you need to enable quantization and choose what it applies on, then xyz grid can change quantization mode
    note: you can also enable 'add time info' to compare performance of different quantization modes
• API
  • Add /sdapi/v1/lora?lora=<lora_name> endpoint that returns full lora info and metadata
  • Add /sdapi/v1/controlnets?model_type=<model_type|all|None> endpoints that returns list of available controlnets for specific model type
  • Set default sampler to Default
• Fixes
  • IPEX with DPM2++ FlowMatch samplers
  • Invalid attention processor with ControlNet
  • LTXVideo default scheduler
  • Balanced offload with OmniGen
  • Quantization with OmniGen
  • Do not save empty params.txt file
  • Override params.txt using SD_PATH_PARAMS env variable
  • Add wheel to requirements due to pip change
  • Case-insensitive sampler name matching
  • Fix delete file with gallery views
  • Add SD_SAVE_DEBUG env variable to report all params and metadata save operations as they happen
  • Fix TAESD model type detection
  • Fix LoRA loader incorrectly reporting errors
  • Fix hypertile for img2img and inpaint operations
  • Fix prompt parser batch size
  • Fix process batch with batch count
  • Fix process batch double image save
  • Fix unapply texture tiling
  • Fix nunchaku batch support
  • Fix LoRA change detection on pipeline type change
  • Fix LoRA load order when it includes text-encoder data
  • Suppress torch empty logging
  • Improve TAESD live preview downscale handling

SD.Next Release 2025-07-29

This is a big one: simply looking at number of changes, probably the biggest release since the project started!

Feature highlights include:

• ModernUI has quite some redesign which should make it more user friendly and easier to navigate plus several new UI themes
  If you're still using StandardUI, give ModernUI a try!
• New models such as WanAI 2.2 in 5B and A14B variants for both text-to-video and image-to-video workflows as well as text-to-image workflow!
  and also FreePix F-Lite, Bria 3.2 and bigASP 2.5
• Redesigned Video interface with support for general video models plus optimized FramePack and LTXVideo support
• Fully integrated nudity detection and optional censorship with NudeNet
• New background replacement and relightning methods using Latent Bridge Matching and new PixelArt processing filter
• Enhanced auto-detection of default sampler types/settings results in avoiding common mistakes
• Additional LLM/VLM models available for captioning and prompt enhance
• Number of workflow and general quality-of-life improvements, especially around Styles, Detailer, Preview, Batch, Control
• Compute improvements
• Wiki & Docs updates, especially new end-to-end Parameters page

In this release we finally break with legacy with the removal of the original A1111 codebase which has not been maintained for a while now
This plus major cleanup of codebase and external dependencies resulted in ~55k LoC (lines-of-code) reduction and spread over ~750 files in ~200 commits!

We also switched project license to Apache-2.0 which means that SD.Next is now fully compatible with commercial and non-commercial use and redistribution regardless of modifications!

And (as always) many bugfixes and improvements to existing features!
For details, see ChangeLog

