Releases: Acly/krita-ai-diffusion
Version 1.7.1
Download krita_ai_diffusion-1.7.1.zip
Changes
- LCM Lora will now be found when downloaded via ComfyUI Manager
- Support safetensors version of the clip vision model (models/clip_vision/SD1.5/model.safetensors)
- Fixed preview toggle (single click) and apply (double click) in the results thumbnail viewer
- Fixed issue with text disappearing in 1-line prompt box
- Improved model detection for external ComfyUI installs and added diagnostics in log file (more info)
Version 1.7.0
Download krita_ai_diffusion-1.7.0.zip
Server Upgrades
The features of this release require new server versions and extensions. If you used an old version of the plugin, you will have the opportunity to upgrade the server automatically after installing the latest plugin version. (The plugin itself is still updated by extracing & overwriting.)
If you maintain your own ComfyUI installation make sure to update to the latest version! See here for model downloads.
The upgrade will remove & reinstall the server code. Models will be migrated and do not have to be re-downloaded.
AI Live Painting with interactive feedback
This release implements the LCM sampler, which requires very few steps and is therefore much faster - albeit at a loss of quality. If you happen to own a high-end GPU (RTX 4070+) this allows for generating images in less than a second. You can select the sampler in the style settings, but more excitingly, there is also a new workspace/tab which gives you a "live" update while you paint.
The AI generated version on the right automatically reflects any changes you make in the canvas. This mode is compatible with all styles, uses LCM sampler by default, and also supports control layers! I'm not sure how useful it is yet, but it's a lot of fun :)
Other Changes
- Use fixed versions for all server packages instead of latest, to avoid issues when they are updated and become incompatible
- Fixed changing style in Upscale workspace having no effect
- Log supported SD versions for better diagnostics
- CFG scale now supports decimals (0.1 steps)
- Separate sampler settings for default (quality) image generation, and live (fast) image generation
- Optimized PNG conversion to avoid (long) UI hang when using large resolutions
Version 1.6.0
Download krita_ai_diffusion-1.6.0.zip
Improved SD XL support
Stable Diffusion XL has been quietly supported for a while, but because it still lacks good inpaint/outpaint capabilities it wasn't integrated as tightly. This release provides installation support, filtering, and better documentation for SD XL. You can read more about SD 1.5 vs SD XL here.
Server installation packages were condensed, and now offer two "workloads": SD 1.5 and SD XL. You need at least one! Recommended checkpoints and control models have also been added for XL.
Style selection now includes default styles for XL, and filters them depending on the installed workloads and checkpoints. That is, you won't see styles if you can't run them.
The settings still show all styles, and show warnings if the respective workload hasn't been installed.
Other Changes
- Fixed creating/deleting styles triggering an assertion
- Automatically hide generation preview when upscaling results are available
- Fixed download errors being silently ignored
- Allow using client with only SD XL installed (or only SD 1.5)
- Fixed issue when inpainting (Kernel size can't be greater than actual input size, #36)
- Fixed validation error when using tiny selection area
- Show SD base model version icons next to checkpoints and styles
- Changed python package installation to call python instead of pip script (#37)
- Fixed bad request / validation error due to changes in 3rd party custom node (#39, #41)
- Added default styles for SD XL
Version 1.5.0
Download krita_ai_diffusion-1.5.0.zip
If you are updating from a previous release: additional custom nodes and models are required. They can be downloaded automatically via the built-in installer.
Upscaling
Now there is a dedicated UI to upscale images. You can switch with the new button in the top-left:
It allows to upscale the current projection (merged visible layers) by up to factor 4x.
- Upscale model can be chosen explicitly
- There are 3 recommended models: Default, Quality, Sharp
- Custom models are also supported (found in
ComfyUI/models/upscale_models)
- The "Refine" option will run diffusion on the upscaled image to improve quality
- 30% strength will stick very closely to the low resolution image
- 40-50% strength will allow freedom for additional details
- Generation runs on tiles, so high resolutions (4k, 8k, more...) are possible without requiring more GPU memory
Upscaling makes use of a new control mode "Blur" (aka. ControlNet Tile). It takes blurry image as input (eg. from bilinear resize) and sharpens and adds detail. You can use it in the image generation UI to selectively upscale/refine areas of the image.
Inpaint Improvements
Trying to add new objects to an existing image in an area has felt unintuitive in some cases. The reason is that a lot of context from the existing image is needed to make sure new content "fits in". But this leads to the text prompt having less impact, and sometimes it felt like it's ignored.
- Method 1: Select area, enter text, 100% strength
- Method 2: Select area, enter text, paint a very crude version of the new object, 80% strength
Method 2 has always been quite reliable and predictable! But it's not as intuitive, so in this version Method 1 is improved.
