# I lost track of my takes after 1,200 H3 renders, so I built a take manager and editor for ComfyUI (open source) — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/84aa5ebdddbceeb2

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-30 14:38 ingested 2026-10-31 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xjh55q/i_lost_track_of_my_takes_after_1200_h3_renders_so/)

## retained text



```
I made a short film with MiniMax H3. It took about 1,200 renders to get 85 keepers, and by the end my output folder was useless. Hundreds of `ComfyUI_00xxx_.mp4` files, and no way to tell which prompt or seed made which, or which version of the prompt it came from. I re-rendered shots I already had, because finding the good take was harder than making a new one.

So I built **Picky**, a custom node pack plus a browser editor (it works like a small video editing app, with a timeline) that runs alongside ComfyUI.

What it does:

* **Takes stay on their shot.** Every render is filed under the scene it belongs to, with the exact prompt text and seed. Old prompt versions are kept, so you can go back.
* **Compare and pick.** Render a shot 4–8 ways, watch the takes side by side, star the good ones and mark one “in the edit”.
* **Build on the pick.** Extend it, re-render it at final quality with the same seed, or upscale it with SeedVR2. The result remembers its parent.
* **Structured prompts.** Shots, camera, dialogue, soundscape. It writes MiniMax’s official prompt format, so you don’t have to hand-write it.
* **Cast.** Set up a character once with reference images and write @ in any prompt.
* **A checker** built from a lot of failed renders. It warns about things H3 tends to misread, such as negations leaking in as burned-in text, or crowding too much into one shot.
* **Agent skill.** Claude Code or Codex can drive the same API, and the edits show up live in the editor.

It’s open weights only and fully local; I run it on one 24 GB card. GPL-3.0.

* Walkthrough (2 min): [https://youtu.be/NMKWWtzNiIM](https://youtu.be/NMKWWtzNiIM)
* Try the editor in your browser. It’s a recorded project, so renders are simulated, but picking, rating and the prompts and checker all work: [https://chenthedesignmaker.github.io/ComfyUI-Picky/demo/](https://chenthedesignmaker.github.io/ComfyUI-Picky/demo/)
* Repo: [https://github.com/chenthedesignmaker/ComfyUI-Picky](https://github.com/chenthedesignmaker/ComfyUI-Picky)

It’s an entry for #OpenSourceHiggsfield, but mostly it’s the tool I needed. Feedback is very welcome, especially from anyone else iterating on H3 or other video models. How do you keep track of your takes now?
```



document fetched 2026-10-31 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xjh55q/i_lost_track_of_my_takes_after_1200_h3_renders_so