ReadMe | ChangeLog | Docs | WiKi | Discord

Details

• License
  • SD.Next license switched from aGPL-v3.0 to Apache-v2.0
    this means that SD.Next is now fully compatible with commercial and non-commercial use and redistribution regardless of modifications!
• Models
  • WanAI Wan 2.2 both 5B and A14B variants, for both T2V and I2V support
    go to: video -> generic -> wan -> pick variant
    optimized support with VACE, etc. will follow soon
    caution Wan2.2 on its own is ~68GB, but also includes optional second-stage for later low-noise processing which is absolutely massive at additional ~54GB
    you can enable second stage processing in settings -> model options, its disabled by default
    note: quantization and offloading are highly recommended regardless of first-stage only or both stages!
  • WanAI Wan T2V models for T2I workflows
    Wan is originally designed for video workflows, but now also be used for text-to-image workflows!
    supports Wan-2.1 in 1.3B and 14B variants and Wan-2.2 in 5B and A14B variants
    supports all standard features such as quantization, offloading, TAESD preview generation, LoRA support etc.
    can also load unet/transformer fine-tunes in safetensors format using UNET loader
    simply select in networks -> models -> reference
    note 1.3B model is a bit too small for good results and 14B is very large at 78GB even without second-stage so aggressive quantization and offloading are recommended
  • FreePix F-Lite in 7B, 10B and Texture variants
    F-Lite is a 7B/10B model trained exclusively on copyright-safe and SFW content, trained on internal dataset comprising approximately 80 million copyright-safe images
    available via networks -> models -> reference
  • Bria 3.2
    Bria is a smaller 4B parameter model built entirely on licensed data and safe for commercial use
    note: this is a gated model, you need to accept terms and set your huggingface token
    available via networks -> models -> reference
  • bigASP 2.5
    bigASP is an experimental SDXL finetune using Flow matching method
    load as usual, and leave sampler set to Default
    or you can use following samplers: UniPC, DPM, DEIS, SA
    required sampler settings: prediction-method=flow-prediction, sigma-method=flowmatch
    recommended sampler settings: flow-shift=1.0
  • LBM: Latent Bridge Matching
    very fast automatic image background replacement methods with relightning!
    simple: automatic background replacement using BiRefNet
    relighting: automatic background replacement with reglighting so source image fits desired background
    with optional composite blending
    available in img2img or control -> scripts
  • add FLUX.1-Kontext-Dev inpaint workflow
  • add FLUX.1-Kontext-Dev Nunchaku support
    note: FLUX.1 Kontext is about 2-3x faster with Nunchaku vs standard execution!
  • support FLUX.1 all-in-one safetensors
  • support for Google Gemma 3n E2B and E4B LLM/VLM models
    available in prompt enhance and process captioning
  • support for HuggingFace SmolLM3 3B LLM model
    available in prompt enhance
  • add fal AuraFlow 0.2 in addition to existing fal AuraFlow 0.3 due to large differences in model behavior
    available via networks -> models -> reference
  • add integrated NudeNet as built-in functionality
    note: used to be available as a separate extension
• Video
  • redesigned Video interface
  • support for Generic video models
    includes support for many video models without specific per-model optimizations
    included: Hunyuan, LTX, WAN, Mochi, Latte, Allegro, Cog
    supports quantization, offloading, frame interpolation, etc.
  • support for optimized FramePack
    with t2i, i2i, flf2v workflows
    LoRA support, prompt enhance, etc.
    now fully integrated instead of being a separate extension
  • support for optmized LTXVideo
    with t2i, i2i, v2v workflows
    optional native upsampling and video refine workflows
    LoRA support with different conditioning types such as Canny/Depth/Pose, etc.
  • support for post load quantization
• UI
  • major update to modernui layout
  • add new Windows-like Blocks UI theme
  • redesign of the Flat UI theme
  • enhanced look&feel for Gallery tab with better search and collapsible sections, thanks to @CalamitousFelicitousness
• WIKI
  • new Parameters page that lists and explains all generation parameters
    massive thanks to @CalamitousFelicitousness for bringing this to life!
  • updated Models, Video, LTX, FramePack, Styles, etc.
• Compute
  • support for SageAttention2++
    provides 10-15% performance improvement over default SDPA for transformer-based models!
    enable in settings -> compute settings -> sdp options
    note: SD.Next will use either SageAttention v1/v2/v2++, depending which one is installed
    until authors provide pre-build wheels for v2++, you need to install it manually or SD.Next will auto-install v1
  • support for torch.compile for LLM: captioning/prompt-enhannce
  • support for torch.compile with repeated-blocks
    reduces time-to-compile 5x without loss of performance!
    enable in settings -> model compile -> repeated
    note: torch.compile is not compatible with balanced offload
• Other
  • Styles can now include both generation params and server settings
