Research update 012 · 21 September 2026
FastH3 V2 has a speed boundary before it has a speed claim.
Community report: one stated Ryzen AI Max+ 395 / Radeon 8060S setup completed a 243-frame FastH3 V2 run at 768×1344 in 33 minutes 9 seconds. At 24 fps, that is about 10.1 seconds of media. This is a local-render lead, not live video.
The number looks different beside the clip
| Source-stated condition | Reported time boundary | Output duration | What the arithmetic says |
|---|---|---|---|
| FastH3 8-Step V2 · 768×1344 · 243 frames · Ryzen AI Max+ 395 / Radeon 8060S · 128 GB unified memory · ComfyUI v0.36.0 | 33 min 9 s end to end | About 10.1 s at 24 fps | About 196 wall-clock seconds for each output second |
The arithmetic is 1,989 seconds divided by 243 ÷ 24. It is useful for a native-canvas local-render decision. It does not establish a general Radeon baseline, a quality result, cold-start behavior, cost, retries, or viewer-visible delay.
Why a smaller canvas may not get the FastH3 path
Community report: the author says FastH3’s VSA, or Video Sparse Attention, stayed dense below 12,288 tokens. In the author’s 640×384 case, the sparse path did not produce the claimed advantage. At the larger 768×1344 canvas, the author says the VSA path engaged.
That turns “FastH3” into a condition card: name the canvas, frame grid, token count, attention backend, GPU and host memory, version, and timing boundary. Without those fields, a speed comparison is not a setup recommendation.
The scope conflict is still waiting beside the timer
Project says: FastVideo’s current eight-forward V2 schedule is T2AV-only and not a distilled reference model. ComfyUI says: its current text-to-video template agrees; its paired image-to-video template says the same checkpoint supports first/last-frame FL2VA. The templates now expose the disagreement directly.
Do not infer that the reported native-canvas render proved a frame-anchored transition. It does not identify an FL2VA run. If the next scene must honour a chosen first or last frame, pin the package and template, run that exact task, and inspect both boundaries before comparing speed.
Use this route as a local-render lead
- Use the report to decide whether a large native-canvas clip is plausible on this named memory-heavy iGPU setup.
- Do not call 33 minutes for about 10 seconds real time, continuous generation, or a response-time measurement.
- Before a download, record the required task, canvas, frame grid, attention backend, model files, GPU or unified memory, host memory, ComfyUI version, and cold/warm boundary.