Comfyui - Anima Experimental Fast Training Node

ComfyUI-AnimaFastTrain is an experimental ComfyUI custom node for quickly training in-memory Anima reference context tokens from one or more reference images.

Instead of creating a LoRA, checkpoint, embedding, or safetensors file, AnimaFastTrain learns temporary context tokens during the workflow and injects them into the Anima model’s cross-attention at generation time.

This makes it useful for experiments with character consistency, visual reference influence, and style transfer without writing trained weights to disk.

Main features:

- Trains lightweight reference context tokens directly inside ComfyUI

- Keeps everything in memory, no model files are saved

- Patches the Anima model during sampling

- Supports up to 3 reference images

- Adjustable token count, training steps, learning rate, dtype, and runtime strength

- Useful for quick style/identity experiments before committing to a full LoRA training run

Included nodes:

- AnimaFastTrain - Train Context Tokens

- AnimaFastTrain - Patch Model

Recommended workflow:

Checkpoint Loader -> LoRA Loader -> AnimaFastTrain Patch Model -> KSampler

Train the context tokens from a reference image, patch the final model after other model/LoRA patches, and then generate normally.

Install by cloning the repo or using comfy manager.

GitHub:

https://github.com/quinteroac/ComfyUI-AnimaFastTrain