Let's look at an example. We have some nature photo, and it's clearly lacking a bear.
| Original | Mask | Prefill for Method 2 |
|---|---|---|
 |
 |
 |
Method 1, OLD behavior - only 1 out of 4 images has a bear :(
Method 2 (unchanged) - note that the prefill also gives some control over shape
Method 1, NEW behavior - bears are now very common indeed :)
Other Changes
- Added option to specify negative text prompt in the generation UI (must be activated in interface settings)
- Added option to change number of text lines for the prompt editor (interface settings)
- Made server installation more robust after crash/abort
- Added IP-adapter support for SDXL, requires
ip-adapter_sdxl_vit-h.pthinComfyUI/custom_nodes/ComfyUI_IPAdapter_plus/models - Support
{prompt}placeholder in style prompts (also updated default styles to use this) - Improved outpainting scenario by excluding image border areas from the IP-adapter reference
- Increased websocket ping timeout for large image transfers when using a remote (internet) server
Version 1.4.0
Download krita_ai_diffusion-1.4.0.zip
Pose Vector Editor
Vector layers can now be used for controlling pose. This makes it easier to do adjustments or create a pose from scratch.
Use this button to automatically generate vector pose from the current image.
Use this button to add a new pose to the selected vector layer.
Editing pose joints will automatically update connections. Moving, scaling and rotating is supported. Keep in mind that Krita's vector layers are very general and allow a lot of additional operations which may not play nice with this feature though.
PoseEditor.webm
Changes
- Added pose editing via vector layers
- Selection area padding is now configurable in diffusion settings.
This controls how much context outside of the selection mask the AI is able to see - Match only the filename when searching for ControlNet models. This allows to organize them in arbitrary subfolders
- Made checkpoint classification more robust (SD 1.5 vs SDXL) - requires custom node update
- Filter out inpaint and refiner checkpoints. Inpaint models currently don't work properly, but also shouldn't be needed
- Automatically exclude control layers from image input when using less than 100% strength
Version 1.3.0
Download krita_ai_diffusion-1.3.0.zip
Reference Images
You can now add arbitrary images from layers as input for generation. It will generate something similar to what's in the image. This works as a replacement or in addition to text prompts.
Here I ask for a cat, and also provide an image of some pegonias. So the results all show a cat and pegonias (or some red flowers at least!).
Multiple images can be added. They don't have to match in size, but should preferrably be square(ish).
Changes
- Added IP-adapter as explicit control mode
- Batch multiple control images
- Fixed platform check for MacOS (there are probably other issues though)
- Fixed installation after PyTorch 2.1.0 release
- Increased value range for LoRA (sometimes used as a "slider" value from -400% to +400%)
- Fixed connection to ComfyUI server behind ngrok
Version 1.2.1
Version 1.2.0
Download krita_ai_diffusion-1.2.0.zip
Support more GPUs and Cloud
This release provides some more options for users without NVIDIA GPU: Windows users can now use non-NVIDIA GPUs via DirectML, and there is a step-by-step guide on how to rent a Cloud GPU which the plugin can connect to.
Updating
For users who update from an older release, it's generally fine to copy the new files and overwrite the old ones. Afterwards, if you installed the server via the plugin, use the Install button to upgrade it.
If you're using your own ComfyUI installation, IPAdapter-ComfyUI was replaced with ComfyUI_IPAdapter_plus.
Changes
- Added actions to cancel queued jobs
- The server is now available as docker image for easy deployment
- Added refresh & open folder buttons to make it easier to install custom checkpoints and lora
- Added DirectML as backend option
- Switch to IPAdapter_plus extension, fixes CPU and DirectML backends when doing 100% strength inpaint
- Improved handling of sporadic websocket disconnect
- Automatically switch to non-inpaint workflows if selection covers entire document
- Added some alternative names for SDXL controlnet files
Version 1.1.0
Download krita_ai_diffusion-1.1.0.zip
More Control
More options to control generation with image layers have been added: Soft Edge, Canny Edge, Depth map, Normal map, Pose from
OpenPose skeletons, Segmentation maps
Use this button to generate a control layer from the current image. Works with all modes. For example if you like the contours of the current image but would like to change colors/shading, this allows you to generate matching line art and use it as control layer for subsequent generation. Note: might be slow on first use, but quite fast afterwards.
You can download required models from the Server Configuration settings.
Other Changes
- Installer now uses Python venv on Linux. This avoids messing with the system Python and some distros require it.
- Improved error reporting for missing resources
- Added option to override VAE in styles
- Line and edge control layers use white background with black lines
- Error message when using non 8-bit RGBA color models
- Added https and wss support by @gergelyszaz in #6
- Control layers with alpha channel work better
- Can now configure ComfyUI command line arguments in settings.json
- Use high quality upscaling for scaling factors >150% in all workflows
- Fixed 'QNetworkReply' has no attribute 'HttpStatusCodeAttribute'
- Added new style and delete style buttons by @gergelyszaz in #9
- Selection grow and feather are now configurable
- Support group layers and nested layers as control images
