Quality qualification¶
Speed and quality are independent axes. A route advances only when the generated MP4 is valid and its intended visual, temporal and audio behavior survives checks appropriate to the claimed optimization.
Qualification stack¶
- Provenance: same prompt, references, seed, canvas, fps, frames, NFE, shifts and model/config revisions for a controlled A/B.
- Media integrity: complete decode, expected frame count and duration, valid stereo audio and no truncated container.
- Numeric comparison: SSIM/PSNR for aligned outputs, motion delta, audio RMS/SDR and silence checks. These metrics do not replace semantic judgment.
- Blind multimodal review: label-balanced contact sheets sampled across the full timeline; score identity, prompt adherence, causality, continuity, composition and artifacts without revealing the method.
- Cross-scene generalization: people, dialogue, mechanical motion, camera movement, fine texture and reference identity before a production claim.
Prompt-treatment experiment¶
Four original, rights-owned prompt variants were run at 1344×768, 24fps, 124 frames, 4 NFE and seed 20260831 on two independent RTX 4090 workers:
| Variant | Generation | Peak VRAM | Blind result vs raw |
|---|---|---|---|
| raw | 154.948 s | 14293.724 MiB | reference anchor |
| local H3 skill | 165.818 s | 14340.082 MiB | 0–0, 3 ties |
| Qwen3.8 constrained rewrite | 151.200 s | 14563.726 MiB | 0–0, 3 ties |
| skill then Qwen3.8 | 154.861 s | 14587.016 MiB | 0–0, 3 ties |
All files were 1344×768 H.264 with 124 frames, 5.175 s duration and 32 kHz stereo AAC. The judge saturated on this already-specific brief and contact sheets cannot assess continuous wheel motion or audio synchronization. The correct conclusion is no demonstrated quality improvement, not that every treatment is universally equivalent. Exact prompts and aggregate evidence are in the example catalog.
Optimization quality examples¶
- Strict Triton RoPE: SSIM 0.994831 with sample-aligned audio unchanged.
- Pattern3 warmup20 cache: SSIM 0.990041 and 17.15% less block work on one 15 s Base40 run.
- Base24: SSIM 0.845058 on the same-seed anchor; one sensitive case lost 18.53% motion and 1.054 dB audio, so it is a fast tradeoff.
- Hybrid INT8: SSIM 0.857694 on one long case; promising speed, insufficient breadth.
- Global INT8+Sage: rejected despite 1.745× because motion and audio regressed materially.
Cross-architecture outputs are not used as pixel truth because CUDA kernels and execution order can change the trajectory even at the same seed.