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

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.

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

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

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

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

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

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

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

Download krita_ai_diffusion-1.44.0.zip

Z-Image Control

The Z-Image-Turbo-Fun-Controlnet-Union model is now supported.

• Adds Pose, Canny, Soft Edge and Depth control layers
• Can be used as an inpaint model when filling selections
  • Use "Generate (Custom)" and uncheck "Seamless" to disable this
• No support for timestep range (start/end) yet, so there is some quality degradation
• See #2188 for downloads and instructions

Qwen Image Layered

Qwen Image Layered is a model that decomposes existing images into transparent layers.

• Use it on an existing flattened image to split it into Krita layers, which can then be edited with traditional tools
• Like other Qwen Image models this is a heavy model and generation times are very slow
• See the Qwen Image Layered section in #2081 on how to download and use this model

Other Changes

• Custom workflows: add text-based number input in addition to sliders (by @Kebolder)
• Custom workflows: allow overwriting existing workflow files when saving a workflow (by @Kebolder)
• Fixed error popup "Upscale model not found" at startup #2209
• Fixed issue where Qwen text encoder was not detected anymore #2200
• Fixed text prompt resize handle not working when negative prompt is disabled #2202
• Fixed wrong style selected when loading a document while not connected to the server yet
• Fixed "Copy Prompt (evaluated)" not setting the negative prompt #2226
• Allow cancelling Cloud jobs with per-document granularity
• Removed built-in SD1.5 Comic/Anime style

Download krita_ai_diffusion-1.43.0.zip

Z-Image (experimental)

Z-Image is a new diffusion model which produces convincing photorealistic images out of the box. Its modest size (larger than SDXL but smaller than Flux1), reasonable generation times and open license make it a promising candidate for future extensions.

Only basic features are supproted for now. Check the discussion for how to use and current status.

Queue improvements

• Counter now reflects the queued jobs for the current document
• The drop-down also reveals how many jobs are queued across the Krita plugin (does not include jobs from other ComfyUI clients)
• Will now display a "waiting" icon when there are queued jobs that are waiting for jobs from other documents
• Cancel now only cancels jobs from the current document (previously canceled all jobs, even those from other ComfyUI clients)
• Fixed some issues were the count wasn't updated correctly

Other Changes

• Support Edit models in the upscale workspace #1896
• Fixed LoRA weight not appearing correctly in metadta tooltip #2169
• Fixed crash when closing a document and re-opening it #2185
• Fixed deleting animation results from history
• Added hidden setting to skip image discard confirmation dialog #2141
  • Edit settings.json and add/modify the following property: "confirm_discard_image": false,
• Fixed managed install UI package status for Flux packages #2178
• Make the search strings for Flux VAE slightly more specific (should no longer match "flux2")

Download krita_ai_diffusion-1.42.1.zip

Simplified Installation

• Added separate tab with brief information about models to quickly choose a recommended set of models
• Selecting individual models is still possible, but optional now
• Added a separate "Welcome" UI for server mode selection, only shown once after installing the plugin
• Server mode selection now occupies less space in the connection settings
• Added Nova Anime XL model as a more consistent or beginner-friendly alternative to NoobAI V-pred

Other Changes

• Fixed "Image with ID not found in cache" error when queing up multiple jobs #2135
• Avoid re-sending images to the server if they're cached
• Fixed "Refine (Custom)" option being available when there is no selection #2156
• Fixed ComfyUI workflow import for paramter nodes without min/max fields #2162
• Fixed slider labels in custom workflow UI when using a higher display scale #2148
• Fixed "File > Create Copy from Current Image" not working properly with the plugin #2164
• Stop re-trying connection after the server mode has changed