Nayelina Z-Anime
Anime Model Based on Z-Image
Overview
Nayelina Z-Anime is an anime-focused model built for Anime Style, designed to be used strictly with Danbooru-style tags.
This is the first public version, so while results are strong, some generations may still require prompt tuning.
Training
- Steps: 30,000
- GPU: RTX 5090
- Tag system: Danbooru tags
- Focus: NSFW anime content
Usage Notes
- Tag-based prompting.
- Works best with clean, structured Danbooru tags.
- Some poses, hands, or compositions may need refinement due to early version limits.
Version
v1.0 — Initial release
Hugginface
https://huggingface.co/nayelina/nayelina_anime
Description
Nayelina Z-Anime -- V1
FAQ
Comments (7)
I've tried this model out, few things of note.
The model responds better to natural language over tag based prompts (presumably because it's based on Z-image), so write out your prompts.
Its NSFW capabilities are fairly poor. Most generations of this type will be full of anatomical inaccuracies and eldritch horrors. Incomparable to even sd1.5 models - just use Illustrious / noob-ai / chroma for NSFW anime generations instead.
The SFW capabilities seem fine though. It's capable of 'suggestive' rather than 'explicit' images. Oftentimes though there will be artifacts on the edges of the screen and whatnot.
Overall this model has a lot of potential that I'm sure will be realized with further training.
Also I'm noticing often backgrounds have this weird pixelation to them rather than a sharp finish.
I recommend anima over those.
As I said, the model was trained with quantization on a 5090. I'll be uploading an update soon. I now have a slightly larger budget to train the model. This is just a sample; the model only has 30k steps.
what tools were used to train this ? (a toml or json of parameters would be appreciated) ? how many pictures too ? from DeepGHS dataset ?
if i were to continue training, what would you recommend ?
Hi, yes, I'll leave all that information on my hugginface. As I mentioned, this was trained on a 5090. I had to optimize many things, which is why the quality is a bit poor. I recently received funding from some people, so I'll probably train this on RunPod soon.
@Nayelina thanks for your reply, i already saw the post and was left hungry for more.
i was expecting to know if you used musubi tuner/one trainer/diffusion pipe or ai toolkit along with batch size and stuff... maybe you used a custom script then ? if you wanna keep it secret i get it, fine by me.