ComfyUI-Anima-3.8B-LoRA-Bridge Introduction
ComfyUI-Anima-3.8B-LoRA-Bridge lets you use LoRAs made for earlier Anima models with Anima 3.8B in ComfyUI. A LoRA is a small set of adjustments that can add a style, character, or other behavior to a model without replacing the main model.
Anima 3.8B has more internal layers than Anima Base or Anima 2.9B. Because of that, ComfyUI’s normal LoRA loader can match an older LoRA’s layer numbers to the wrong places in Anima 3.8B. The bridge translates those layer numbers before ComfyUI applies the LoRA, so supported older LoRAs reach the corresponding inherited layers.
The extension provides two loading nodes: one for model-only LoRAs and one for LoRAs that also include text-encoder weights. It detects the source layout automatically; you do not need to choose a conversion mode.
How ComfyUI-Anima-3.8B-LoRA-Bridge Works
Anima’s layer structure grew in two stages:
Anima Base: 28 blocks → Anima 2.9B: 40 blocks → Anima 3.8B: 52 blocks
Think of the blocks as pages in a book. If new pages are inserted, a note that originally pointed to page 2 may need a different page number in the expanded edition. Older Anima LoRAs have a similar problem: their adjustments refer to block numbers from the model they were made for.
The bridge identifies whether a LoRA follows the 28-block Anima Base layout or the 40-block Anima 2.9B layout. It then maps supported LoRA keys to the matching inherited blocks in Anima 3.8B and passes the result to ComfyUI’s standard LoRA patcher. LoRAs already using the native 52-block layout pass through unchanged.
This mapping is deliberately limited:
- LoRA patches are applied only to inherited DiT blocks.
- The 24 blocks added during the two expansions remain untouched.
- The changed internal text adapter in Anima 3.8B remains untouched by Base and 2.9B LoRAs.
The author checked the inherited Anima Base mapping 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. This supports the block mapping; it does not guarantee that every LoRA will produce exactly the same artistic result on different Anima models.
ComfyUI-Anima-3.8B-LoRA-Bridge Features
Choose the loader that matches your LoRA
Load Anima LoRA on 3.8B (Model Only) is for LoRAs that adjust only the image model, including standard style, character, Turbo, RL, and other model-only Anima LoRAs. Its practical inputs mirror ComfyUI’s model-only LoRA loader:
model: the Anima 3.8B model to adjust.lora_name: the LoRA you want to use.strength_model: how strongly to apply its model adjustments. Changing this value changes the LoRA’s influence on the result.
Load Anima LoRA on 3.8B is the variant for a LoRA that also contains text-encoder weights. It accepts and returns both MODEL and CLIP, so both can continue through your workflow.
Automatic support for Anima layouts and LoRA keys
The bridge supports full 28-block Anima Base and 40-block Anima 2.9B layouts. It recognizes ComfyUI generic keys, such as diffusion_model.blocks.0...lora_A.weight, and Kohya-style keys, such as lora_unet_blocks_0_.... Native 52-block Anima 3.8B LoRAs pass through without block remapping.
You do not need to identify those key formats by eye. Select your LoRA in the node and let the bridge detect its source layout.
Place it before nodes that use the model
Connect the bridge after UNETLoader and before the sampler. If another node reads the model to build conditioning, give that node the bridge’s final MODEL output too:
UNETLoader → Anima 3.8B LoRA Bridge → KSampler
|
+──────→ Qwen3.5 Unified Prompt (model input)
For a Qwen3.5 workflow, use the companion comfyui-anima-3-8B nodes and connect the bridge’s final MODEL output to both the Unified Prompt node and the sampler. See the API-format example workflow for a concrete connection example.
ComfyUI-Anima-3.8B-LoRA-Bridge Models
The bridge does not include model files or LoRAs, and its nodes are not separate image-generation models. The model names below describe the LoRA layouts it can handle:
| LoRA source layout | Blocks | What the bridge does |
|---|---|---|
| Anima Base | 28 | Maps supported LoRA keys to the corresponding inherited blocks in Anima 3.8B. |
| Anima 2.9B | 40 | Maps supported LoRA keys from its expanded layout to the corresponding inherited blocks in Anima 3.8B. |
| Anima 3.8B | 52 | Passes native LoRA keys through unchanged. |
For example, if you have a character LoRA made for Anima Base, use the bridge to apply its supported adjustments to the matching inherited layers of Anima 3.8B. If your LoRA was made for Anima 3.8B itself, the bridge does not need to renumber its blocks.
Troubleshooting ComfyUI-Anima-3.8B-LoRA-Bridge
My Qwen3.5 workflow does not seem to use the bridged model
- Check that the bridge is connected after
UNETLoader. - Connect the bridge’s final
MODELoutput to the sampler. - Connect that same final
MODELoutput to the Qwen3.5 Unified Prompt node’s model input. Do not leave that node connected directly to the unmodified loader output. - Compare your connections with the example workflow.
Which loader node should I choose?
Use Load Anima LoRA on 3.8B (Model Only) for a model-only LoRA. If the LoRA also contains text-encoder weights, use Load Anima LoRA on 3.8B and continue the workflow with its MODEL and CLIP outputs.
Do I need to select “Base,” “2.9B,” or “3.8B” manually?
No. Source-layout detection is automatic. A native 52-block LoRA passes through without block remapping.
Will an older LoRA change the new Anima 3.8B blocks?
No. The bridge applies older LoRA patches only to inherited DiT blocks. The 24 expansion blocks and Anima 3.8B’s changed internal text adapter remain untouched by Base and 2.9B LoRAs.
Will the result look identical to using the LoRA on its original model?
Not necessarily. The bridge maps supported adjustments to their matching inherited layers, but Anima 3.8B also contains additional blocks. The author’s tensor check verifies the inherited Base block mapping, not identical final images across models.
Learn More about ComfyUI-Anima-3.8B-LoRA-Bridge
- Anima 3.8B model card: background on the model used with the bridge.
- Anima 2.9B expansion manifest: the recorded expansion layout behind part of the mapping.
- comfyui-anima-3-8B: companion nodes for the Qwen3.5 workflow described above.
- Example API workflow: a reference for connecting the bridge in ComfyUI.
The extension’s code is MIT licensed. Model files and LoRAs are not included and remain subject to their own licenses.
