Hardcore Workflow: Generating 3D Gaussian Splats from MiniMax-H3 Videos
questcoast · reddit · 2026-08-09
A developer shared a comprehensive workflow for converting MiniMax-H3 video clips into 3D Gaussian Splat models. The process cleverly integrates multiple AI models with traditional 3D/video processing tools:
- Asset Generation: Used Nano Banana 2 to generate reference images of a car from different angles against a green screen, improving symmetry using a mirror tool.
- Video Transitions: Used the MiniMax-H3 reference model in ComfyUI to generate 3-second transitions between views based on the reference images, upscaling to 1080p.
- Video Processing: Stitched clips in DaVinci Resolve, applied deflicker, denoise, and despill, exporting 1314 frames alongside mask files.
- Alignment & Training: Aligned features in RealityScan, exported in COLMAP format, and trained for 30k iterations using the MCMC strategy in LichtFeld.
- Final Editing: Exported the PLY and performed final edits in SuperSplat.
More from coding & agent
- Multi-Agent Crash: Blind Tool Calls on Missing Context Raise Fault Tolerance Concerns — h34ra · 2026-08-09
- Google Whitepaper Explores the New Paradigm of AI Writing and Reviewing Code — Saboo_Shubham_ · 2026-08-09
- Geek Project: Multi-Model 'Group Chat' Orchestrator with Shared Context Window — aegersz · 2026-08-09
- PixOS: Developer Uses mini-swe-agent to Build Retro Pyxel GUI Harness — babayada · 2026-08-09
- Open-Source Multi-Agent Chat Framework: Agents Decide When to Speak — Emotional_Scale9702 · 2026-08-09
- Claude Connects to Hundreds of MCP Servers: 10 Worth Adding for Video and Design — socialwithaayan · 2026-08-09