Version 1.41.1

Download krita_ai_diffusion-1.41.1.zip

Installer Changes

• Removed pycairo dependency #2134 #2139 #2140
  It was causing installation problems on Linux and macOS systems that were unable to build it.
• Prompt to re-install if a previous upgrade attempt failed (rather than starting the old server backup)

Version 1.41.0

Download krita_ai_diffusion-1.41.0.zip

Text Prompts: Wildcards

Wildcards can be used in text to randomly select between various options. For example, a {red|blue} car will generate either a a red car or a blue car everytime you hit Generate. See Wildcards in Prompts for what is supported so far.

What type of 1girl are you? Now you no longer have to decide.

Text Prompts: Reference Layers

You can now embed reference images via the text prompt with <layer:name of a layer>. For latest Qwen Image Edit it allows to refer to layers by name. See Reference Layers in Prompts for more information.

Image Metadata

• Fixed issues where steps / CFG were not stored to saved PNG files #2012
• Changed prompt written to PNG files to be the fully expanded (including style)
• Added info about control layers to tooltip #1996
• Added info about wildcard choices to tooltip
• Changed "Copy Info to Clipboard" to JSON format. It now has a lot of additional information that does not fit into the tooltip.

Other Changes

• New Guide: Edit Models
• Improved performance of image transfer between Krita and ComfyUI #1265
• Added confirmation dialog when discarding images from history
• Disabled batch size for Qwen Nunchaku (it doesn't work yet) #2114
• More robust model detection with progress indicator when connecting #2079 #1796
• Allow the plugin to connect to remote ComfyUI server over Zrok (by @21dh114172)
• Fixed not being able to override architecture for Flux & Qwen in style settings
• Fixed consistency issues in edit mode (strength slider, copy prompt)

Version 1.40.1

Download krita_ai_diffusion-1.40.1.zip

Changes

• Fixed context area for Fill / Expand / Remove-Content being smaller than intended #2095
• Fixed an issue where text encoder models were not found on custom ComfyUI installs #2082
• Fixed validation error "CLIPLoader: Value not in list" related to custom workflow import #2087
• Fixed Illustrious models not being selectable in the installer #2096
• Fixed Qwen image generation without prompt #2091

Version 1.40.0

Download krita_ai_diffusion-1.40.0.zip

Qwen Image (experimental)

Qwen image models are some of the largest open models that can run locally. They offer high precision and great text understanding, but require a lot of GPU power to run. They support a large variety of content (subjects and styles) and understand many languages without external translation.

There are also Qwen edit models for instruction based changes.

See the Discussion for supported features and model downloads. Thanks @Danamir for the PR!

Edit Toggle

You can now link a regular image model with an edit model, to swap between tasks more easily.

First, make sure you have Styles for both your image model (eg. Illustrious) and the edit model (eg. Flux Kontext). Then link them in the Style settings.

Now when you have the image Style active, you can toggle Edit mode. It remembers its own prompt and reference images.

Other Changes

• Added Indonesian translations by @DDAANN222
• Hide alternative Flux models in installer if they're not recommended/supported by hardware #2048
• Replace WaveSpeed cache with ComfyUI's EasyCache (custom node is no longer installed)
• Updated default styles for more recent models and edit model links
• Only require clip-vision-vit-h model for SD1.5 / SDXL, it is no longer needed to connect
• Increased limits for maximum history size
• Improved typing performance in text prompt box for large documents
• Fixed compatibility with latest ComfyUI #2077 #2080
• Fixed naming collision when running download_models.py script
• Fixed incorrect sha256 for SD1.5 face Lora

Version 1.39.2

Download krita_ai_diffusion-1.39.2.zip

Fixes an installation issue due to broken dependency package on Windows #2025 #2028

Version 1.39.0

Download krita_ai_diffusion-1.39.1.zip

Changes

• Added option (default: off) to save generation metadata to images saved directly from the history (by @sfisher)
• Added option (default: PNG) to select image format when saving images directly from the history (by @nolan778)
• Clearing the generation history now also removes the preview layer #1926
• Prompts can now contain comments: text after a # is ignored until the next line break
• The negative prompt text box is now disabled when a model doesn't use negative prompts
• Fixed a hang/crash when images finish faster than they can be saved to document history. This usually only happened when producing the same image due to fixed seed.
• Prioritize Flux ControlNet models over Flux Tools LoRA #2000
• Edit models (like Flux Kontext) can now be used for batch generation in the animation workspace
• Custom graphs: Mask returned by the "Krita Selection" node is now fully opaque if there is no selection or the entire image is selected #1997
• Custom graphs: "Krita Selection" node now has a bool output to indicate if there is an active selection (requires nightly version of comfyui-tooling-nodes)
• External server: Retry connection a few times at Krita startup when ComfyUI isn't running yet #2004
• External server: No longer demand Clip Vision (ViT-H) model for base models that don't use it (Flux, Illu, Chroma) #2007
• Managed install: Added Flux Nunchaku fp4 models for users of NVIDIA 5000 series GPUs #1991
• Managed install: show a warning message when trying to install CUDA backend without a supported GPU

Version 1.38.0

Download krita_ai_diffusion-1.38.0.zip

Flux Nunchaku

Flux models can now be downloaded and installed via the plugin's installer. Support was added for SVDQuant/Nunchaku quantized models, which are smaller and faster.

