Z Image Introduces Trajectory-Sensitivity Quantization Theory
Zestyclose_Bake3680 · reddit · 2026-08-19
The author released the HSWQ Hybrid ConvRot NVFP4 quantization method and ComfyUI loader for Z Image. While file sizes remain similar to Convrot INT8, this approach significantly reduces VRAM usage and boosts processing speed.
Technically, it abandons previous static weight importance theories (like Histogram MSE, SVD, or Cosine) in favor of Trajectory-Sensitivity. This dynamic metric ranks layers based on the actual divergence caused by their quantization errors after propagating through the full model and sampler. It accounts for inter-layer dependencies, non-linear amplification, and error cancellation, making it applicable to any iterative sampling system. A Krea2 version is currently in development.
More from Infra
- LLM Midtraining Guide: Data Mixing and Overfitting Prevention — cwolferesearch · 2026-08-19
- Brex Report: Infrastructure Vendors Top Fastest-Growing List — marcbhargava · 2026-08-19
- AI safety monitoring could consume 20% of inference compute — ohlennart · 2026-08-19
- RTX 6000 Blackwell test: 300W Max-Q card power locked, cannot overclock — MelodicRecognition7 · 2026-08-19
- Report: China allows small batches of Nvidia H200 chips to aid AI firms — The Decoder · 2026-08-19
- User reports $200 OpenAI Pro credits lasted only 12 hours for heavy Codex usage — ___Patrice___ · 2026-08-19