High-fidelity ConvRot INT8 and ConvRot NVFP4 quantization for SDXL, Flux1.dev, and Z Image Turbo diffusion models. HSWQ uses sensitivity and importance analysis instead of naive uniform cast.
- ConvRot INT8 (SDXL V3.1): ComfyUI-compatible
int8_tensorwisepack with FULL ConvRot on remaining Linear/Conv2d after DualMonitor + V4 weighted-histogram FP16 protection under a fixed 300 MiB budget. Keep ratio is 0 (r0); critical layers stay FP16 via automatic analysis, not a keep-ratio percentage. Pack path matchesnative_convert_int8_convrot.py. - ConvRot NVFP4 (SDXL): ComfyUI Load Diffusion Model
nvfp4pack with FULL ConvRot (Linear→NVFP4, Conv2d→INT8int8_tensorwise) after DualMonitor + V4 pack-MSE FP16 protection under a fixed 600 MiB budget. Keep ratio is 0 (r0); calib writes NVFP4.input_scale. Script:hswq_convert_nvfp4_convrot_1.0.py. - Z Image INT8 (HSWQ): Development and public release ended. For Z Image, native ConvRot INT8 already reaches roughly SSIM > 0.99 in general, so a separate HSWQ Z Image 8-bit line is no longer developed or published. Use native ConvRot INT8 for Z Image 8-bit; HSWQ INT8 work continues for SDXL.
Technical details (FP8): md/HSWQ_ Hybrid Sensitivity Weighted Quantization.md — FP8 development has ended; this document is retained as a technical asset.
Technical details (INT8 FP16-protect / pack overview — ConvRot pack guide not published yet): md/HSWQ_INT8_SDXL_Technical_Guide.md
ComfyUI Loader for ConvRot INT8 / INT8: To use these models in ComfyUI, please use this custom node: ComfyUI-nunchaku-unofficial-loader ComfyUI Loader for ConvRot NVFP4: To use these models in ComfyUI, please use this custom node: ComfyUI-HSWQ-Loader-and-Tools
- SDXL (ConvRot INT8): How to quantize SDXL ConvRot INT8
- SDXL (ConvRot NVFP4): How to quantize SDXL ConvRot NVFP4
- Z Image (native ConvRot INT8): How to quantize Z Image — HSWQ-specific Z Image development has ended; this How-to only introduces the general ConvRot INT8 quantization method.
Benchmark results:
- SDXL (ConvRot INT8): MSE / SSIM
- SDXL (ConvRot NVFP4): MSE / SSIM
| Feature | ConvRot INT8 (SDXL V3.1) | ConvRot NVFP4 |
|---|---|---|
| Compatibility | ComfyUI int8_tensorwise / QUANT_ALGOS compatible |
ComfyUI Load Diffusion Model / QUANT_ALGOS nvfp4 compatible |
| File format | INT8 weights + scale (int8_tensorwise); SDXL V3.1 packs remainder with FULL ConvRot |
Linear NVFP4 + Conv2d INT8 (int8_tensorwise); FULL ConvRot on eligible layers |
| Image quality (SSIM) | 0.94–0.98 | 0.92-0.98 |
| Mechanism | Absmax + DualMonitor / V4 FP16 protect (r0); then FULL ConvRot on Linear/Conv2d remainder | Absmax + DualMonitor / V4 FP16 protect (r0, 600 MiB); FULL ConvRot (Linear→NVFP4, Conv2d→INT8) |
| Keep ratio | 0 (fixed) | 0 (fixed) |
| Benchmark | Measurable | Measurable |
| Use case | SDXL ConvRot INT8 distribution / kitchen loaders | SDXL ConvRot NVFP4 distribution / native ComfyUI load |
Note (Z Image 8-bit): HSWQ Z Image INT8 development and publication ended. Native ConvRot INT8 is sufficient for Z Image (typically SSIM > 0.99). HSWQ INT8 remains the SDXL path.
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 → kept in FP16 when selected by HSWQ. ConvRot INT8 / NVFP4 (SDXL): keep ratio 0; FP16 set comes from automatic analysis / budget ranking, not a keep-ratio %.
- Importance: V1 uses per-channel input mean-abs; V4 uses per-element SVD leverage × RMS magnitude hybrid → weights of the weighted histogram. Technical details: Dual Monitor System — Technical Guide.
-
Rigorous grid / pack simulation
- ConvRot INT8 (SDXL): natural absmax pack point for the symmetric INT8 grid; V4 weighted-histogram MSE ranks FP16 protection candidates (does not choose pack amax). SDXL V3.1 then applies FULL ConvRot (Hadamard rotate → channelwise absmax) on remaining Linear/Conv2d, identical to
native_convert_int8_convrot.py. - ConvRot NVFP4 (SDXL): absmax pack point for Linear→NVFP4 and Conv2d→INT8; V4 pack-MSE ranks FP16 protection under the 600 MiB budget, then FULL ConvRot on eligible remainder.
