# Bad at video prompts? Let Gemma write them for MiniMax H3-Fast Video generation + 2K Upscale — retained record

> Retained reddit record from reddit:comfyui.

Canonical: https://highsignal.app/data/records/b69656616adb1572

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:comfyui)  published 2026-10-13 09:55 ingested 2026-10-14 02:42

[original source ↗](https://www.reddit.com/r/comfyui/comments/1x4shs4/bad_at_video_prompts_let_gemma_write_them_for/)

## retained text



```
I kept struggling to write good prompts for MiniMax H3, so I built a workflow where Gemma does that part for me. You write a short idea (“a woman reads a letter by the window and starts crying”), and Gemma expands it into a detailed cinematic prompt before generation.

**What it does:**

* Image-to-video with native synced audio (MiniMax H3)
* **Gemma prompt enhancer:** short idea in any **Language**, cinematic prompt out
* Gemma also generates the video title and file name automatically, so your output folder stays organized
* 2K upscale with SeedVR2 that keeps the original audio and fps
* Two-stage pipeline with a latent upscale for sharper detail. Stage 2 runs in 3, 4 or 5 steps depending on speed vs quality
* Optional last-frame control

**Hardware:** tested on an RTX 5090 laptop (24 GB VRAM, 64 GB RAM). A 15-second clip takes about 10 minutes.

Video Tutorial: [https://www.youtube.com/watch?v=8XgMWIo75u8](https://www.youtube.com/watch?v=8XgMWIo75u8)

**FREE WORKFLOW IN THE VIDEO TUTORIAL DESCRIPTION**

Happy to answer questions. I’m curious whether others are using an LLM to write their video prompts, or still writing them by hand.
```



document fetched 2026-10-14 00:42 · canonical https://reddit.com/r/comfyui/comments/1x4shs4/bad_at_video_prompts_let_gemma_write_them_for
