Research update 009 · 18 September 2026
FastH3 just got a ComfyUI path. Its official instructions disagree on the one feature that may matter most.
FastVideo released FastH3 8-Step V2 and ComfyUI now has a current template for it. That makes the route easier to find. Before you plan a chained scene, read the scope twice: FastVideo calls the recipe text-to-video with audio only; ComfyUI says the same checkpoint also accepts first and last frames.
This is a different FastH3 recipe, not a new checkbox on Preview
Project says: FastVideo’s 15 September release is an eight-forward, data-free DMD2 MiniMax H3 checkpoint with 80% Video Sparse Attention. Its schedule page fixes eight denoising rungs and says the script rejects a different grid. The earlier FastH3 Preview route remains four-forward.
Any speed, quality, GPU, or cost number for a four-step Preview, a dense community conversion, or a B200 queue does not transfer automatically. The checkpoint and schedule belong on the same line as the result.
ComfyUI makes the package visible. It does not settle the task scope.
ComfyUI says: its current FastH3 8-Step V2 image-to-video template links a Comfy-layout diffusion model with named MiniMax H3 video/audio VAEs and a Qwen3-VL encoder. It states that no connected images produces T2VA, while a first and/or last frame produces FL2VA; Ref2VA is excluded.
Project says: FastVideo’s current schedule page calls that 8-Step V2 recipe T2AV-only. These are not cosmetic wording differences. A first/last-frame route can anchor a transition; a text-only route cannot. The pool found no dated clarification or run that reconciles them.
The pre-download card
| Ask this | What the sources say today | Do not infer |
|---|---|---|
| Which model? | FastH3 8-Step V2, an eight-forward FastVideo checkpoint. | That it behaves like four-step Preview or another FastH3 conversion. |
| Which task? | FastVideo says T2AV only; the current ComfyUI template says T2VA plus first/last-frame FL2VA, not Ref2VA. | That a first/last-frame transition works until a pinned run proves it. |
| Which workflow shape? | ComfyUI’s template names 24 fps, a 17k+5 frame grid, a 768-pixel short edge, and a 768×1344 cap. | Consumer VRAM, time-to-ready, quality, cost, reliability, or a live buffer. |
Use it as a local compatibility lead—not a live-video claim
A new template is useful because it tells a builder which files and nodes to inspect. It says nothing about queue depth, playout, retries, a viewer’s change reaching a scene, or output arriving before playback ends.
Start with one pinned five-second control. Record the ComfyUI build and template revision, exact checkpoint and component files, task, frames, schedule, attention path, GPU and system memory, cold and warm time, output, and error. If you need FL2VA, make that the first test—not an assumption attached after the download.