• Added built-in style templates for Flux Kontext
• Flux styles will use the new "Flux - Turbo" sampler for Live generation (requires installing Turbo LoRA)

Users with a custom ComfyUI setup will have to install Nunchaku to use it. You can find recommended Flux models here. Please also update comfyui-tooling-nodes!

Important

Nunchaku requires a NVIDIA GPU (RTX 2000 series or newer).

Other Changes

• Fixed some issues with server upgrades from older versions #1924
  • if you still encounter issues, please back up your models and re-install
  • occasionally a fresh install is required to make sure new features are compatible
• Added subtle progress indicator to Live previews #1837
• Added option in performance settings to disable multi-threading
• Added option "history_format" in settings.json to change file format in which image history is stored to .kra files
• Upgraded to PyTorch 2.8.0 (thanks to @SHoogstad for testing XPU)
• No longer require Python to be installed on the system (Linux/MacOS)

Version 1.37.0

Download krita_ai_diffusion-1.37.2.zip

Chroma (experimental)

Chroma is a diverse open-source model based on the Flux architecture. It can generate photography and various art styles.

It's currently still in a "raw" state. When optimized versions, finetunes and control models come out in the future it will allow to cover more of the plugin features.

Model download (manual)

Download one of these and put it in /models/diffusion_models (the unet folder also works):

• BF16: chroma-unlocked-v50.safetensors
• FP8: chroma-unlocked-v50_float8_e4m3fn_learned_svd.safetensors
• Quantized versions (GGUF)

You also need VAE and text encoder (T5). They are the same as Flux, if you have that running you can skip.

• ae.safetensors → /models/vae
• t5xxl_fp8_e4m3fn_scaled.safetensors → /models/text_encoders

The Flux discussion has some more information. The prerequisites are the same, but note that LoRA, Control models, etc. are not compatible.

Important

Make sure ComfyUI & comfyui-tooling-nodes are up-to-date if you're running a custom server!

How to Use

There is a built-in Chroma style with recommended settings. Use natural language to prompt.

Other Changes

• Added Italian translation (by @FabioLe23)
• Added option to cancel and replace queued up jobs when clicking "Generate", can be triggered with Ctrl+Click (by @FeepingCreature)
• Fixed "Enqueue at Front" option causing all later jobs to no longer be tracked #1917
• Improved performance when typing prompts or tweaking sliders
• Improved performance when saving large images to generation history
• Managed server: Fixed an issue where old Python environments were used even after upgrading to a new version #1842
• Managed server: Create empty folder for model directories to make it easier to manually add models later
• Managed server: Improved initial package selection in installer UI
• Managed server: Try to fix MacOS PyTorch install
• Improved name-matching for T5 text encoder to avoid confusion with some other T5 variants #1866 #1852
• Provide layer image to custom animation workflows even if it's not part of an animation timeline #1882
• Improve error messages for validation and timeout

Version 1.36.1

Download krita_ai_diffusion-1.36.1.zip

Flux Kontext (experimental)

Flux Kontext is an edit model: you give it instructions to change your image.

How to use

• You always need an existing image! The canvas will be used as input, even at 100% strength.
• Use the prompt to tell the model what to change. Prompting guide
• You can use selections to restrict changes - this helps with larger images
• Add a "Reference" control layer for additional image input (don't overdo it)

Model download (manual)

• flux1-dev-kontext_fp8_scaled.safetensors → /models/diffusion_models

Also needed in case you don't have them already for Flux:

• ae.safetensors → /models/vae
• clip_l.safetensors → /models/text_encoders
• t5xxl_fp8_e4m3fn_scaled.safetensors → /models/text_encoders

See also the ComfyUI documentation and the Flux discussion for more information.

Version 1.35.1

Download krita_ai_diffusion-1.35.1.zip

Changes

• Allow SDXL Xinsir Union ControlNet to be used as fallback for Illustrious models #1803
• Added a keyboard shortcut for the custom graph workspace #1812
• Raise an error when trying to generate with the canvas as input, but all layers are invisible
• Fixed managed installation for DirectML #1829
• Disable XPU extension when launching in CPU mode #1813
• Improved XPU performance for large resolutions #1801 (by @SHoogstad)
• Increased connection timeouts #1796
• Add a default VAE for Illustrious models (when not using integrated or custom VAE)
• Fixed system locale detection for launching and communicating with processes