ComfyUI-Anima-2.9B-loraPatch

ComfyUI-Anima-2.9B-loraPatch

English | 日本語

A load-time patch that applies LoRAs trained on anima-base-v1.0 / anima-preview (28 blocks) to Anima-2.9B (40 blocks) with the correct layer correspondence. No workflow nodes are added.

Why it is needed

Anima-2.9B is a block expansion (depth upscale) of anima-base-v1.0: 12 new blocks were inserted between the original 28. So base block j no longer lives at index j:

| base | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| | 2.9B | 0 | 1 | 3 | 4 | 6 | 7 | 9 | 10 | 12 | 13 | 15 | 16 | 18 | 19 |

| base | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---| | 2.9B | 20 | 22 | 23 | 25 | 26 | 28 | 29 | 31 | 32 | 34 | 35 | 37 | 38 | 39 |

(Measured by direct weight comparison; the 28 mapped blocks are bit-identical to the base model.)

Without this patch, a 28-block LoRA applied to 2.9B hits the wrong layers silently — blocks.2 through blocks.27 all exist in the 40-block model, so no "lora key not loaded" warning is ever emitted. Only blocks 0 and 1 land correctly. This manifests as "the LoRA doesn't work" or "raising strength degrades the image instead".

Installation

1. Put this folder into ComfyUI/custom_nodes/
2. Restart ComfyUI
3. Apply LoRAs with the standard LoRA Loader nodes as usual; remapping happens automatically only when the conditions below are met

No extra pip installs, no workflow changes.

Note: nothing appears in the node list (Add Node menu). This patch takes effect automatically at startup; the log lines below are the only way to confirm it is active.

To uninstall, delete the folder or append .disabled to its name.

Requirements

Anima-2.9B has to load with all 40 blocks first, which needs one of:

• ComfyUI-Anima-2.9B-blocksPatch or the original ComfyUI-Anima-2.9B — required on any released ComfyUI build (v0.32.0 or earlier, Desktop, portable)
• or a ComfyUI checkout that includes PR #15555 (merged 2026-08-12, master only so far), which fixes the block count natively

This patch cannot help a model that was already truncated to 28 blocks — the remap needs the real 40-block layout to target.

How it works

Wraps comfy.lora.load_lora and rewrites the block indices in the key_map values (the LoRA-key → model-key table). Both the kohya format (lora_unet_blocks_5_...) and the peft / AI-Toolkit format (diffusion_model.blocks.5....) resolve through the same key_map, so one rewrite fixes both.

Remapping is applied only when all of the following hold:

• the model is Anima (its key_map references diffusion_model.llm_adapter.*)
• the model has exactly blocks 0..39 (the layout this mapping was measured on)
• the LoRA covers exactly blocks 0..27

Anything else — a 28-block model, a native 40-block LoRA, a non-Anima model, an unknown layout — passes through unchanged. Text-encoder keys (lora_te_*) and llm_adapter keys are never touched.

Verifying it works

Startup log:

[Anima 2.9B LoRA Patch] installed (base block j -> 2.9B block [0, 1, 3, 4, 6, 7]...). Set ANIMA_LORA_REMAP=0 to disable.

When a 28-block LoRA is applied to 2.9B:

[Anima 2.9B LoRA Patch] 28-block LoRA on 40-block Anima-2.9B: remapped 1040 target keys to the expanded block layout.

(The count varies per LoRA.)

Environment variables

| Variable | Default | Meaning | |---|---|---| | ANIMA_LORA_REMAP | 1 | Set to 0 to disable remapping |

Caveats

• Layer correspondence becomes correct, but the residual stream now passes through 12 new blocks and a retrained llm_adapter, so the LoRA will not behave identically to how it does on the base model. Start around strength 0.5–0.7
• A future LoRA trained natively on 2.9B that deliberately targets only the lower 28 blocks would be misdetected (an unusual configuration; disable via the environment variable above if it ever happens)

See also

storyAura/Anima2.9B-Lora-weight-conversion takes the alternative approach of converting the LoRA files ahead of time (its insertion indices match this patch). That is useful outside ComfyUI or when distributing converted files. Running both is safe — a converted LoRA no longer covers exactly blocks 0..27, so this patch passes it through untouched.

Development notes

This project was vibe-coded with Claude Opus 5 (Anthropic) in a Claude Code session: the investigation, the weight measurements and the code were all produced by the model under human direction.

The block mapping is not guesswork. It was derived by comparing the actual .safetensors weights of Anima-2.9B-preview-v1 against anima-base-v1.0 tensor by tensor, which is how the 28 bit-identical blocks and the 12 insertion points were identified. The gating logic was then checked against 50 real Anima LoRAs: all 50 resolved every block key to the correct expanded index, with no false positives on 28-block models or on synthetic native 40-block LoRAs.

That said, please read the code before trusting it — it is short and single-purpose on purpose. Bug reports and corrections are welcome in the issue tracker.

License

GPL-3.0, matching ComfyUI itself, since this patch hooks ComfyUI internals directly.