comfyui-anima-3-8B

Minimal ComfyUI nodes for the lylogummy/Anima-3.8B model, its Qwen3.5 4B text encoder, progressive-cross adapter, and timestep-aware Semantic Connector v2. All inference functionality is included in this package; comfyui-qwen35-anima is not required.

This package registers:

• AnimaQwen35Loader — Load Qwen3.5 4B (Anima)
• AnimaQwen35UnifiedPrompt — Qwen3.5 Unified Prompt (Anima)
• Anima38BV2Loader — anima.3-8B-v2
• Anima38BV2Prompt — anima.3-8B-v2 Prompt

Installation

Use a current ComfyUI build with native Anima support (comfy.ldm.anima). Update ComfyUI first if that module is missing.

Clone the repository into ComfyUI/custom_nodes:

git clone https://github.com/GumGum10/comfyui-anima-3-8B.git

Install its dependencies. From ComfyUI/custom_nodes in the Windows portable build:

..\..\python_embeded\python.exe -m pip install -r comfyui-anima-3-8B\requirements.txt

For other ComfyUI installations:

python -m pip install -r comfyui-anima-3-8B/requirements.txt

Restart ComfyUI after installation.

Model files

For the v2 release, the DiT and fixed-strength Semantic Connector v2 are one bundled checkpoint. The native Qwen3 0.6B encoder and VAE remain shared base-Anima dependencies. Subfolders are allowed.

The older separate-model path remains available as a backup:

The adapter dropdown identifies compatible .safetensors files by checkpoint metadata, so the adapter may use a different filename. Supported architectures:

• anima_progressive_qwen35_cross_adapter_v1
• anima_qwen35_quality_anchored_semantic_connector_v2

Semantic Connector v2 is evaluated inside every denoising step because its semantic extraction is timestep-aware. The extension includes the required Anima sampler patch and keeps a 64-entry conditioning-run registry so ComfyUI can safely reuse cached conditioning across repeated generations.

Workflow

For v2:

1. Load the bundle with anima.3-8B-v2.
2. Load qwen_3_06b_base.safetensors with CLIPLoader, using type stable_diffusion.
3. Load qwen35_4b.safetensors with Load Qwen3.5 4B (Anima).
4. Connect all three outputs to anima.3-8B-v2 Prompt and enter one prompt.
5. Send expanded to the sampler. Strength is fixed at the trained value of 1.0; both text encoders are released after their conditioning is produced.

For the backup separate-adapter path:

1. Load Anima-3.8B.safetensors with ComfyUI's UNETLoader.
2. Load qwen_3_06b_base.safetensors with CLIPLoader, using type stable_diffusion. This is the native_clip input.
3. Load qwen35_4b.safetensors with Load Qwen3.5 4B (Anima).
4. Connect the model and both CLIP outputs to Qwen3.5 Unified Prompt (Anima).
5. Select the adapter, enter one prompt, and set adapter_strength.
6. Send expanded to the sampler's positive conditioning. The native output is available for comparisons or fallback workflows.

A complete API-format example is included at examples/api_workflow.json.

Building the release bundle

bundle_v2.py validates that the connector's recorded native-adapter hash matches the selected DiT before writing a combined checkpoint. It refuses to overwrite an existing output:

python bundle_v2.py `
  --base path\to\Anima-base.safetensors `
  --adapter path\to\semantic_connector_v2.safetensors `
  --output path\to\Anima-3.8B-v2.safetensors

Bundled tokenizer

The official Qwen3.5 tokenizer is included in qwen35_tokenizer/; users do not need to download it separately. The node loads this bundled copy before checking other local locations and never downloads tokenizer files at runtime.

The tokenizer files are unmodified copies from Qwen/Qwen3.5-4B, revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a. See THIRD_PARTY_NOTICES.md for attribution and license information.

Verified configuration

The package has been tested with the 52-block Anima 3.8B checkpoint, qwen_3_06b_base.safetensors, qwen35_4b.safetensors, progressive-cross v1, and the 608M-parameter Semantic Connector v2. A standalone ComfyUI startup test is run with only comfyui-anima-3-8B enabled to prevent accidental dependencies on other custom-node packs.