xFutureTechx

/

2026_Loras

Like 2

Follow

Future Tech 8

OH OH SECODN COMMENT LMAO THE CAMERON X'VOOR ANIMA ONE DOES NOT WORK XD I REPEAT: FOX BOY ON ANIMA DOES NOT WORK XD i dunno why but it made sense cause it was only 90 mb XD

Hmm, in a nutshell, "that's a bug" 😅. Or rather, different tools still expect different keys and formats for Anima LoRAs, and compatibility support isn't fully sorted out across all software yet:

I investigated both CAMKNXAnima and RIVALSANIMA, and the results were rather interesting.

The ~90 MB file size isn't the problem. Both are structurally valid rank-32 LoRAs.

Here's what I reproduced using Anima Base v1.0:

• ComfyUI v0.40.0: Both original LoRAs loaded successfully and visibly affected image generation.
• Diffusers 0.41.0 + PEFT 0.20.0: Both original Kohya-format LoRAs were silently ignored. No exception, zero adapter layers loaded, and the generated images were pixel-identical to the no-LoRA baseline.
• After conversion: Both LoRAs worked in Diffusers, with all 280 LoRA modules loaded and visible effects on generation.

Interestingly, this also affects RIVALSANIMA, even though it was trained on Anima Base v1.0. CAMKNXAnima was trained on Preview3, so model-version differences may separately affect character fidelity.

I also tested a conversion approach that uses ComfyUI-style keys and folds the LoRA's alpha / rank scaling into the weights. It worked in both tested backends.

I put together three resources:

• README and usage instructions
• Anima LoRA compatibility technical report
• Python conversion script

This doesn't necessarily identify the cause in your particular setup, since I don't know which software or versions you're using.

But if you're using Diffusers, the LoRA format alone could explain why it appears not to work. The same issue may affect other Anima LoRAs in the repository, too.

The good news is that retraining may not be necessary. Converting the existing LoRA can be enough.

...........Preview 3 should still work for Anima base 1 tho.

That's what's wild, and Rivals ANIMA DID WORK in comfyUI when i tested it.

CamKNX i tested on Arc, and it just WOULDNT work.

That's WILD though, that means something went NUTS in the middle of me training and the software added a bug --

I'll get back to you on that because whilst YES i've quit training (HEY YO XD LIKE I KNOW I SOHULDNVE HAVE RSD"ed out but pfft at least i know what stupid trainer i used XD)

.. it was KOHYA-SS BASED but it was a REALLY lego-scoped nextjs style training system i started and i promise you the backend was never fixed, and was always fighting shit.

In fact, those LoRAs worked perfectly with Base v1.0. They also worked properly in terms of their effect on the output images. In the latest ComfyUI, they all worked fine as-is. The LoRAs are “essentially” fine and aren’t broken at all.

However, when running them with Diffusers 0.41.0 or earlier (the latest as of today)—and possibly also when running them with older versions of ComfyUI, ForgeNeo, or other software—the LoRA keys will be ignored because they don’t match expectations unless the data is converted to ComfyUI format beforehand.

The fact that “converting to ComfyUI format makes it work fine with Diffusers 0.41.0” is “almost purely coincidental.” A real fix is in the works on the Diffusers side, but it doesn’t seem to have been incorporated into the release version yet.
On the other hand, on the ComfyUI side, the Anima LoRA created for Diffusers isn’t being applied correctly, and it seems the fix hasn’t been implemented there either.

Simply put, it’s like, “Ugh, this is terrible (today)🤢.” 😆

In fact, those LoRAs worked perfectly with Base v1.0. They also worked properly in terms of their effect on the output images. In the latest ComfyUI, they all worked fine as-is. The LoRAs are “essentially” fine and aren’t broken at all.

However, when running them with Diffusers 0.41.0 or earlier (the latest as of today)—and possibly also when running them with older versions of ComfyUI, ForgeNeo, or other software—the LoRA keys will be ignored because they don’t match expectations unless the data is converted to ComfyUI format beforehand.

The fact that “converting to ComfyUI format makes it work fine with Diffusers 0.41.0” is “almost purely coincidental.” A real fix is in the works on the Diffusers side, but it doesn’t seem to have been incorporated into the release version yet.
On the other hand, on the ComfyUI side, the Anima LoRA created for Diffusers isn’t being applied correctly, and it seems the fix hasn’t been implemented there either.

Simply put, it’s like, “Ugh, this is terrible (today)🤢.” 😆

Which means there's something likely wrong in how the training system i'd developed DID the lora because all the other loras work, which ... is a testament to why i quit doing it because MATH LOL

Oh, there’s probably nothing wrong with your Trainer. Or rather, it’s not the Trainer’s fault—it’s a minor slip-up on the part of the generation software or the base model distributor. Well, Anima is still growing, so it seems to change quite a bit.
Actually, both of your LoRAs are fine. However, during the development of the generation software, more than four “dialects” of the standard LoRA format (in this case, just regarding how key names and alphas are handled) ended up being created. These dialects are technically correct, but there are versions of the software that don’t recognize them—a problem that was particularly common in the early days of computing. 😅

For now, the reasonable takeaway is:

1. It’s not your fault; Anima simply has many LoRA dialects.
2. If you just leave it alone, the developers will likely fix the software so that any dialect works without issues.
3. You can use it normally right now by manually converting the dialects for each version.
   That’s probably the best way to look at it.

it because MATH

Hahaha. I’m the type who thinks about running away the moment I see a mathematical formula, too. Or I’ll ask ChatGPT, “What kind of relationship does this formula represent? Can you explain it in Japanese?” or “Can you rewrite it in Python?” But in most cases, I still end up having no choice but to run away!😆