    see Styles docs for details
  • TAESD is now default preview type since its the only one that supports most new models
  • support TAESD preview and remote VAE for HunyuanDit
  • support TAESD preview and remote VAE for AuraFlow
  • support TAESD preview for WanAI
  • SD.Next now starts with locked state preventing model loading until startup is complete
  • warn when modifying legacy settings that are no longer supported, but available for compatibilty
  • warn on incompatible sampler and automatically restore default sampler
  • XYZ grid can now work with control tab:
    if controlnet or processor are selected in xyz grid, they will overwrite settings from first unit in control tab,
    when using controlnet/processor selected in xyz grid, behavior is forced as control-only
    also freely selectable are control strength, start and end values
  • Batch warn on unprocessable images and skip operations on errors so that other images can still be processed
  • Metadata improved parsing and detect foreign metadata
    detect ComfyUI images
    detect InvokeAI images
  • Detailer add expert mode where list of detailer models can be converted to textbox for manual editing
    see docs for more information
  • Detailer add option to merge multiple results from each detailer model
    for example, hands model can result in two hands each being processed separately or both hands can be merged into one composite job
  • Control auto-update width/height on image upload
  • Control auto-determine image save path depending on operations performed
  • autodetect V-prediction models and override default sampler prediction type as needed
• SDNQ
  • use inference context during quantization
  • use static compile
  • rename quantization type for text encoders default option to Same as model
• API
  • add /sdapi/v1/lock-checkpoint endpoint that can be used to lock/unlock model changes
    if model is locked, it cannot be changed using normal load or unload methods
• Fixes
  • allow theme type None to be set in config
  • installer dont cache installed state
  • fix Cosmos-Predict2 retrying TAESD download
  • better handle startup import errors
  • fix traceback width preventing copy&paste
  • fix ansi controle output from scripts/extensions
  • fix diffusers models non-unique hash
  • fix loading of manually downloaded diffuser models
  • fix api /sdapi/v1/embeddings endpoint
  • fix incorrect reporting of deleted and modified files
  • fix SD3.x loader and TAESD preview
  • fix xyz with control enabled
  • fix control order of image save operations
  • fix control batch-input processing
  • fix modules merge save model
  • fix torchvision bicubic upsample with ipex
  • fix instantir pipeline
  • fix prompt encoding if prompts within batch have different segment counts
  • fix detailer min/max size
  • fix loopback script
  • fix networks tags display
  • fix yolo refresh models
  • cleanup control infotext
  • allow upscaling with models that have implicit VAE processing
  • framepack improve offloading
  • improve prompt parser tokenizer loader
  • improve scripts error handling
  • improve infotext param parsing
  • improve extensions ui search
  • improve model type autodetection
  • improve model auth check for hf repos
  • improve Chroma prompt padding as per recommendations
  • lock directml torch to torch-directml==0.2.4.dev240913
  • lock directml transformers to transformers==4.52.4
  • improve install of sentencepiece tokenizer
  • add int8 matmul fallback for ipex with onednn qlinear
• Refactoring
  note: none of the removals result in loss-of-functionality since all those features are already re-implemented
  goal here is to remove legacy code, code duplication and reduce code complexity
  • obsolete original backend
  • remove majority of legacy a1111 codebase
  • remove legacy ldm codebase: /repositories/ldm
  • remove legacy blip codebase: /repositories/blip
  • remove legacy codeformer codebase: /repositories/codeformer
  • remove legacy clip patch model: /models/karlo
  • remove legacy model configs: /configs/*.yaml
  • remove legacy submodule: /modules/k-diffusion
  • remove legacy hypernetworks support: /modules/hypernetworks
  • remove legacy lora support: /extensions-builtin/Lora
  • remove legacy clip/blip interrogate module
  • remove modern-ui remove only-original vs only-diffusers code paths
  • refactor control processing and separate preprocessing and image save ops
  • refactor modernui layouts to rely on accordions more than individual controls
  • refactore pipeline apply/unapply optional components & features
  • split monolithic shared.py
  • cleanup /modules: move pipeline loaders to /pipelines root
  • cleanup /modules: move code folders used by pipelines to /pipelines/<pipeline> folder
  • cleanup /modules: move code folders used by scripts to /scripts/<script> folder
  • cleanup /modules: global rename modules.scripts to avoid conflict with /scripts
  • override gradio installer
  • major refactoring of requirements and dependencies to unblock numpy>=2.1.0
  • patch insightface
  • patch facelib
  • patch numpy
  • stronger lint rules
    add separate npm run lint, npm run todo, npm run test, npm run format macros

