h3.c 的 fast 模式到底删了什么：六个计算轴不能混成一个开关

很多推理工具把调整收进一个 fast=true。按理说，速度变快后使用者只看到总耗时下降。却不知道程序少算了哪些部分，也不知道画面出错时应该撤回哪个开关。

h3.c 固定版本 8974cc0 把这件事拆得很具体。它允许减少去噪步数、复用完整 DiT 输出、减少活跃层、复用 Transformer core、缩短中间视频序列，或者降低内部画布。每个选项删掉的计算不同，保留的信息不同，失效方式也不同。

这篇不回答“哪个 preset 最快”。它只建立一张预算表：改动发生在哪个轴。哪些部分仍保持新鲜，代价会先出现在哪里还有如何做单变量验收。

先把 reference、default 和 aggressive 分开

固定 README 提供了一组清楚的对照口径：

控制轴	慢参考	常用设置	激进设置	实际减少的工作
去噪时间 steps
DiT forward 次数完整去噪器复用 reuse
真实 velocity 求值次数网络深度 layers
每次 forward 的活跃 block Core residual 复用 core-reuse
Transformer core 刷新次数中间序列关闭可选 token-reduction
中间 block 的目标视频 token 内部画布输出尺寸 → DiT 与 VideoVAE 的空间网格

This table is used to distinguish different computational axes. Skipping a denoising step。dropping five layers,reusing a core residual segment,or merging half of video tokens will all save compute,but y break different information.

少 steps：直接把时间轨迹采稀

The flag –steps N always means executing N denoising passes. Reducing it doesn’t preserve some positions and make m cheaper;it directly uses a sparser trajectory.

This optimization first affects scene construction,motion propagation and post‑detail cleanup. In README’s × and frame author experiments,a four‑step path took about X s versus Y s for pass reference. Full‑video SSIM values were roughly Z and W respectively. It proves that speed differences come hand in hand with visible quality differences—quality can’t be judged by “still recognizing subject” alone.

A low‑step setting is best suited for previewing composition. When verifying you should first inspect subject count,camera motion and overall structure rar than just checking sharpness on final frame.

Whole reuse：保留时间点，少算 velocity

The flags –reuse N1 –reuse N2` still progress along an Euler transition with N points but do not run a full DiT at every point. In this version under X steps it performs Y true DiT evaluations and uses linear extrapolation from neighboring sigma intervals for rest.

This removes full denoiser evaluation while relying on velocity being smooth 娱乐ween adjacent sigma ranges. The first and last evaluations remain true to limit drift but cannot guarantee stable mid‑motion,limbs or detail.

• Sparse steps change solution grid.
• A whole reuse keeps transition counts but replaces part of model evaluation with extrapolation.

Core reuse：边界每步刷新。昂贵主要少刷新

The flag –core-reuse N` does not reuse complete outputs. According to h3.h it lowers Transformer core recomputation frequency while still refreshing timestep‑sensitive patches and final heads each step.

current noisy latent → patch / timestep input per step
→ reuse or recompute Transformer core residual
→ final head per step

This retains more time‑point information than whole reuse and relies on assumption that Transformer core residual changes smoothly over adjacent timescales. Note that both options are mutually exclusive in this implementation—y cannot be stacked arbitrarily as separate multipliers.

少 layers：每次都跑，但网络变浅

The flags –layers N1 –layers N2)` are not just mechanical cut‑offs of last layers. The source protects essential beginning and ending blocks while selecting active blocks based on AdaLN gate signals in checkpoints. Unactivated block weights and schedule tensors are discarded,thus reducing both compute and resident transformer storage.

This changes network depth per forward pass—each time point still runs but with fewer layers may lose object relationships,conditional compliance or detail repair ability. Verification should compare results under identical steps,reuse settings,seed values against baseline layer counts to attribute differences correctly.

Token reduction：只缩目标视频中段，并保留 full-resolution bypass

The flag –token-reduction K L M…` is often misunderstood as generic resize. README and h3_dit.c show it only pairs horizontally adjacent tokens within middle DiT blocks of target video suffixes two by two while keeping text/audio/condition/reference tokens at original length plus maintaining full resolution as bypass.

When restoring each original token keeps its own value plus updates learned from paired branches. This reduces intermediate block processing length without simply shrinking final video. The design preserves high‑frequency detail within individual tokens but paired branches learn shared updates—so composition may still deviate from close path. In fixed device tests authors reported denoise time dropping from X s to Y s after adding layers + reuse + token reduction. Composition may change as noted in README. These numbers are author reports under specific conditions—no local replication here. .

This section is based solely on author documentation—not reproduced locally by me—or cross‑device generalization checks. For production readiness consult furr testing or vendor guidance before enabling this option on end‑user workflows.

---

Author Reported – No Local Runtime – Unknown Production Suitability

---

---