Version 1.51.0

Download krita_ai_diffusion-1.51.0.zip

For Krita 5 (recommended: Krita 5.3.1)

Control layers for Anima

Supported control modes:

• for official v1.0: Scribble, Line Art, Unblur
• for old preview models: Depth, Pose

Fill, Expand, etc. at full strength now requires the inpaint control model and gives much better results.

Please check the discussion for model downloads! PR by @Sen-sou

Other Changes

• Added ROCm as managed install option for supported AMD GPUs #76 #2315 #2493
• Fixed wrong assignments of <layer:name> references in prompts
• De-duplicate references to the same layer in more cases so they are only passed to the model once
• Fixed crash when a single-channel layer is used as input (eg. transparency mask) #2492
• Fixed the upscale strength slider being permanently disabled after selecting an edit model #2499
• Fixed issues when opening a document, especially with custom workflows #2479
• Changed color match functionality to detect when there is no color information outside the selection #2483
• Improved error message when trying to use reference layers in the prompt, but the model does not support them
• Improved file handling when come files are inaccessible (eg. access denied due to permissions)

Version 1.50.0

Download krita_ai_diffusion-1.50.0.zip

New Models (Preview)

Anima

• 2B parameter model for Anime and 2D, still being trained
• Supports tag based and natural language prompts
• Preview models can now be used for generation (#2450 by @FeepingCreature)
• See #2477 for downloads and more information

ERNIE Image

• 8B parameter model good for precise instructions and text
• Base and Turbo models can now be used for generation (#2459 by @eruanno123)
• See #2478 for downloads and more information

Other Changes

• Fixed error when switching to Live workspace related to seed input box #2454
• Fixed selection modes not available after opening an existing document #2469 #2474
• Don't add automatic padding in custom workflows when selection mode is "mask" #2411

Version 1.49.1

Download krita_ai_diffusion-1.49.1.zip

for Krita 5.2 or 5.3

Changes

• Fixed "refine" with less than 100% strength not working correctly for Flux 2 Klein #2392
• Use multiple-of-16 resolutions for internal resize to avoid artifacts at bottom/right borders #2436
• Custom workflows: Added a button to reset all parameters to default values #2402
• Fixed style not updating when switching between multiple documents that use different styles
• Fixed UI showing welcome state after opening a document instead of switching to image generation
• Fixed Flux 1 styles missing their checkpoint on non-CUDA installs
• Managed install now tries to install an older version of PyTorch for NVIDIA GTX 10xx #1871 #2396 #2426
• Changed the fixed seed widget to support larger numbers (up to 2^32) (by @LennyArchive)
• Fixed missing access token for custom ComfyUI authentication (by @sohaieb)
• Custom workflows: don't crop the preview image, allow workflows to return smaller or larger results #2418 #2419
• Custom workflows: fixed that some padding was still applied to selections masks even when "mask bounds"
• Assume SDXL-like checkpoints are Illustrious architecture when SDXL workload has not been installed to avoid warnings
• Refer to "workloads" on the connection screen in settings #2400

Version 1.49.0

Download krita_ai_diffusion-1.49.0.zip

Favourite Styles

When switching Styles, the most recent ones are now listed on top.

• Number of recent Styles can be configured in Interface settings (set to 0 to disable)
• Renamed some built-in styles to encourage their use over old/obsolete models

Prompt Hightlighting

• Special instructions in prompts are now highlighted
• Hex color codes are no longer treated as comments

Other Changes

• Added a button to toggle edit mode in the live workspace
• Flux 2 styles can now be linked as edit styles #2369
• Enabled "Scribble" control mode for Z-Image
• Custom graph parameters are now remembered for each individual workflow #2362
• Fixed selected custom graph now being restored when loading a document in some cases #2363
• Fixed animation frame batch generation with Flux 2 in edit mode #2372
• Disabled auto-exclusion of control and preview layers in custom graph live mode to avoid interruption #2376
• Fixed bad results with some models when using CFG with an empty negative prompt #2382
• Fixed spacing for Generate and Queue buttons on high-DPI displays
• Fixed issue with "Expand" selections when they were not overlapping with existing content

Version 1.48.0

Download krita_ai_diffusion-1.48.0.zip

Flux 2 Fill & Expand improvements

Improved behavior when using Fill or Expand with Flux 2 4B, especially without prompt.