SD.Next Release 2025-08-15

New release two weeks after the last one and its a big one with over 150 commits!

• Several new models: Qwen-Image (plus Lightning variant) and FLUX.1-Krea-Dev
• Several updated models: Chroma, SkyReels-V2, Wan-VACE, HunyuanDiT
• Plus continuing with major UI work with new embedded Docs/Wiki search, redesigned real-time hints, wildcards UI selector, built-in GPU monitor, CivitAI integration and more!
• On the compute side, new profiles for high-vram GPUs, offloading improvements, parallel-load for large models, support for new torch release and improved quality when using low-bit quantization!
• SD.Next Model Samples Gallery: pre-generated image gallery with 60 models (45 base and 15 finetunes) and 40 different styles resulting in 2,400 high resolution images!
  gallery additionally includes model details such as typical load and inference times as well as sizes and types of each model component (e.g. unet, transformer, text-encoder, vae)
• And (as always) many bugfixes and improvements to existing features!

ReadMe | ChangeLog | Docs | WiKi | Discord

Note: Change-in-behavior - locations of downloaded HuggingFace models and components are changed to allow for de-duplication of common modules and switched from using system default cache folder to models/huggingface
SD.Next will warn on startup on unused cache entries that can be removed. Also, to take advantage of de-duplication, you'll need to delete models from your models/Diffusers folder and let SD.Next re-download them!

