ComfyUI Anima Legacy / Universal LoRA Loader

A ComfyUI compatibility loader for using LoRAs across different generations of the Anima DiT architecture.

This project does not introduce or claim authorship of Anima 3.8B, the 52-block expansion, the Qwen 0.6B + Qwen 3.5 dual-encoder approach, expanded conditioning, or Semantic Connector / Semantic Connector v2.

Those architectures and techniques originate from the original Anima 3.8B / Anima Expanded 3.8B work.

Credits / Attribution

Anima Expanded 3.8B
https://civarchive.com/models/2910974/anima-expanded-38b

Anima 3.8B
https://civarchive.com/models/2880480/anima-38b

Original Hugging Face repository
https://huggingface.co/lylogummy/Anima-3.8B

The contribution of this custom node is limited to LoRA compatibility and runtime block remapping across different Anima generations.

The original version was created to use legacy 28-layer Anima LoRAs with 40-layer Anima 2.9B / expanded Anima checkpoints.

Starting with v1.4, the loader adds compatibility for Anima 3.8B / 52-layer checkpoints and can automatically handle 28-layer, 40-layer, and 52-layer Anima LoRAs.

No LoRA file is permanently converted or modified. The compatibility remapping is performed at runtime when the LoRA is loaded in ComfyUI.

Versions

v1.4 — Universal 28L / 40L / 52L Loader

Recommended for Anima 3.8B / 52-layer checkpoints.

New node: Anima38AutoLoRALoader

Display name: Anima 3.8B Universal LoRA Loader (Auto 28L / 40L / 52L)

The loader automatically selects the required conversion path:

Legacy Anima 28L LoRA → 28L → 40L → 52L → Anima 3.8B

Anima 2.9B 40L LoRA → 40L → 52L → Anima 3.8B

Anima 3.8B 52L LoRA → Direct load → Anima 3.8B

The existing legacy loader node IDs are retained for workflow compatibility.

v1.0_28L_to_40L

Original compatibility version for:

Legacy Anima 28L LoRA → 28L → 40L → Anima 2.9B

Use this version if you only need legacy 28-layer LoRAs with Anima 2.9B / 40-layer checkpoints.

v1.4 Main Features

Automatic source LoRA detection: 28L / 40L / 52L
Legacy 28L → 40L → 52L compatibility
Native Anima 2.9B 40L → 52L compatibility
Native Anima 3.8B 52L direct loading
Optional weak LoRA propagation into newly inserted blocks
Sharp / anti-blur oriented propagation modes
Manual source architecture override
Detection report with architecture, confidence, and conversion path
Compatibility with older workflow JSON files
No permanent modification of the original LoRA
No checkpoint or LoRA weights included

Automatic LoRA Type Detection

With source_type = auto, the v1.4 loader examines the LoRA structure before loading it.

Basic Detection Rules

Highest block index 40–51 → Anima 3.8B / 52L LoRA

Highest block index 28–39 → Anima 2.9B / 40L LoRA

Only block indices 0–27 → Legacy 28L LoRA candidate

If the tensor structure alone is ambiguous, the loader also checks available safetensors metadata and filename hints.

The report shows:

Detected architecture / Detection confidence / Detected block indices / Selected conversion path / Number of converted LoRA entries

Important Note About Partial LoRAs

Some LoRAs only train a limited number of blocks.

For example, a real Anima 3.8B LoRA may contain only blocks 0–15. In that case, its block structure alone can look identical to a legacy 28-layer LoRA.

Because this cannot always be determined perfectly from tensor keys alone, v1.4 also provides a manual override:

source_type: auto / 28L / 40L / 52L

Normally, use auto.

If the report shows a low-confidence detection for a partial LoRA, manually select the correct architecture.

