THIS TOOK X Hrs 35 Mins TO MAKE ON A RTX 3050 6GB VRAM USING MINIMAX H3
retained text
**60 seconds of AI video, 100% local on an RTX 3050 (6 GB)** *Work in progress, and the quality isn’t where I want it yet. Here’s how it was made:* 1. **Model:** MiniMax H3 reference-to-video, using ComfyUI’s native H3 nodes with the int8 quantized model and the Qwen3-VL text encoder. 2. **References:** two character images, re-used in every shot to keep the characters consistent. 3. **Prompts:** the official six-section reference prompt format (subject definitions, summary, retention analysis, detailed description, soundscape, music). No negative prompts, because H3 doesn’t use them. 4. **Shots:** six shots of 260 frames (10.8 s) at 864×480 and 25 steps, with the res\_multistep sampler and the simple scheduler. 5. **Continuity:** each shot starts from the last 22 frames and audio of the previous one, using the H3 Add Guide node (MiniMaxH3AddGuide) at frame 0. 6. **Automation:** a small Python script sends the six shots to ComfyUI’s API one after another, saves each result, trims the 22-frame overlaps, and encodes one 60-second MP4 with audio. 7. **Runtime:** about 6Hrs 35Mins hours, unattended. Known issues: soft detail, and some flicker. Still tuning.
document fetched 2026-11-12 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xtjanb/this_took_x_hrs_35_mins_to_make_on_a_rtx_3050_6gb