Details

• Models
  • Qwen-Image
    new image foundational model with 20B params DiT and using Qwen2.5-VL-7B as the text-encoder!
    available via networks -> models -> reference
    note: this model is almost 2x the size of Flux, quantization and offloading are highly recommended!
    recommended params: steps=50, attention-guidance=4
    also available is pre-packaged Qwen-Lightning
    which is an unofficial merge of Qwen-Image with Qwen-Lightning-LoRA to improve quality and allow for generating in 8-steps!
  • FLUX.1-Krea-Dev
    new 12B base model compatible with FLUX.1-Dev from Black Forest Labs with opinionated aesthetics and aesthetic preferences in mind
    available via networks -> models -> reference
  • Chroma
    great model based on FLUX.1 and then redesigned and retrained by lodestones
    update with latest HD, HD Flash and HD Annealed variants which are based on v50 release
    available via networks -> models -> reference
  • SkyReels-V2
    SkyReels-V2 is a genarative video model based on Wan-2.1 but with heavily modified execution to allow for infinite-length video generation
    supported variants are:
    • diffusion-forcing: T2I DF 1.3B for 540p videos, T2I DF 14B for 720p videos, I2I DF 14B for 720p videos
    • standard: T2I 14B for 720p videos and I2I 14B for 720p videos
  • Wan-VACE
    basic support for Wan 2.1 VACE 1.3B and 14B variants
    optimized support with granular guidance control will follow soon
  • HunyuanDiT-Distilled
    variant of HunyuanDiT with reduced steps and improved performance
    Torch
  • Set default to torch==2.8.0 for CUDA, ROCm and OpenVINO
  • Add support for torch==2.9.0-nightly
• UI
  • new embedded docs/wiki search!
    Docs search: fully-local and works in real-time on all document pages
    Wiki search: uses github api to search online wiki pages
  • updated real-time hints, thanks @CalamitousFelicitousness
  • add Wilcards UI
    in networks display
  • every heading element is collapsible!
  • quicksettings reset button to restore all quicksettings to default values
    because things do sometimes get wrong...
  • configurable image fit in all image views
  • rewritten CivitAI downloader
    in models -> civitai
    hint: you can enter model id in a search bar to pull information on specific model directly
    hint: you can download individual versions or batch-download all-at-once!
  • redesigned GPU monitor
    • standard-ui: system -> gpu monitor
    • modern-ui: aside -> console -> gpu monitor
    • supported for nVidia CUDA and AMD ROCm platforms
    • configurable interval in settings -> user interface
  • updated models tab
    • updated models -> current tab
    • updated models -> list models tab
    • updated models -> metadata tab
  • updated extensions tab
  • redesigned settings -> user interface
  • gallery bypass browser cache for thumbnails
  • gallery safer delete operation
  • networks display indicator for currently active items
    applies to: styles, loras
  • apply privacy blur to hf and civitai tokens
  • image download will now use actual image filename
  • increase default and maximum ui request timeout to 2min/5min
  • hint: card layout
    card layout is used by networks, gallery, civitai search, etc.
    you can change card size in settings -> user interface
• Offloading
  • changed default values for offloading based on detected gpu memory
    see offloading docs for details
  • new feature to specify which modules to offload always or never
    in settings -> model offloading -> offload always/never
  • new highvram profile provides significant performance boost on gpus with more than 24gb
  • new offload during pre-forward option
    in settings -> model offloading
    switches from explicit offloading to implicit offloading on module execution change
  • new diffusers_offload_nonblocking exerimental setting
    instructs torch to use non-blocking move operations when possible
• Features
  • new T5: Use shared instance of text encoder option
    in settings -> text encoder
    since a lot of new models use T5 text encoder, this option allows to share
    the same instance across all models without duplicate downloads
    note this will not reduce size of your already downloaded models, but will reduce size of future downloads
  • Wan select which stage to run: first/second/both with configurable boundary ration when running both stages
    in settings -> model options
  • prompt parser allow explict BOS and EOS tokens in prompt
  • Nunchaku support for FLUX.1-Fill and FLUX.1-Depth models
  • update requirements/packages
  • use model vae scale-factor for image width/heigt calculations
  • SDNQ add modules_dtype_dict to quantize Qwen Image with mixed dtype
  • prompt enhance
    add allura-org/Gemma-3-Glitter-4B, Qwen/Qwen3-4B-Instruct-2507, Qwen/Qwen2.5-VL-3B-Instruct model support
    improve system prompt
  • schedulers add Flash FlowMatch
  • model loader add parallel loader option
    enabled by default, selectable in settings -> model loading
  • filename namegen use exact sequence number instead of next available
    this allows for more predictable and consistent filename generation
  • network delete new feature that allows to delete network from disk
    in networks -> show details -> delete
    this will also delete description, metadata and previews associated with the network
    only applicable to safetensors networks, not downloaded diffuser models
• Wiki
  • Models page updated with links to original model repos and model licenses, thanks @alerikaisattera
  • Updated Model-Support with newly supported models
  • Updated Offload, Prompting, API pages
• API
  • add /sdapi/v1/checkpoint POST endpoint to simply load a model
  • add /sdapi/v1/modules GET endpoint to get info on model components/modules
  • all generate endpoints now support sd_model_checkpoint parameter
    this allows to specify which model to use for generation without needing to use additional endpoints
• Refactor
  • change default huggingface cache folder from system default to models/huggingface
    sd.next will warn on startup on unused cache entries
  • new unified pipeline component loader in pipelines/generic
  • remove LDSR
  • remove api-only cli option
• Docker
  • update cuda base image: pytorch/pytorch:2.8.0-cuda12.8-cudnn9-runtime
  • update official builds: https://hub.docker.com/r/vladmandic/sdnext-cuda/tags
• Fixes
  • refactor legacy processing loop
  • fix settings components mismatch
  • fix Wan 2.2-5B I2V workflow
  • fix Wan T2I workflow
  • fix OpenVINO
  • fix video model vs pipeline mismatch
  • fix video generic save frames
  • fix inpaint image metadata
  • fix processing image save loop
  • fix progress bar with refine/detailer
  • fix api progress reporting endpoint
  • fix openvino backend failing to compile
  • fix zluda with hip-sdk==6.4
  • fix nunchaku fallback on unsupported model
  • fix nunchaku windows download links
  • fix Flux.1-Kontext-Dev with variable resolution
  • use utf_16_be as primary metadata decoding
  • fix sd35 width/height alignment
  • fix nudenet api
  • fix global state tracking
  • fix ui tab detection for networks
  • fix ui checkbox/radio styling for non-default themes
  • fix loading custom transformers and t5 safetensors tunes
  • add mtime to reference models
  • patch torch version so 3rd party libraries can use expected format
  • unified stat size/mtime calls
  • reapply offloading on ipadapter load
  • api set default script-name
  • avoid forced gc and rely on thresholds
  • add missing interrogate in output panel

