To put it simply, this is a Lora Block Weight that can be used with ComfyUI.
●Qwen-image 1.0
This custom node allows you to specify which layers to use when applying LoRA. You can extract and apply only the layers containing the necessary information.
The CKPT used must be BF16 or FP8. GGUF and other formats cannot be used.
The setting is specified using four parameters: strength, center layer number, width, and shape. Rather than specifying strength for each layer, you specify which layers to use. You could also just line up 60 sliders.
If the base is 15 and the range is 4, the range of 15±4 (11 to 19) will be targeted.
Shape 0 uses the specified value itself, and other values ??are set to 0.
Shape 1 interpolates linearly from half the specified width beyond the specified width.
Shape 2 rounds the corners like a COS curve.
Shapes 3 and 4 change the way the slope is adjusted between the front and back.
The LoRA file reflecting the information specified here will be saved in the custom node folder under the name QwenLayered_current.safetensors. It will be overwritten each time, so please retrieve it as needed.
When in use, it consumes about 28GB of VRAM, so it may be a bit difficult even in a 24GB environment. It should probably work even with 16GB. If it's too slow, it's best to stop the preview and display the saved LoRA file.