- ConvRot INT8 (SDXL): natural absmax pack point for the symmetric INT8 grid; V4 weighted-histogram MSE ranks FP16 protection candidates (does not choose pack amax). SDXL V3.1 then applies FULL ConvRot (Hadamard rotate → channelwise absmax) on remaining Linear/Conv2d, identical to
-
Weighted MSE Optimization — Finds parameters that minimize quantization error using an importance-weighted histogram (not a plain frequency histogram).
- V1 / Fast: per-channel importance (activation mean-abs) drives the histogram. Technical details: Weighted Histogram MSE — Technical Guide.
- V4 (SVD × RMS hybrid): per-element importance blends SVD structural leverage (L(i,j)=(U_i\cdot\sigma)^2\cdot(V_j)^2) with RMS magnitude; (\alpha) tilts toward SVD on heavy-tailed layers. Used by SDXL ConvRot INT8 and ConvRot NVFP4 for FP16-candidate ranking at the absmax pack point. Technical details: HSWQ V4 SVD-RMS — Technical Guide.
- Script:
quantize_sdxl_hswq_v3.1.py(SDXL ConvRot INT8; 300 MiB FP16 budget; FULL ConvRot default ON). - SDXL V3.1 order: (1) FP16 keep via DualMonitor + analyze + V4 under the 300 MiB budget → (2) remaining Linear/Conv2d FULL ConvRot INT8 (
native_convert_int8_convrot.pypack path). Card 1 / Card 2 forced OFF. - Tensorwise / channelwise: format tag
int8_tensorwise. ConvRot layers use rotate → channelwise absmax; non-ConvRot remainder uses per-tensor absmax unless Card 3 is enabled. - Card 3 (
--per_channel_int8): per-output-channel amax / scale for non-ConvRot plain packs (SDXL). - Keep ratio: 0 (fixed) — FP16 protection is automatic (analyze Hard VETO + DualMonitor + V4 ranking inside the FP16 budget), not a percentage keep-ratio.
- Z Image INT8 (HSWQ): Ended — no further HSWQ Z Image 8-bit development or Hugging Face publication. Prefer native ConvRot INT8 for Z Image (typically SSIM > 0.99).
- Script:
hswq_convert_nvfp4_convrot_1.0.py(SDXL ConvRot NVFP4; 600 MiB FP16 budget; FULL ConvRot default ON). - Order: (1) FP16 keep via DualMonitor + analyze + V4 under the 600 MiB hard ceiling → (2) remaining eligible layers FULL ConvRot — Linear → NVFP4, Conv2d → INT8 (
int8_tensorwise). - Loader / format: ComfyUI Load Diffusion Model / QUANT_ALGOS
nvfp4; NVFP4.input_scalefrom PTQ calib (--calib_file). - Keep ratio: 0 (fixed) — FP16 protection is automatic (analyze Hard VETO + DualMonitor + V4 pack-MSE ranking inside the FP16 budget), not a percentage keep-ratio.
- Image quality (SSIM): 0.92-0.98.
- Samples: 32 (recommended) — number of calibration samples (same for FP8 and ConvRot INT8).
- Steps: 25 — number of inference steps per sample during calibration (same for FP8 and ConvRot INT8).
- ConvRot group size (SDXL V3.1): power of 4 (default from script CLI);
--no-convrotdisables FULL ConvRot (plain pack only).
| Model | SSIM (Avg) | File size | Compatibility |
|---|---|---|---|
| Original FP16 | 1.0000 | 100% | High |
| Naive FP8 | 0.75–0.93 | 50% | High |
| HSWQ ConvRot INT8 | 0.94–0.98 | 68% (FP16 mixed) | High (ComfyUI INT8) |
| HSWQ ConvRot NVFP4 | 0.92-0.98 | 60% (FP16 mixed) | High (ComfyUI NVFP4) |
HSWQ ConvRot INT8 targets SSIM 0.94–0.98; HSWQ ConvRot NVFP4 targets SSIM 0.92–0.98. Both keep full loader compatibility on their respective formats.
Version history and release notes are in CHANGELOG.md.
This project is built upon the following repositories:
| Repository | In-repo path | Upstream |
|---|---|---|
| ComfyUI | ComfyUI-master/ |
@Comfy-Org — The most powerful and modular diffusion model GUI, API and backend with a graph/nodes interface. |