If you want to use my model online you can do so here on Tensor Art. Currently only has Illustrious version 3.

I guess I lied.

I wanted to be part of the cool gang and hop on the NAI v-pred craze. I still think Illustrious is better. Here's what I cooked up.

The prompts I used to generate the sample images were lost to the void, I guess.

NoobAI version:

As with any of these models, use Euler or Euler A samplers. Comfy user should use ModelSamplingDiscrete and set it to v-prediction. I'll add more info here as I continue to research but for now, resources.

You can find CSV files filled with tags the model understands here. Both danbooru and e621. Which... just means prompting like you would an Illustrious model.

ComfyUI users can use the example workflow here. Just drag the image into comfy to open the workflow.

General guide for using NoobAI models effectively.

Illustrious version:

Triple model merge. It's a first merge attempt, but it came out alright I'd say. Understands Danbooru and e621 tags.

It's an EPS model but it seems to work fine without EPS sampling, but I'd still recommend you use the correct settings for whichever tool you use. (ComfyUI/A1111/etc.)

I do not use A1111, but for ComfyUI you have to use the node ModelSamplingDiscrete or an equivalent custom node and set it to EPS sampling.

Use Euler A sampler, ~25-40 steps and ~4-8 guidance scale (CFG).

Positive Illustrious prompts:

(subject, {1boy/1girl/2boys/2girls/etc.}, artist/style).

Describe image with natural language. {A portrait of ...}.

(tags, quality tags).

Negative Illustrious prompts:

bad quality, worst quality, worst detail, sketch, censor.

Pony version:

Test model, I wanted to see if merging two of my favorite models would increase prompt adherence while still keeping its imagination.

It has a bit of a watermark problem, but they don't seem to be too intrusive and are usually easily fixed with a bit of inpainting.

Positive Pony prompts:

score_9, score_8_up, score_7_up, rating_tag, BREAK

Negative Pony prompts:

score_6, score_5, score_4, etc.

But please experiment, I'd love to hear your thoughts!