SD.Next Release 2025-08-20

A quick service release with several important hotfixes, improved localization support and adding new Qwen model variants...

ReadMe | ChangeLog | Docs | WiKi | Discord

• Models
  • Qwen-Image-Edit
    Image editing using natural language prompting, similar to Flux.1-Kontext, but based on larger 20B Qwen-Image model
  • Nunchaku-Qwen-Image
    if you have a compatible nVidia GPU, Nunchaku is the fastest quantization engine, currently available for Flux.1, SANA and Qwen-Image models
    note: release version of nunchaku==0.3.2 does NOT include support, so you need to build nunchaku from source
• SD.Next Model Samples Gallery
  • updated with new models
• Features
  • new setting -> huggingface -> download method
    default is rust as new xet is known to cause issues
  • support for flux.1-kontext lora
  • support for qwen-image lora
  • new setting -> quantization -> modules dtype dict
    used to manually override quant types per module
• UI
  • new artwork for reference models in networks
    thanks @liutyi
  • updated localization for all 8 languages
  • localization support for ModernUI
  • single-click on locale rotates current locale
    double-click on locale resets locale to en
  • exclude ModernUI from list of extensions
    ModernUI is enabled in settings, not by manually enabling extension
• Docs
  • Models and Video pages updated with links to original model repos, model licenses and original release dates
    thanks @alerikaisattera
• Fixes
  • nunchaku use new download links and default to 0.3.2
    nunchaku wheels: https://huggingface.co/nunchaku-tech/nunchaku/tree/main
  • fix OpenVINO with offloading
  • add explicit offload calls on prompt encode
  • error reporting on model load failure
  • fix torch version checks
  • remove extra cache clear
  • enable explicit sync calls for rocm on windows
  • note if restart-needed on initial startup import error
  • bypass diffusers-lora-fuse on quantized models
  • monkey-patch diffusers to use original weights shape when loading lora
  • guard against null prompt
  • install hf_transfter and hf_xet when needed
  • fix ui cropped network tags
  • enum reference models on startup
  • dont report errors if agent scheduler is disabled

SD.Next release 2025-09-15

What's new? Big one is that we're (finally) switching the default UI to ModernUI, for both desktop and mobile use!
StandardUI is still available and can be selected in settings, but ModernUI is now the default for new installs

What's else? Chroma is in its final form, there are several new Qwen-Image variants and Nunchaku hit version 1.0!
Also, there are quite a few offloading improvements and many quality-of-life changes to UI and overal workflows
And check out new history tab in the right panel, it now shows visualization of entire processing timeline!

