Downloads · 30 days
62
19% of all-time downloads
Bl4ckSpaces/Ilustrious-Nova-V3.6-PRO
Ilustrious-Nova-V3.6-PRO is a text-to-image model from Bl4ckSpaces. Use it when you need an image from a text prompt. It is set up for diffusers. The card lists the license as openrail++.
V3.6 PRO is the definitive anatomy fix version. It utilizes high-precision FP32 filtering and DARE-style rescaling to transplant IllumiYume's superior anatomy into NoobAI's pristine art style without the blurring issu…
Downloads · 30 days
62
19% of all-time downloads
All-time downloads
322
Public
Repo size
13.9 GB
Likes
0
Public
Click a slice to open those files.
.safetensors6.9 GB · 100%
From the Hugging Face model README
V3.6 PRO is the definitive anatomy fix version. It utilizes high-precision FP32 filtering and DARE-style rescaling to transplant IllumiYume's superior anatomy into NoobAI's pristine art style without the blurring issues common in standard merges.
Unlike standard mergers that operate in FP16, this model upcasted all tensors to Float32 before calculating deltas. This prevents "rounding noise" (micro-grain) from accumulating in the final image.
Instead of a static threshold, we used dynamic quantile filtering:
After injection, every modified layer was mathematically rescaled to match the Standard Deviation of the original NoobAI layer. This restores contrast and sharpness immediately, eliminating the "washed out" look of mixed models.
The VAE is locked to NoobAI original. No merging happened on the VAE to prevent latent space glitches.