Update

A new model vpred v065S iter based on NoobAI vpred v0.65S version has been added. You need ComfyUI, reForge or a1111 dev branch to use it. This model version has better knowledge and prompt comprehension but may be less stable than previous epsilon versions.

All samples are pure txt2img with euler (not euler a) sampler with simple scheduler, 40 steps, 6 CFG (comfy workflows are embedded). Other samplers may not work well with vpred model. Because euler is pretty basic sampler I found that 35+ steps are required for acceptable hands drawing.

This model is a drop-in replacement for the vanilla NoobAI-XL. It aims to generate slightly better results while maintaining the original model's style, composition, and flexibility as closely as possible. Improvements include reduced body horrors and hallucinations, better anatomy, enhanced backgrounds, and more detailed outputs.

Based on my limited testing, the model has at least partially succeeded, with over 50% of generations appearing subjectively better while preserving the original NoobAI composition and style. I have uploaded several comparison images with prompts directly copied from the NoobAI gallery. However, due to the inherent randomness in AI generation, you should conduct your own testing to evaluate the model's performance.

Usage

You don't need to use Pony "score_x" and "source_x" tags. Start with simple "masterpiece" and "worst quality" in positives and negatives and work from there. For prompting, use booru tags. In tag names replace "_" with a space and escape brackets "(" to "\(" and ")" to "\)". Note that this model does not change the "default" style of NoobAI. As with the vanilla model, you should use artist tags to improve image aesthetics. Just copy the artist tag to the prompt; there's no need to add extra "style of" or "artist:". Most artists with 500+ posts work more or less, some better than others. Including multiple artist tags with different weights can produce interesting results too.

For sample images, I used the Euler A sampler with the AYS (Align Your Steps) scheduler at 30 steps with a 6.5 CFG. I think this is a good starting point, but any settings compatible with NoobAI should work as well. I believe short, concise prompts work best with this model. Extra-long prompts with a lot of magic seem to do nothing useful while limiting the model's flexibility.

All sample images are pure txt2img and have a1111 metadata or ComfyUI workflows embedded if you are interested.

LoRAs compatibility should be the same as with the vanilla NoobAI.

Merge Recipe

Nothing fancy was done. The weights difference from IterComp 1.0 (and SPO for iter+spo version) vs. SDXL 1.0 was selectively applied to the Unet blocks of NoobAI-XL epsilon-pred v0.75, with the goal of minimizing resulting style and composition differences. Original TE from NoobAI-XL was left unmodified.

Drop the checkpoint into ComfyUI, and you should see the merging workflow. Feel free to modify it and share if you like. I'll be very interested to see your results.