ReadMe | ChangeLog | Docs | WiKi | Discord | Sponsor

Details

• Models
  • Chroma final versions: Chroma1-HD, Chroma1-Base and Chroma1-Flash
  • Qwen-Image InstantX ControlNet Union support
    note qwen-image is already a very large model and controlnet adds 3.5GB on top of that so quantization and offloading are highly recommended!
  • Qwen-Lightning-Edit and Qwen-Image-Distill variants
  • Nuchaku variants of Qwen-Image-Lightning, Qwen-Image-Edit, Nunchaku-Qwen-Image-Edit-Lightning
  • Nunchaku variant of Flux.1-Krea-Dev
    if you have a compatible nVidia GPU, Nunchaku is the fastest quantization & inference engine
  • HunyuanDiT ControlNet Canny, Depth, Pose
  • KBlueLeaf/HDM-xut-340M-anime
    highly experimental: HDM Home-made-Diffusion-Model is a project to investigate specialized training recipe/scheme
    for pretraining T2I model at home based on super-light architecture
    requires: generator=cpu, dtype=float16, offload=none, both positive and negative prompts are required and must be long & detailed
  • Apple FastVLM in 0.5B, 1.5B and 7B variants
    available in captioning tab
  • updated SD.Next Model Samples Gallery
• UI
  • default to ModernUI
    standard ui is still available via settings -> user interface -> theme type
  • mobile-friendly!
  • new History section in the right panel
    shows detailed job history plus timeline of the execution
  • make hints touch-friendly: hold touch to display hint
  • improved image scaling in img2img and control interfaces
  • add base model type to networks display, thanks @Artheriax
  • additional hints to ui, thanks @Artheriax
  • add video support to gallery, thanks @CalamitousFelicitousness
  • additional artwork for reference models in networks, thanks @liutyi
  • improve ui hints display
  • restyled all toolbuttons to be modernui native
  • reodered system settings
  • dynamic direction of dropdowns
  • improve process tab layout
  • improve detection of active tab
  • configurable horizontal vs vertical panel layout
    in settings -> user interface -> panel min width
    example: if panel width is less than specified value, layout switches to verical
  • configurable grid images size
    in settings -> user interface -> grid image size
  • gallery now includes reference model images
  • reference models now include indicator if they are ready or need download
• Offloading
  • balanced
    • enable offload during pre-forward by default
    • improve offloading of models with multiple dits
    • improve offloading of models with impliciy vae processing
    • improve offloading of models with controlnet
    • more aggressive offloading of controlnets with lowvram flag
  • group
    • new offloading method, using type=leaf works on a similar level as sequential offloading
      and can present siginificant savings on low-vram gpus, but comes at the higher performace cost
• Quantization
  • option to specify models types not to quantize: settings -> quantization
    allows for having quantization enabled, but skipping specific model types that do not need it
    example: sd, sdxl
  • sdnq
    • add quantized matmul support for all quantization types and group sizes
    • improve the performance of low bit quants
  • nunchaku: update to nunchaku==1.0.0
    note: nunchaku updated the repo which will trigger re-download of nunchaku models when first used
    nunchaku is currently available for: Flux.1 Dev/Schnell/Kontext/Krea/Depth/Fill, Qwen-Image/Qwen-Lightning, SANA-1.6B
  • tensorrt: new quantization engine from nvidia
    experimental: requires new pydantic package which may break other things, to enable start sdnext with --new flag
    note: this is model quantization only, no support for tensorRT inference yet
• Other
  • refactor reuse-seed and add functionality to all tabs
  • refactor modernui js codebase
  • move zluda flash attenion to Triton Flash attention option
  • remove samplers filtering
  • allow both flow-matching and discrete samplers for sdxl models
  • cleanup command line parameters
  • add --new command line flag to enable testing of new packages without breaking existing installs
  • downgrade rocm to torch==2.7.1
  • set the minimum supported rocm version on linux to rocm==6.0
  • disallow zluda and directml on non-windows platforms
  • update openvino to openvino==2025.3.0
  • add deprecation warning for python==3.9
  • allow setting denoise strength to 0 in control/img2img
    this allows to run workflows which only refine or detail existing image without changing it
  • Detailer allow manually setting processing resolution
    note: this does not impact the actual image resolution, only the resolution at which detailer internally operates
• Fixes
  • normalize path hanlding when deleting images
  • unified compile upscalers
  • fix OpenVINO with ControlNet
  • fix hidden model tags in networks display
  • fix networks reference models display on windows
  • fix handling of pre-quantized flux models
  • fix wan use correct pipeline for i2v models
  • fix qwen-image with hires
  • fix omnigen-2 failure
  • fix auraflow quantization
  • fix kandinsky-3 noise
  • fix infiniteyou pipeline offloading
  • fix skyreels-v2 image-to-video
  • fix flex2 img2img denoising strength
  • fix flex2 contronet vs inpaint image selection, thanks @alerikaisattera
  • fix some use cases with access via reverse-proxy
  • fix segfault on startup with rocm==6.4.3 and torch==2.8
  • fix wildcards folders traversal, thanks @dymil
  • fix zluda flash attention with enable_gqa
  • fix wan a14b quantization
  • fix reprocess workflow for control with hires
  • fix samplers set timesteps vs sigmas
  • fix detailer missing metadata
  • fix infiniteyou lora load with

