MiniMax H3 ecosystem roundup: Fizgig 4.1.2 trains character+voice LoRAs in one run
optimisticalish · reddit · 2026-08-19
A MiniMax H3 community roundup:
- Fizgig 4.1.2 (LoRA trainer): can now train a single character+voice LoRA from photos, video clips and voice recordings in one run, locally in 16GB VRAM; fixed Turbo/official-workflow compatibility and passed a security audit. A 20-minute YouTube tutorial is available.
- ComfyUI-MiniMaxH3-SingleFrame: two still-image nodes, notably interpolating a plausible 'middle frame' between given first/last frames, no Core patching or special VAE needed.
- H3 Prompt Journal: prompt-writing demos for complex camera physics — three-person occlusion-linked orbital long takes, dual-subject speed contrast, single-subject three-pose sweeps.
- Video → H3 Prompt: end-to-end pipeline turning a video file into a paste-ready H3 prompt, packaged as a Codex skill.
- MiniMax Music: new rvq-encoder-169m ONNX encoder (676MB) converts audio back into MiniMax generation codes, enabling track continuation, section replacement, or re-performing the same song.
- RTX 3060 benchmark: testing the official reference-to-video workflow on 12GB; key caveat — 'two 3060s are still not one big card', with parallel jobs bottlenecked by system RAM, not VRAM.
More from coding & agent
- Kinex: a harness that connects coding agents to any physical robot for scalable learning — daibond_alpha · 2026-10-16
- Every publishes 8,000-word Codex oral history, agent-native so your AI can dig the raw interviews — danshipper · 2026-10-16
- Developer rebuilds GitHub in 4 days: Keel makes agents declare intent before changing code — threepointone · 2026-10-16
- Stop using small models as judges with overloaded context, dev warns — xeophon · 2026-10-16
- Meta's open-weight Muse Spark Open hits Ollama with 1M-token context and Claude Code support — ollama · 2026-10-16
- Monid launches: one balance, pay-per-call access to 1,700+ tools for AI agents — HeyAmit_ · 2026-10-16