Training a Cyberpunk LoRA for MiniMax H3 and Fixing Its Joint Audio-Video Code
tetsuoai · x · 2026-08-12
A developer has fine-tuned a LoRA adapter for the MiniMax H3 video model, trained on cyberpunk film data to evoke a strong Ghost in the Shell and Blade Runner aesthetic.
During the process, the author also fixed the training and inference code for H3’s joint audio-video path. This ensures the audio branch is correctly preserved and conditioned during adapter training and generation. The results are looking highly promising, and the author plans to release the LoRA, the training code, and the audio fixes on Hugging Face tomorrow.
More from coding & agent
- Seamless UI Editing: Using Claude Design within Claude Code's In-App Browser — daniel_mac8 · 2026-08-12
- Roboflow's Open-Source Trackers Library Adds McByte for Occlusion Handling — burny_tech · 2026-08-12
- Open-Sourced Hackathon Search Agent Seeks Advice on Evaluation Architecture and Open-Source Models — AggressiveAerie9031 · 2026-08-12
- Ostris AI Toolkit adds support for LTX 2.5 model — ostrisai · 2026-08-12
- GraphQLite: An SQLite Extension Bringing Graph Database Capabilities with Cypher — tom_doerr · 2026-08-12
- Open-Source Terminal TUI to Find Available Single-Word Domains — iannuttall · 2026-08-12