antirez / antirez/h3.c

Questions about M4 Max support, 720p 15‑second generation time, and GUI

未关闭
#5 7 条评论 1 个 reaction 已指派 0 人 在 GitHub 查看
主要语言
C
星标
2.7k
派生
210
PR 合并指标
30 天内没有已合并 PR

描述

Hello,

I am very interested in using h3.c on my Mac with an M4 Max chip and 64GB of unified memory. Before diving in, I have three practical questions that I hope you or the community could help clarify.

1. **M4 Max (64GB) compatibility and performance**
The README mentions that current optimizations are incremental for M3 Max and M5 Max. Has there been any testing on the M4 Max? Are there any known performance figures (e.g., time per step for a 512×512, 22‑frame run) or special environment variables (like `H3_NAX` or Metal flags) that might improve performance on this chip? Also, does the 64GB memory size handle the full model comfortably, especially when using the `--show` preview (which adds ~10 GiB extra memory)?

2. **Estimated generation time for a 15‑second 720p video**
The benchmarks in the README show impressive speeds for short 512×512 clips. However, I would like to generate longer, higher‑resolution videos – for example, 15 seconds at 720p (roughly 362 frames at 24 fps, e.g., 1344×768 or 1280×720). Could you provide an approximate generation time on a typical M5 Max or M4 Max, assuming default settings (e.g., 20 steps, 45 layers, reuse 2)? If no direct numbers exist, any rough estimate or scaling rule (e.g., time proportional to frames × pixels) would be very helpful. I am also open to using internal render resolution or token reduction to speed things up, but I would like to know if the quality‑time trade‑off is acceptable for such long videos.

3. **Graphical user interface (GUI) support**
I know there are community‑made ComfyUI nodes (like H3 Studio, ComfyUI‑H3‑Conditioning‑Cache‑AI‑Drama‑Production‑Suite, etc.). Does the official h3.c project plan to provide its own integrated GUI in the future? Alternatively, do you recommend any particular third‑party frontend that offers a user experience similar to ComfyUI (e.g., node‑based workflow, real‑time preview, easy parameter tuning)? If the official stance is to keep it command‑line only, that is fine, but I would appreciate some guidance on which external tool is most actively maintained and supports the full feature set (including first/last frame, audio references, etc.).

Thank you for developing this excellent project. I look forward to any insights you can share.

Best regards,
Slzx1

贡献指南

这个仓库没有索引到贡献指南

评估

这个 Issue 还没有评估数据。

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。