Video and Workflow Export - Star7

The whole "workflow included" culture in ComfyUI runs on one trick: SaveImage stuffs the node graph into the PNG as text chunks, so the image is the project file. Drag it back in, the whole pipeline reconstructs. Videos never got that treatment - VideoHelperSuite gives you a file, but the graph that made it is gone the moment you close the tab.

Video and Workflow Export - Star7 is the fix. It sits at the end of a MiniMax H3 video workflow and repackages the finished MP4 so the workflow travels with it. It's part of the same minimax-h3-chunk-star7 pack as the VRAM-chunking core node, but it's a standalone utility: no model patching, no attention tricks, just metadata plumbing.

How it works

Two inputs, and only one of them is really yours to think about. 视频文件 accepts a file path or the filenames output from VideoHelperSuite (the node is built to read VHS output directly). 导出方式 is an export-mode dropdown with four options - fair warning, the UI labels are Chinese-only (this node has no English localization, unlike the pack's prompt loader):

• 视频内置工作流 (default) - remuxes the video with ffmpeg and injects the workflow and prompt as metadata tags in the container, without re-encoding (-c copy, so it's fast and lossless). MP4/MOV also get faststart so it streams nicely.
• 仅视频 - remuxes and strips all metadata. Use this when you want a clean file for sharing.
• 视频 + 工作流 JSON - writes a <video>.workflow.json sidecar next to the file.
• 视频 + 工作流 PNG - embeds the workflow and prompt into a PNG as tEXt chunks, the classic ComfyUI convention. If VHS produced a workflow PNG it reuses that; otherwise it extracts a frame from the video.

There are no outputs - it's an output node, and that's the point. It grabs the current workflow and prompt from the running graph automatically (hidden inputs you never set), then overwrites the video in place. Two hard rules from the code: the video must live in ComfyUI's output/ or temp/ directory, and ffmpeg has to be reachable - it uses the imageio-ffmpeg binary ComfyUI ships with, falling back to a system ffmpeg.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/star7code/minimax-h3-chunk-star7.git

then restart. Or search "MiniMax H3 Activation Chunk - Star7" in ComfyUI Manager. No extra pip installs - ffmpeg comes from ComfyUI's own dependency.

Common issues

• "No workflow to write" (the error is in Chinese) - you ran the node API-only, or outside a normal UI run, so there's no graph to serialize. Run it from the UI and it goes away.
• File locked by a player or previewer. This one bites on Windows. The node replaces the file in place and retries the write-back 30 times, but if your video player has the file open, it gives up and tells you so. Close the player and re-run.
• Video outside output//temp/ gets rejected outright - it refuses to write somewhere it can't safely overwrite.
• Metadata isn't indestructible. The workflow survives byte-for-byte copies only; if a hosting site re-encodes your MP4, the tags are gone. Keep the original file.

The payoff loop

This node pairs with the pack's Prompt Load - Star7. Export embeds the workflow and prompt into the MP4; later, drop that same MP4 back onto the prompt loader and it reads the prompt straight out of the video metadata. Generate, export, forget the graph - and still recover the prompt weeks later. For anyone doing long H3 runs on a low-VRAM card where every second of compute is precious, a file that remembers its own setup is worth having.