NOTE: Only work well with Anima base-v1.0 or similar models, heavily fine-tuned experimental models like AnimaYume v1.0 (beta) might not work well.
UPDATE: AnimaYume v1.0 Base (Final) can be used now.
Supported input resolutions for v2/v2b:
- 768px (trained)
- 1024px (trained)
- 1536px (tested, prone to error)
Supported input resolutions for V1.0:
- 512px (trained)
- 768px (tested)
- 1024px (slight error/glitch expected)
With help of local llm (Qwen3.6 27B), I managed to create extensions for forge neo that add latent reference image input. you can check it here.
Another weird LoRA experiment.
I've been playing around with Mirumo0u0's sd-scripts fork, specifically looking at their conditional training. Honestly, I was just curious if this could actually work for an Edit LoRA if I fed it before and after pregnant pairs, which I gathered from multiple H game CG last year for a failed attempt to train an SDXL ADDifT LoRA.
Turns out, it does, and it works.
This is an edit LoRA for pregnant edit. Because it was trained with target pairs that only change the belly, it wont edit the unrelated regions. It will produce some artifacts because of full latent sampling, but it's visually identical. This LoRA needs ComfyUI-Cosmos-Reference to work, and it only edits this single concept, nothing else. This might be only useful if you are lazy to draw masks for inpainting.
Description
Trained on 768 and 1024 px, so they might support up to 1536px.
FAQ
Comments (8)
wait, you can edit with anima? funny that im just finding out this now.
kinda, not as free as other model, just edit it in one concept only.
this one just for cheap preggo edit and done, cant do anything else.
@PertaliteMeister still should be useful. I have no idea why, but when doing pregnancy in anima, it tends to make the woman bigger than the man in a couple pose. Maybe its the vore in the dataset or something.
This LoRA is amazing! Do you plan on making LoRAs with other concepts as well?
most likely not, because this one datasets already prepared well enough since last year, and it still needs 20hr to train one iteration, and v2 doesn't meant the 2nd one lmao.
what probably i will do is training this concept using other method, like ControlNet-LLLite (most people use), or EasyControl (newest one that claim to be fastest), if one of them give me confidence on how fast they trained, I might explore more concepts next.
@PertaliteMeister Ah, I see! 20 hours is a lot of work. I hope you find a faster training method that works for you. Thanks again for this amazing LoRA!
could you please illustration how the workflow should be built to use this lora? It can't work on my recent one :(