Anima LoRA Rename (Diffusers to Non-Diffusers)

Anima - the 2B-parameter anime-focused model from Circlestone Labs built on NVIDIA's Cosmos-Predict2 - has an unusually easy, unusually popular training pipeline, and that's exactly what creates the problem this node fixes. People train Anima LoRAs through more than one tool: Anima TrainFlow (a Gradio front end over modified sd-scripts) and the Anima Standalone Trainer are both in active use, and they don't necessarily export tensor keys in the same convention. Some pipelines output LoRA weights in HuggingFace's "diffusers" key-naming style; ComfyUI's native LoRA loading (and most Anima workflows built around it) expects the older, non-diffusers naming the rest of the local-diffusion ecosystem standardized on. Load a diffusers-format LoRA into a workflow expecting the other convention and, at best, it does nothing - no error, just silence, because the keys don't match anything the loader recognizes.

What it does

One job, done specifically: it loads a diffusers-format Anima LoRA, reverses its key names to the non-diffusers convention, and saves the result as a new safetensors file. The node's own description is exactly that direct - "loads a Diffusers Anima LoRA, reverses its key names to the non-diffusers format, and saves it as a new safetensors file." This is not a general-purpose rename tool; it's a fixed mapping built for one specific format mismatch, on one specific model family.

The inputs and outputs that matter

Two fields. lora_name picks the source LoRA from a dropdown. output_filename (default anima_lora_non_diffusers) names the result. That's it - no manual key-mapping required, because the diffusers-to-non-diffusers translation is exactly what the node already knows how to do. Output is a single output_path string, and it's a save node (is_output_node: true) - running it writes the converted file to disk.

Installing it

ComfyUI Manager: search Model Utility Toolkit, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-ModelUtils

Restart ComfyUI. No downloads needed beyond the LoRA file itself.

Where people get burned

The tell that you need this node at all is a LoRA that loads without any error but visibly does nothing at inference - no style shift, no likeness, nothing, even at strength 1. On Anima specifically, before assuming the LoRA is broken or badly trained, check whether it came from a pipeline that exports diffusers-format keys. If you're not sure, LoRAMetaKeys (elsewhere in this pack) will show you the actual tensor names - diffusers-style keys look visibly different (dotted, HuggingFace-module-path style) from the non-diffusers convention ComfyUI expects.

Run this node on a LoRA that's already in the correct non-diffusers format and you'll get a file with wrong or garbled key names out the other end - this is a one-directional fix for one specific mismatch, not something to run reflexively on every Anima LoRA you download. And it's worth knowing the underlying compatibility question hasn't fully settled even once naming is fixed: the community's own testing on whether Anima LoRAs trained against one preview version carry cleanly to a later base release is mixed, some report near-identical results, others report detail loss - so a rename fixing the format doesn't guarantee the LoRA behaves identically to how it did on the checkpoint version it was trained against.