High-fidelity FP8 quantization for SDXL, Flux1.dev, and Z Image Turbo diffusion models. HSWQ uses sensitivity and importance analysis instead of naive uniform cast. It offers two modes: standard-compatible (V1) and high-performance scaled (V2). V2 requires a dedicated loader and is not usable at the current time.
Technical details: md/HSWQ_ Hybrid Sensitivity Weighted Quantization.md
SDXL models: Hugging Face — Hybrid-Sensitivity-Weighted-Quantization-SDXL-fp8e4m3
- SDXL: How to quantize SDXL
- Z Image Turbo: How to quantize Z Image Turbo
Benchmark results: SDXL (MSE / SSIM)
| Feature | V1: Standard Compatible | V2: High Performance Scaled |
|---|---|---|
| Compatibility | Full (100%), any FP8 loader | Requires dedicated loader — not usable at present |
| File format | Standard FP8 (torch.float8_e4m3fn) |
Extended FP8 (weights + .scale metadata) |
| Image quality (SSIM) | ~0.98 (max) | Unmeasurable (no dedicated loader) |
| Mechanism | Optimal clipping (smart clipping) | Full-range scaling (dynamic scaling) |
| Benchmark | Measurable | Currently unmeasurable (no dedicated loader) |
| Use case | Distribution, general users | Unavailable until a dedicated loader exists |
File size is reduced by about 30–40% vs FP16 while keeping best quality per use case.
-
Dual Monitor System — During calibration, two metrics are collected:
- Sensitivity (output variance): layers that hurt image quality most if corrupted → top 10–25% kept in FP16 (for SDXL and ZIT, 10% often gives sufficient quality).
- Importance (input mean absolute value): per-channel contribution → used as weights in the weighted histogram. Technical details: Dual Monitor System — Technical Guide.
-
Rigorous FP8 Grid Simulation — Uses a physical grid (all 0–255 values cast to
torch.float8_e4m3fn) instead of theoretical formulas, so MSE matches real runtime. -
Weighted MSE Optimization — Finds parameters that minimize quantization error using the importance histogram. Technical details: Weighted Histogram MSE — Technical Guide.
- V1 (
scaled=False): No scaling; only the clipping threshold (amax) is optimized. Output is standard FP8 weights. Use this mode — full compatibility with any FP8 loader. - V2 (
scaled=True): Weights are scaled to FP8 range, quantized, and inverse scaleSis stored in Safetensors (.scale). Requires a dedicated loader; not usable at the current time.
- Samples: 32 (recommended) — number of calibration samples.
- Steps: 25 — number of inference steps per sample during calibration.
- Keep ratio: 10–25% — keeps critical layers in FP16. For SDXL and ZIT, 10% often gives sufficient quality.
- Latent: 32–256, default 128 — calibration latent size (H/W). Use
--latent 32for faster calibration,--latent 256for higher fidelity.
| Model | SSIM (Avg) | File size | Compatibility |
|---|---|---|---|
| Original FP16 | 1.0000 | 100% | High |
| Naive FP8 | 0.75–0.93 | 50% | High |
| HSWQ V1 | 0.86–0.98 | 60-70% (FP16 mixed) | High |
| HSWQ V2 | — (currently unmeasurable) | 60-70% (FP16 mixed) | Not usable (no dedicated loader) |
HSWQ V1 gives a clear gain over Naive FP8 with full compatibility. V2 would offer higher quality but requires a dedicated loader; benchmark is currently unmeasurable and V2 is not usable at the current time.
Version history and release notes are in CHANGELOG.md.
This project is built upon the following repositories:
- ComfyUI — The most powerful and modular diffusion model GUI, API and backend with a graph/nodes interface by @Comfy-Org.