Beta Testing

Is this better than v5? It is more realistic, but does it have greater artistic power? All comments are very much appreciated, but if you can, please help me answer:

Yes - in a measurable sense, e.g. "1 in 5 gens are better".
No - with e.g. "specific prompt" makes better gen on model X, considering:

1. prompt following, character accuracy

2. anatomy, hands, poses, positions

3. interesting composition, seed variety, bias (gender, ethnicity, etc)

4. lighting, color, special effects

5. faces, emotions, expressions

I mostly test with the Euler sampler and beta scheduler, to reduce noise. However, other combinations like dpm_2 sampler and kl_optimal seem to give great detail.

I make the merges with ComfyUI, I test new block merge permutations using a high resolution, 20-25 steps, cfg 1.2 + 0.8 perturbed attention guidance to make a 'total cfg' in the 2-3 range, then save new permutations that seem better for that prompt. After a few weeks, thousands of permutations, I've usually got a new version. I test prompts with and without realistic/photorealistic, a few art styles, with and without masterpiece, absurdres.
I keep the clip_l, clip_g and vae identical to the original animagine. I try to keep as much weight from the original unet, in particular that which controls composition, prompt following and lighting.

License

This model adopts the original CreativeML Open RAIL++-M License from Stability AI without any modifications or additional restrictions. The license terms remain exactly as specified in the original SDXL license, which includes:

• ✅ Permitted: Commercial use, modifications, distributions, private use

• ❌ Prohibited: Illegal activities, harmful content generation, discrimination, exploitation

• ⚠️ Requirements: Include license copy, state changes, preserve notices

• 📝 Warranty: Provided "AS IS" without warranties

Please refer to the original SDXL license for the complete and authoritative terms and conditions.