28 → 40 Block Mapping

0→0, 1→1, 2→3, 3→4, 4→6, 5→7, 6→9, 7→10, 8→12, 9→13, 10→15, 11→16, 12→18, 13→19, 14→20, 15→22, 16→23, 17→25, 18→26, 19→28, 20→29, 21→31, 22→32, 23→34, 24→35, 25→37, 26→38, 27→39

Inserted Anima 2.9B Blocks

2, 5, 8, 11, 14, 17, 21, 24, 27, 30, 33, 36

40 → 52 Expansion

For Anima 3.8B, twelve additional blocks are inserted into the 40-layer layout.

Inserted Anima 3.8B Blocks

3, 7, 11, 15, 19, 23, 27, 31, 35, 39, 43, 47

The original 40 Anima 2.9B blocks are placed into the remaining positions in order.

Therefore a legacy LoRA follows:

28L → 40L → 52L

instead of attempting a direct 28 → 52 conversion.

Recommended Starting Settings

Legacy 28L LoRA → Anima 3.8B 52L

Model strength: 0.75

28 → 40: copy_prev / 0.18 / attn_only

40 → 52: copy_prev / 0.12 / attn_only

Native Anima 2.9B / 40L LoRA

Model strength: 0.75

40 → 52: copy_prev / 0.08–0.16 / attn_only

Native Anima 3.8B / 52L LoRA

Source type: auto

No block expansion is performed. The LoRA is loaded directly at the selected strength_model.

Sharp / Anti-Blur Modes

copy_prev

Copies the LoRA contribution from the previous source block.

Recommended starting mode for sharper results.

sharp_blend

Approximately 70% previous block / 30% next block.

blend

Approximately 50% previous block / 50% next block.

This may produce smoother transitions, but can also make the result softer.

copy_next

Uses the next source block.

disabled

No synthetic LoRA contribution is applied to inserted blocks.

Propagation Scope

attn_only

Only attention-related LoRA components are propagated.

Recommended for the first test.

attn_and_mlp

Propagates both attention and MLP-related components.

all

Propagates all supported LoRA components.

Workflow Compatibility

The following older node IDs remain available:

Anima29LegacyLoRALoader

Anima29LegacyLoRALoaderSharp

Anima38LegacyLoRALoader

Anima38LegacyLoRALoaderSharp

v1.4 Universal Node

Anima38AutoLoRALoader

Compatibility Alias

Anima38UniversalLoRALoader

This allows older workflow JSON files to continue loading while new Anima 3.8B workflows can use automatic LoRA architecture detection.

What This Project Does — and Does Not Do

This custom node provides runtime LoRA compatibility and block remapping.

It does not create or replace the underlying Anima architecture.

In particular, this project does not claim authorship of:

Anima 3.8B
the Anima 3.8B 52-block architecture
the 40-block → 52-block Anima expansion concept
the Qwen 0.6B + Qwen 3.5 dual-encoder approach
expanded conditioning
Semantic Connector
Semantic Connector v2
the underlying Anima 3.8B model or text-conditioning architecture

For those developments, please refer to the original projects listed in the Credits / Attribution section above.

The purpose of this custom node is specifically to make LoRAs trained for different Anima block layouts more convenient to use with expanded Anima checkpoints.

Installation

Extract the package so that the node is located here:

ComfyUI/custom_nodes/ComfyUI-Anima29-LegacyLoRA/__init__.py

If an older version is already installed, replace the existing ComfyUI-Anima29-LegacyLoRA folder.

Then fully restart ComfyUI.

Important Notes

This custom node does not retrain the LoRA.

It performs compatibility remapping only when the LoRA is loaded into ComfyUI.

Results may differ from the original source architecture because newly added Anima blocks did not exist when older LoRAs were trained.

Compatibility also depends on how the LoRA was trained. Unusual LoRA key formats or highly selective partial-block LoRAs may require manual source_type selection.

Automatic detection is intended as a convenience feature, not a guarantee that every unusual or partial LoRA can be identified correctly.

This is an experimental compatibility tool intended to make LoRAs from multiple Anima generations easier to use with expanded Anima checkpoints.

Credits

The Anima 3.8B architecture, expanded architecture work, dual-encoder conditioning approach, and related conditioning techniques are credited to their original developers.