Download krita_ai_diffusion-1.30.0.zip
Illustrious / NoobAI
Illustrious XL is a diffusion model trained extensively on (mostly anime-themed) illustration content. NoobAI XL adds further training on top. Both are based on the SDXL architecture, but weights have diverged significantly enough to require their own set of control models.
This release adds a separate "Illustrious" workload:
- NoobAI Control models (ControlNet, IP-Adapter) are used for styles using Illustrious architecture
- Managed install offers checkpoint and control models as download option
- Built-in style "Anime (Noob AI XL)" with recommended settings
- Supports both epsilon and v-prediction
- Seamless fill is disabled by default (there is no inpaint model so far)
Important
Most Illustrious checkpoints are detected as "SDXL" because they cannot be distinguished. You can configure
the correct architecture manually in Style settings > Checkpoint configuration (advanced) > Diffusion Architecture
The v-prediction models are detected automatically (no configuration needed).
User Guide
The documentation that was previously on the Wiki has moved to a new place: docs.interstice.cloud
There is a lot of new content too:
Something that was frequently requested but not easily possible on the Wiki was a full list of models recognized by the plugin: Model Database
Regional LoRA
LoRA which are added to text prompts of regions are now applied to that region only. Previously LoRA could only be applied to the entire image.
The initial implementation either requires a lot of VRAM or runs very slowly! Only use it if you really need it.
Other Changes
- Improved compatibility with custom tag lists (for auto-completion) by @BetaDoggo
- Lowered the default minimum resolution for SDXL to avoid resizing where it isn't required #1470
- Style settings are now locked when a built-in style is selected (with an option to make a copy to edit it)
- Made CFG rescale for v-prediction models configurable in style.json files (no UI for now)
- Reorganized model packages for the managed server install to be grouped by base model
- Upgraded managed install to PyTorch 2.5.1
- Fixed an assertion when there is a node with a "type" parameter in a custom workflow #1473
- Fixed mismatch between image and mask size when using resolution multiplier #1532
- Fixed an infinite recursion in custom graph workspace which could lead to hangs and broken UI when opening documents
- Fixed issues when active layer change is emitted before the layer is part of the node tree #1475
- Fixed empty error box being shown after a connection error is resolved
- Fixed long layer names stretching the animation workspace UI
- Moved model database to a separate JSON file (resource paths are still in resources.py)
- Added style models (eg. Flux depth/canny) to extra_model_paths