| # Methods and measurement notes | |
| ## Why one-step error and not video PSNR? | |
| H3's sampler is chaotic: computing only the AdaLN in fp32 instead of bf16 already moves a 20-step clip to | |
| ~17 dB PSNR against the shipped model. End-of-run pixel metrics therefore cannot rank recipes. Each recipe is | |
| instead fed the exact recorded inputs of every step and compared with the BF16 output of that step. | |
| ## Measurement notes | |
| - **Noise floor ~1–2%.** Replays are bitwise exact within one container but differ by 1–2% across GPU hosts, and | |
| any bf16-level change lands at that level. Compare recipes within a run; each run reports its own floor. | |
| - **Size** is the approximate transformer checkpoint size with the rank-12 AdaLN (NVFP4 counted as 4.5 bits per | |
| weight including block scales). | |
| - **Simulation vs ComfyUI.** NVFP4 activation rounding is implementation-defined: ComfyUI's CUDA and Triton kernels | |
| agree with each other but differ from ModelOpt's simulation by ~2% on a quantized activation. Rankings should | |
| carry over; final quality is judged in ComfyUI itself (below). | |