Original	Old (v1.47.0)
New (v1.48.0)	New + Color Match

Color match

• Works for all models, but strongest effect on edit models
• Can be disabled in Diffusion settings
• Requires a selection. Matches colors outside of the selection area.

Note

External/custom ComfyUI installs

• Outpaint LoRA is used for Fill/Expand: download → models/loras/flux-2-klein-4B-outpaint-lora.safetensors
• Use "Generate (Custom)" to use regular inpainting (no edit, no LoRA)
• Remember to update comfyui-inpaint-nodes

Other Changes

• Added ControlNet support for Qwen image model by @krigeta #2332
• Fixed Lora widget showing incomplete list or displaying wrong text #2357
• Strength now fully applies to feather modifiers for selections [documentation]
• Fixed selection blend mask being too large when it connects with the borders of the canvas
• Custom workflows: Fixed regular/live sampling parameters passed to custom workflows #2341
• Custom workflows: Added a mask output to KritaCanvas custom node #2354
• Custom workflows: Sliders and number edit widgets now have more sensible step values #2342
• Fixed prompt metadata when using an empty prompt
• Bring back "Add Content" fill mode that was accidentaly removed in a prior version
• Added alternative Flux2 4B model download for non-CUDA managed install
• Deprecated DirectML support. It is still supported, but not recommended - AMD users are usually better off installing ComfyUI manually with ROCm support
• Added a button to collect diagnostics and system information, which can be copied into bug reports
• Removed Omni upscale models 3x/4x from required models list
• Fixed missing extension when creating a default style

Version 1.47.0

Download krita_ai_diffusion-1.47.0.zip

Flux 2 and Z-Image support

Flux 2 [klein] and Z-Image are now supported "officially", with automatic local installation and availability on cloud.

Left: generated with Z-Image | Right: series of edits with Flux 2 klein

Managed Install

Online Service [Interstice.cloud]

• Added Flux 2 klein
• Added Z-Image Turbo
• To be removed: Flux 1 schnell will no longer be available after 2026/03/01
• Plugin can now queue the next request already while it is receiving result images, allows faster batch generation

Other Changes

• Custom workflows: added new node "Krita Style & Prompt" by @leha-games #2276
  • get prompt processing (style merge, loras and wildcards) in custom workflows [documentation]
• Limit selection blend (blur for alpha compositing) to outside of user selection #2299
• Fixed reference images for edit not added to batch generation #2300
• Fixed tiled upscale with Flux 2 (was not working in some cases because of tile size mismatch)
• Added edit toggle checkbox in custom inpaint menu #2307
• Fixed prompt weight adjustment (Ctrl+Up/Down) when prompt text contains non-ASCII characters (eg. emojis) #2306
• Fixed potential cause for assertion popup when opening a document #2298
• Improved match of "Qwen2.5-VL-7B" for Qwen-Image text encoder #2311
• Added "res multistep" sampler to built-in sampler options
• Mapped special inpaint modes (add/remove object, replace background) to edit instructions when using Flux 2
• Managed install: don't modify shell/environment when installing uv
• Fixed change of edit prompt not being saved to document
• Fixed resource ID for flux VAEs in models.json and download_models script

Version 1.46.1

Download krita_ai_diffusion-1.46.1.zip

This release contains some tweaks to selection handling after more testing and feedback.