SD.Next Release 2025-10-17

It's been a month since the last release and number of changes is yet again massive with over 300 commits!
Highlight are:

• Torch: ROCm on Windows for AMD GPUs
  if you have a compatible GPU, performance gains are significant!
• Models:
  a lot of new stuff with Qwen-Image-Edit including multi-image edits and distilled variants,
  new Flux, WAN, LTX, HiDream variants, expanded Nunchaku support and new SOTA upscaler with SeedVR2
  plus improved video support in general, including new methods of video encoding
• Quantization:
  new SVD-style quantization using SDNQ offers almost zero-loss even with 4bit quantization
  and now you can also test your favorite quantization on-the-fly and then save/load model for future use
• Other: support for Huggingface mirrors, changes to installer to prevent unwanted torch-cpu operations, improved previews, etc.

Details for 2025-10-17

• Models
  • WAN 2.2 14B VACE
    available for text-to-image and text-to-video and image-to-video workflows
  • Qwen Image Edit 2509 and Nunchaku Qwen Image Edit 2509
    updated version of Qwen Image Edit with improved image consistency
  • Qwen Image Pruning and Qwen Image Edit Pruning
    pruned versions of Qwen with 13B params instead of 20B, with some quality tradeoff
  • Tencent FLUX.1 Dev SRPO
    SRPO is trained by Tencent with specific technique: directly aligning the full diffusion trajectory with fine-grained human preference
  • Nunchaku SDXL and Nunchaku SDXL Turbo
    impact of nunchaku engine on unet-based model such as sdxl is much less than on a dit-based models, but its still significantly faster than baseline
    note that nunchaku optimized and pre-quantized unet is replacement for base unet, so its only applicable to base models, not any of fine-tunes
    how to use: enable nunchaku in settings -> quantization and then load either sdxl-base or sdxl-base-turbo reference models
  • HiDream E1.1
    updated version of HiDream-E1 image editing model
  • LTXVideo 0.9.8
    updated version of LTXVideo t2v/i2iv model
  • SeedVR2
    originally designed for video restoration, seedvr works great for image detailing and upscaling!
    available in 3B, 7B and 7B-sharp variants, use as any other upscaler!
    note: seedvr is a very large model (6.4GB and 16GB respectively) and not designed for lower-end hardware, quantization is highly recommended
    note: seedvr is highly sensitive to its cfg scale, set in settings -> postprocessing
    lower values will result in smoother output while higher values add details
  • X-Omni SFT
    experimental: X-omni is a transformer-only discrete auto-regressive image generative model trained with reinforcement learning
• Features
  • Model save: ability to save currently loaded model as a new standalone model
    why? SD.Next always prefers to start with full model and quantize on-demand during load
    however, when you find your exact preferred quantization settings that work well for you,
    saving such model as a new model allows for faster loads and reduced disk space usage
    so its best of both worlds: you can experiment and test different quantization methods and once you find the one that works for you, save it as a new model
    saved models appear in network tab as normal models and can be loaded as such
    available in models tab
  • Qwen Image-Edit multi-image editing
    requires qwen-image-edit-2509 or its variant as multi-image edits are not available in original qwen-image
    in ui control tab: inputs -> separate init image
    add image for input media and control media
    can be
  • Cache-DiT
    cache-dit is a unified, flexible and training-free cache acceleration framework
    compatible with many dit-based models such as FLUX.1, Qwen, HunyuanImage, Wan2.2, Chroma, etc.
    enable in settings -> pipeline modifiers -> cache-dit
  • Nunchaku Flux.1 PulID
    automatically enabled if loaded model is FLUX.1 with Nunchaku engine enabled and when PulID script is enabled
  • Huggingface mirror in settings -> huggingface
    if you're working from location with limited access to huggingface, you can now specify a mirror site
    for example enter, https://hf-mirror.com
• Compute
  • ROCm for Windows
    support for both official torch preview release of torch-rocm for windows and TheRock unofficial torch-rocm builds for windows
    note that rocm for windows is still in preview and has limited gpu support, please check rocm docs for details
  • DirectML warn as end-of-life
    torch-directml received no updates in over 1 year and its currently superseded by rocm or zluda
  • command line params --use-zluda and --use-rocm will attempt desired operation or fail if not possible
    previously sdnext was performing a fallback to torch-cpu which is not desired
  • installer if --use-cuda or --use-rocm are specified and torch-cpu is installed, installer will attempt to reinstall correct torch package
  • installer warn if cuda or rocm are available and torch-cpu is installed
  • support for torch==2.10-nightly with cuda==13.0
• Extensions
  • Agent-Scheduler
    was a high-value built-in extension, but it has not been maintained for 1.5 years
    it also does not work with control and video tabs which are the core of sdnext nowadays
    so it has been removed from built-in extensions: manual installation is still possible
  • DAAM: Diffusion Attentive Attribution Maps
    create heatmap visualizations of which parts of the prompt influenced which parts of the image
    available in scripts for sdxl text-to-image workflows
• Offloading
  • improve offloading for pipelines with multiple stages such as wan-2.2-14b
  • add timers to measure onload/offload times during generate
  • experimental offloading using torch.streams
    enable in settings -> model offloading
  • new feature to specify which models types not to offload
    in settings -> model offloading -> model types not to offload
• UI
  • connection monitor
    main logo in top-left corner now indicates server connection status and hovering over it shows connection details
  • separate guidance and detail sections
  • networks ability to filter lora by base model version
  • add interrogate button to input images
  • disable spellchecks on all text inputs
• SDNQ
  • add SVDQuant quantization method support
  • make sdnq scales compatible with balanced offload
  • add int8 matmul support for RDNA2 GPUs via triton
  • improve int8 matmul performance on Intel GPUs
• Other
  • server will note when restart is recommended due to package updates
  • interrupt will now show last known preview image
    keep incomplete setting is now save interrupted
  • logging enable debug, docs and api-docs by default
  • logging add detailed ram/vram utilization info to log
    logging frequency can be specified using --monitor x command line param, where x is number of seconds
  • ipex simplify internal implementation
  • refactor to use new libraries
  • styles and wildcards now use same seed as main generate for reproducible results
  • api new endpoint POST /sdapi/v1/civitai to trigger civitai models metadata update
    accepts optional page parameter to search specific networks page
  • reference models additional example images, thanks @liutyi
  • reference models add model size and release date, thanks @alerikaisattera
  • video support for configurable multi-stage models such as WAN-2.2-14B
  • video new LTX model selection
  • replace pynvml with nvidia-ml-py for gpu monitoring
  • update loopback script with radon seed option, thanks @Rabanti
  • vae slicing enable for lowvram/medvram, tiling for lowvram, both disabled otherwise
  • attention remove split-attention and add explicitly attention slicing enable/disable option
    enable in settings -> compute settings
    can be combined with sdp, enabling may improve stability when used on iGPU or shared memory systems
  • nunchaku update to 1.0.1 and enhance installer
  • xyz-grid add guidance section
  • preview implement configurable layers for WAN, Qwen, HV
  • update swagger /docs endpoint style
  • add [epoch] to filename template
  • starting [seq] for filename template is now higher of largest previous sequence or number of files in folder
• Video
  • use shared T5 text encoder for video models when possible
  • use shared LLama text encoder for video models when possible
  • unified video save code across all video models
    also avoids creation of temporary files for each frame unless user wants to save them
  • unified prompt enhance code across all video models
  • add job state tracking for video generation
  • fix quantization not being applied on load for some models
  • improve offloading for ltx and wan
  • fix model selection in ltx tab
• Experimental
  • new command line flag enables new pydantic and albumentations packages
  • modular pipelines: enable in settings -> model options
    only compatible with some pipelines, invalidates preview generation
  • modular guiders: automatically used for compatible pipelines when modular pipelines is enabled
    allows for using many different guidance methods:
    CFG, CFGZero, PAG, APG, SLG, SEG, TCFG, FDG
• Wiki
  • updates to AMD-ROCm, ZLUDA, LoRA, DirectML, SDNQ, Quantization, Prompting, LoRA pages
  • new Stability-Matrix page
• Fixes
  • Microsoft Florence 2 both base and large variants
    note this will trigger download of the new variant of the model, feel free to delete older variant in huggingface folder
  • MiaoshouAI PromptGen 1.5/2.0 in both base and large variants
  • fix prompt scheduling, thanks @nolbert82
  • ui: fix image metadata display when switching selected image in control tab
  • framepack: add explicit hf-login before framepack load
  • framepack: patch solver for unsupported gpus
  • benchmark: remove forced sampler from system info benchmark
  • xyz-grid: fix xyz grid with random seeds
  • reference: fix download for sd15/sdxl reference models
  • fix checks in init/mask image decode
  • fix hf token with extra chars
  • image viewer refocus on gallery after returning from full screen mode
  • fix attention guidance metadata save/restore
  • vae preview add explicity cuda.sync