This LoRA can help some models (that are not dedicated furry models) to create female deer anthros.

ZiT version

Experimental LoRA, just as all the other versions. But I would consider it the best version so far.

Integrate

anthropomorphic deer woman

into your prompt.

It was trained on Z-image “de-turbo” released by Ostris to avoid the negative effects that come with training distilled model. It can be used with the original Z-image Turbo model as well as with the de-distilled model.

Flux version

Version 2

Enhanced version with bigger and better dataset.

Human bias of V1 were greatly reduced. Much better fur detail.

The LORA is still experimental: PLEASE READ TEXT BELOW.

Use c3rv1d deer woman (humanoid c3rv1d deer woman for V2) in your prompt. Sometimes however this trigger is not enough to get good results. Flux still has strong human bias, describing the deer features can help.

Prompt for the same seed comparison:

a real photo of a c3rv1d woman, the c3rv1d deer woman is wearing a cute pink dress, the focus is on her upper body, the scene takes place at the beach.

Acknowledgment:

I’m aware that flux-dev is able to create humans with deer head without a LORA. But that was not the goal of the training. It is supposed to generate something that actually looks like a furry and not just a human with an animal head. If you look closely at the example images you can see that the skin has fur texture.

Also the LORA allows you to reduce all the deer features you would have to describe to the trigger, at least in many cases. The effect is very visible in the same seed example. It is not particularly cherry-picked. I only generated 5 pairs and they were all were very similar.

Even if you describe all deer features the base model would with 90% probability only generate a human with a deer head and ignore all the other aspects.

The purpose of the LORA is to address this problem.

Also I want to add that I didn’t create this LORA to use is standalone. I made to use it in combination with realistic SDXL models: Make a rawgen with flux and inpaint with SDXL.

The RTX3060 trained LORA did in fact turn out better than I expected, so it can actually be used alone.

I hope the LORA is useful somehow! Please share your results!

Known issues:

It has problems with generating hooves. This seems to be a general flux problem, I don’t know a single furry flux LORA that doesn’t have this problem.

It may create nudity even if not prompted, because the fully synthetic dataset is partially NSFW. Also it has bias towards bigger breasts sizes.