Changes

• Fixed selection feather size being larger than intended and sometimes exceeding crop size #2287
  • If you run a custom ComfyUI server, please update comfyui-inpaint-nodes to get this fix!
• Make selection padding take minimum feather amount into account
• Scale selection feather size with denoise strength (less feather when using low strength)
• Use lower initial resolution in 2-pass generation for inpaint models #2263
• Don't do 2-pass generation if the resolution gap between low-pass and high-pass is tiny
• Fixed selection masks not working in the first step of diffusion in some cases due to numerical issues
• Fixed control range slider not being displayed even when supported #2285 (by @nabileisa)
• Fixed not being able to switch back from edit mode (with linked edit style) #2282
• Match style checkpoints between Windows/Linux path conventions #2271
• Managed install: pin numpy on initial install #2283
• Managed install: clear previous errors when using one of the "Manage" actions
• Handbook: update custom workflow example
• Handbook: add information about Flux 2

Version 1.46.0

Download krita_ai_diffusion-1.46.0.zip

Flux 2 klein (experimental)

Now supported as preview.

Flux 2 "klein" is a smaller variant of the Flux 2 model by Black Forest Labs. While visual fidelity is not quite as good as Z-Image, it is very interesting because it combines generation and editing in a single model. It often beats other edit models at a fraction of the size and computation cost.

Check the discussion for instructions on how to use.

Inpaint Overhaul

This version makes changes to how Fill/Expand/Add/Remove work, and how selections are handled in general.

• "Selection Grow" setting was removed. Please adjust the "Selection Feather" setting accordingly (recommended: 10%).
• "Selection Blend" setting was added. It defines the feather/blur size to use for alpha compositing in pixels.
• Added tiny fixed size dilation to avoid issues with outlines of object selections
• Added minimum feather size in pixels to deal with small selections
• Note: setting Feather/Blend to 0 will disable all selection processing by the plugin

Feather is used as denoising mask and should be generous to integrate changes into the image. Blend is used for compositing and can be much smaller.

Z-Image Changes

Fill with Z-Image can now produce very convincing results for realistic images.

• Recommended: use the 2601 version of Z-Image-Turbo-Fun-Controlnet-Union-2.1
• Support "Lite" versions of Z-Image union controlnet (smaller, faster, and great quality)
• Support "Tile" controlnet model for "Unblur" and upscaling
• Improved masks passed to Z-Image controlnet when using selections

Other Changes

• Managed install: fixed issue with prompt translation #2257
• Managed install: fixed issues with numpy version when installing "Face" control layers #2262
• Managed install: fixed old version of tooling nodes being installed
• Managed install: use torch-2.8.0+cu128 on Linux to be compatible with Nunchaku builds #2272 #2256
• Fixed Z-Image output with empty prompt
• Skip encoding negative prompts when CFG is 1
• Improved VAE model detection for Z-Image

Version 1.45.0

Download krita_ai_diffusion-1.45.0.zip

Custom Workflows

This update brings various features to improve custom inpaint and upscale workflows.

Selection node

• Can now define a crop region to only transfer part of the image and mask
• Context crop area is based on the selection size, with optional extra padding
• Crop area affects the selection mask, and the Canvas node (but not the Layer nodes)
• Documentation | Example Workflow

Output node

• Can now specify a pixel offset where the result is inserted into the canvas
• Results can be named (used as layer name when applied)
• Added options to interpret batches of images as separate results, or animation frames, or individual layers
• Added option to resize canvas to the output size when it is applied in Krita (by @Kebolder)
• Documentation

Fixes

• Fixed Parameter node min/max being reset to default when it's set to 0
• Fixed detection of integer widget for Parameter node

Note

Remember to update comfyui-tooling-nodes if you are using a custom install!

Other Changes

• Support Z-Image Nunchaku SVDQ model (by @fukc-gihtub)
  • Note: no Lora or ControlNet for now, use the regular model or GGUF if you need that
• Detect Qwen Image Edit 2511 by name and set the correct default architecture #2245
• Improved performance of Style Lora UI #2247
• Managed install: Upgrade PyTorch to 2.9.1 and Nunchaku to 1.1.0
• Fixed manual typing of installation path in managed server UI
• Fixed switch widget style not being consistent
• Better initial settings window size for higher screen resolutions

Version 1.44.1

Download krita_ai_diffusion-1.44.1.zip

Changes

• Fixed a bug where only the first image in a batch could be selected and applied #2239 #2241
• Managed install: improve robustness of downloads and allow timeout
• Managed install: resolved some issues where errors were not reported or logged
• Managed install: try to use default system shell to install uv on Linux