WAI-illustrious-HSWQ_v17.0_base_checkpoint

WAI-illustrious-HSWQ

<Gallery />

Model description

WAI-illustrious-SDXL HSWQ Enhanced Fast Like FP8. Stable Like Full Precision. A performance-focused merge built on the HSWQ FP8 edition of WAI-illustrious-SDXL v17.0, enhanced with DMD2, Stabilizer IL, and SPO-SDXL_4k-p_10ep. This model was created with a simple goal: Deliver high-quality anime image generation while maintaining the speed and efficiency advantages of FP8 quantization. By combining modern optimization techniques with carefully selected merges, this checkpoint aims to provide strong prompt adherence, improved stability, and excellent visual quality even at relatively low step counts.

What's Inside Base Model

WAI-illustrious-SDXL v17.0 Additional Merges

DMD2

Stabilizer IL

SPO-SDXL_4k-p_10ep Each component was selected to improve a different aspect of generation quality:

DMD2 helps achieve stronger results at lower sampling steps.

Stabilizer IL improves generation consistency and reliability.

SPO-SDXL_4k-p_10ep enhances composition, structure, and detail retention. The result is a model optimized for both rapid iteration and everyday image generation.

What is HSWQ? HSWQ (Hybrid Sensitivity-Weighted Quantization) is an advanced quantization technique designed to reduce model size and VRAM requirements while preserving image quality. Traditional FP8 quantization often applies the same level of compression throughout the entire network. While efficient, this can sometimes introduce quality degradation in layers that are particularly important for image generation. HSWQ takes a smarter approach. Instead of treating every layer equally, it analyzes layer sensitivity and applies quantization more selectively. Critical layers receive greater protection while less sensitive components can be compressed more aggressively. This allows the model to retain much of the visual quality and prompt understanding of the original checkpoint while benefiting from the efficiency of FP8. Benefits of HSWQ

Lower VRAM requirements

Faster loading times

Faster ...

Credits

Original model by: unexpectedlyprovided Civitai: Link