AnimaRefLora

Reference-based character identity transfer for Anima (the 2B anime DiT by CircleStone Labs / Comfy Org): give it one reference image of a character, and it generates new images of that character following your prompt — new poses, outfits, and scenes — without per-character LoRA training.

• Code, training scripts, and ComfyUI plugin: github.com/crazysheep924/AnimaRefLora
• Technical write-up: see the blog/ directory in the GitHub repo.

All example characters used in the project are AI-generated original designs.

Files

anima-reflora/idinject_500k.animaref.safetensors   ← the release checkpoint (single-file bundle)

The .animaref.safetensors bundle is adapter-only (it is not a merged model). One file packs everything inference needs:

Component	Contents
lora.*	LoKr adapter for the Anima DiT (dim 512 / alpha 512)
ref_conditioner.*	reference/identity conditioning modules
crepa_projector.*	CREPA feature projector
metadata	feature config + RoPE reference-position layout (read automatically by the loader)

Requirements

You also need the base models, from circlestone-labs/Anima:

• anima-base-v1.0.safetensors (DiT)
• the Qwen3-0.6B text encoder (file must be named model.safetensors — see plugin README)
• qwen_image_vae.safetensors (VAE)

Verified on Anima Base v1.0 only. Support for the community layer-expanded Anima variants (2.9B/3B-class) has not been confirmed.

Usage (ComfyUI)

Install the plugin from the GitHub repo (comfyui/ComfyUI-AnimaRefLora), then place the models as:

ComfyUI/models/diffusion_models/anima-base-v1.0.safetensors
ComfyUI/models/text_encoders/model.safetensors
ComfyUI/models/vae/qwen_image_vae.safetensors
ComfyUI/models/anima_reflora/idinject_500k.animaref.safetensors

Node graph: Anima RefLora Loader → Anima Ref Encode (reference image) → Anima RefLora Sampler. An optional Anima Extra LoRA node accepts regular Anima style LoRAs (sd-scripts native and diffusers/PEFT keys, DoRA included).

Usage (CLI)

The GitHub repo's anima_reflora.local_ref_ab_infer module provides the same inference path used for the project's REF evaluations; see the repo README for setup and commands.

Training summary

• Trained on image pairs derived from Danbooru, with a T=3 frame contract (head-crop reference / full reference / noised target) and a disjoint RoPE layout separating reference and target positions.
• The release checkpoint corresponds to ~500K training steps, including anti-copy objectives (diff-weighted loss, tag-level caption dropout) and identity-accessory injection; full recipes and provenance are documented in the GitHub repo's README.

License

The adapter weights in this repository are released under Apache-2.0. Note that the Anima base model itself is distributed under the CircleStone Labs Non-Commercial License (and the NVIDIA Open Model License as a Cosmos-Predict2 derivative) — using these adapters together with the base model is subject to those base-model terms.