ComfyUI-Anima-3.8B-LoRA-Bridge

ComfyUI Anima 3.8B LoRA Bridge

日本語

ComfyUI nodes for applying original 28-block Anima Base LoRAs and 40-block Anima 2.9B LoRAs to the matching inherited layers in the 52-block Anima 3.8B model.

Why this is needed

Anima 3.8B was produced through two interleaved layer expansions:

Anima Base: 28 blocks -> Anima 2.9B: 40 blocks -> Anima 3.8B: 52 blocks

The normal ComfyUI LoRA loader matches block numbers directly. After the insertions, a Base LoRA key such as blocks.2 no longer belongs on block 2 in the expanded model. This node remaps every supported key through both expansion layouts before handing the result to ComfyUI's standard LoRA patcher.

The safe behavior is fixed and has no experimental switches:

• LoRA patches are applied only to inherited DiT blocks.
• The 24 expansion blocks remain untouched.
• Anima 3.8B's changed internal text adapter remains untouched for Base/2.9B LoRAs.
• Source layout is detected automatically.

The inherited Base mapping was checked against the official Anima Base v1.0 and Anima 3.8B checkpoints: all 560 tensors in the 28 inherited DiT blocks were bit-identical at their mapped destinations.

Installation

From ComfyUI/custom_nodes:

git clone https://github.com/Lakeside529/ComfyUI-Anima-3.8B-LoRA-Bridge.git

Restart ComfyUI. No additional Python packages are required.

Nodes

Load Anima LoRA on 3.8B (Model Only)

Use this for standard style, character, Turbo, RL, and other model-only Anima LoRAs. It has the same three practical inputs as ComfyUI's model-only loader:

• model
• lora_name
• strength_model

Load Anima LoRA on 3.8B

Use this variant when a LoRA also contains text-encoder weights. It accepts and returns both MODEL and CLIP.

Workflow placement

Place the bridge after UNETLoader and before both the sampler and any node that reads the model to build conditioning:

UNETLoader -> Anima 3.8B LoRA Bridge -> KSampler
                              |
                              +-------> Qwen3.5 Unified Prompt (model input)

When using the Qwen3.5 workflow, install the companion comfyui-anima-3-8B nodes and connect the bridge's final MODEL output to both the Unified Prompt node and the sampler. An API-format example is provided in examples/api_workflow.json.

Supported LoRA key formats

• ComfyUI generic keys such as diffusion_model.blocks.0...lora_A.weight
• Kohya-style keys such as lora_unet_blocks_0_...
• Full 28-block Anima Base and 40-block Anima 2.9B layouts
• Native 52-block LoRAs pass through unchanged

Technical sources

• Anima 2.9B expansion manifest
• Anima 3.8B model card
• The 40-to-52 insertion positions are also stored in the Anima 3.8B checkpoint metadata.

License

The code in this repository is MIT licensed. Model files and LoRAs are not included and remain subject to their respective licenses.