https://github.com/blepping/ComfyUI-ApplyResAdapterUnet
ComfyUI node to apply the ResAdapter Unet patch for SD1.5 models
https://github.com/blepping/ComfyUI-ApplyResAdapterUnet
Last synced: 4 months ago
JSON representation
ComfyUI node to apply the ResAdapter Unet patch for SD1.5 models
- Host: GitHub
- URL: https://github.com/blepping/ComfyUI-ApplyResAdapterUnet
- Owner: blepping
- License: gpl-3.0
- Created: 2024-03-26T19:03:25.000Z (over 1 year ago)
- Default Branch: main
- Last Pushed: 2024-03-26T19:29:48.000Z (over 1 year ago)
- Last Synced: 2025-02-26T16:44:19.661Z (4 months ago)
- Language: Python
- Size: 6.2 MB
- Stars: 31
- Watchers: 1
- Forks: 1
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
- awesome-comfyui - **ComfyUI-ApplyResAdapterUnet**
README
# ComfyUI-ApplyResAdapterUnet
ComfyUI node to apply the ResAdapter Unet patch for SD1.5 models.
See https://github.com/bytedance/res-adapter for explanation and link to download the LoRA and unet patch.
Note: I am not affiliated with ResAdapter. This is an (experimental) personal project. I believe it works correctly, but no guarantees!
## Usage
### SDXL
For SDXL, you only need the LoRA (as far as I know) so a dedicated node is unnecessary: just load the LoRA as usual. You don't need this repo.
### SD 1.5
* Put the `resolution_normalization.safetensors` model in `models/unet`
* Patch the model with the `ApplyResAdapterUnet` node, load the `resolution_lora.safetensors` LoRA normally.
You can experiment with different unet and LoRA strengths.
I haven't tested it extensively, but at resolutions above 1024x1024 using full strength doesn't seem to work well (and in fact may be worse than nothing).
It's also possible to combine ResAdapter with other techniques such as Kohya Deep Shrink (AKA `PatchModelAddDownScale`).
## Example Workflow
Workflow with included ComfyUI metadata:
(Simple demonstration, I made no effort to get a pretty picture.)