ComfyUI-AnimaFastTrain Injects Ephemeral Visual Identity

    
        By vramkickedin    
     | 
    
            June 2, 2026 at 11:10 am        
    
     | 
    
        2 min read

ComfyUI-AnimaFastTrain is a new set of custom nodes that lets you train temporary reference tokens for Anima models directly inside ComfyUI. These nodes work completely in memory without saving any files to disk. They enable fast experiments with character consistency, visual reference influence, and style transfer.

Developer quinteroac released the project as an experimental tool for the AI open-source community. It provides a way to inject reference image traits into generated images during sampling. The nodes are designed for quick, private experimentation without permanent storage.

In-memory training for quick experiments

Key Features

Trains context tokens directly from reference images.
Uses in-memory processing with no disk saves.
Two custom nodes for training and patching.
Configurable training steps and learning rate.
Per-block context injection across 28 blocks.
Control runtime strength with reference context scale.
Install via ComfyUI Manager or git clone.

This tool is ideal for prosumer GPU owners who want to experiment with character consistency without complex setups. Privacy-conscious professionals will appreciate that all processing stays local and leaves no file traces. Small agencies can use it to rapidly test visual style transfer and reference-based image generation.

Developer notes and limitations

The developer notes that the nodes are intentionally memory-only, meaning they do not save any weight caches or safetensors files. Because the patch is applied just before sampling and restored after, you must connect it in a specific graph order after any LoRA nodes. Future updates may expand functionality, but currently this remains an experimental release for tinkering.