awdqwdasdg/Anima-Lightning-Comfyui

Disclaimer

This conversion dropped four tensors that did not match the original implementation, so quality may not be as intended.

This does end up reducing the quality, and will have to be solved with a custom node (wip).

Custom node (still experimental; read below): https://github.com/awdqwdasdg/ComfyUI-Anima-Lightning

Usage

A: Download an entire model, add the custom node* right after the Load Diffusion Model node, and run at 4 steps.

B: Download the LoRA, load whatever base model** at your preferred precision, add the custom node* in somewhere, and run at 4 steps

C: Download the LoRA, load whatever base model** at your preferred precision, add the custom node* in somewhere, mix the lightning lora with an extracted turbo-1.1 lora***, and run at 4 steps.

• * The custom node (still experimental) does seem to improve things in some cases, but in others it degrades the image. The outputs look passable without it, so mess around with/without it and come to your own conclusions.
• ** Tested with Anima-2.9B, and it does work but you will need some other custom node to remap the layers to load properly.
• *** Mixing the distill loras does work to some extent, but also introduces a slew of issues including some uncommon, but egregious, anatomy issues.
• The custom node will automatically download anima_tdm_k_branch.safetensors within it's own folder.
• You can use NegPIP (https://github.com/pamparamm/ComfyUI-ppm) to use negative tags at CFG 1 by adding negative weighted tags to your positive prompt (eg; '(loli:-1)' within the positive prompt).

Notes

The orig files are based on the original uploaded versions of the models. Files not tagged are the current, and recommended, versions of the model.

Info

Converted from https://huggingface.co/aina-tech/Anima-Lightning using a script by GLM-5.3-Flash

int8 model created with https://github.com/silveroxides/convert_to_quant

Lora extracted from anima-base-v1.0 using the 'LoraExtractKJ' node from https://github.com/kijai/ComfyUI-KJNodes