Comfyui-Anima-IPadapter Clones Character Looks Without Training

The Comfyui-Anima-IPadapter custom node brings IP-Adapter support to the Anima DiT model inside ComfyUI. It lets users inject reference image features directly into the generation process using decoupled cross-attention. This enables consistent character appearance across multiple images without fine-tuning the base model.

Wenaka2004 released the open-source node under the circlestone-labs GitHub organization, making it simple to replicate a character’s look by providing a reference photo. The tool uses Qwen3-VL embeddings to extract visual features, so you don’t need to train a separate adapter for each style. It is designed for ComfyUI users who want fast, local character-consistent generation without complex setups.

Character-consistent generation for Anima

Privacy-conscious professionals working with local image generation will appreciate that everything runs on their own hardware. Small creative studios can use it to maintain character consistency across storyboards or concept art without cloud costs. Prosumer GPU owners with at least 8GB of VRAM can run the on-the-fly encoding option, though pre-computed embeddings work with even less memory.

Developer notes and known limitations

The repository has no official release yet; it is distributed as source code to clone directly into ComfyUI’s custom_nodes folder. Option B for real‑time embedding requires loading the Qwen3-VL-Embedding-2B model, which adds roughly 4 GB of extra VRAM — a potential obstacle for lower‑end GPUs. The project is licensed under Apache-2.0, allowing both personal and commercial use without restrictions.

"Injects reference image features into Anima's DiT via decoupled cross-attention, enabling character-consistent image generation." — Source: GitHub