Underwater AI music video with MiniMax H3: full ComfyUI two-pass workflow released
Cheap_Credit_3957 · reddit · 2026-11-22
Creator vrgamegirl19 used her open-source ComfyUI music-video builder to produce a full underwater AI music video with MiniMax H3. Despite the model likely never seeing underwater singing, environments, lighting and motion came out well, with only occasional lip-sync glitches.
- Output at 2K 1920x1080 via two-pass sampling: Pass 1 with resmultistep, 25 steps; Pass 2 adds latent upscaling plus a rank-16 refiner LoRA at strength 1.0, 3 steps, denoise 0.20
- The shareable workflow JSON, ComfyUI nodes, setup tutorial and Discord community are all on GitHub
More from Multimodal
- Two AI models duel in Super Smash Bros using only game frames and telemetry — anpaure · 2026-11-22
- AI Film Team Open-Sources 50,000-Character Claude Skill for Anamorphic Shots — jamestagg · 2026-11-22
- Fuzzy-textured Christmas AI imagery exploration — hewarsaber · 2026-11-22
- MiniMax H3 voice-reference dialog sounds clippy; low-res seed workflow shared — sitefall · 2026-11-22
- Microsoft patents AI that rewrites livestream speech and lip movements in near real time — TinfoilTricorn · 2026-11-22
- Blind-taste-test workflow: two Gemini models grade Claude Code's video edits — aryanXmahajan · 2026-11-22