Recreating Retro Anime OP with MiniMax H3: 15-min Generation on RTX 5090
GamerVick · reddit · 2026-08-09
A creator shared a workflow using the MiniMax H3 model to produce a retro Japanese anime-style opening sequence (inspired by Naruto). The process combines multi-image references, a structured reference-video timeline, and music from Suno, achieving a stylized graphic noir/jazz aesthetic.
The author also highlighted a local inference bottleneck: generating a 15-second 720p video takes about 15 minutes on an RTX 5090 with 64GB RAM. Despite enabling sol attention and sage attention optimizations, the speed remains a hurdle. The creator avoided Turbo LoRA to prevent quality degradation and asked the community for further speed-up solutions.
More from Infra
- Epoch AI: Global AI Chips to Reach 200 Million by 2028, Doubling Every 9 Months — nordicinst · 2026-08-09
- AI Hardware Debate: Bandwidth Matters More Than Compute? RTX 5090 Over DGX Spark — AccBalanced · 2026-08-09
- OpenAI Exec Predicts Local AI Agents Will Be Obsolete in Months, Replaced by Cloud Infrastructure — 机器之心 · 2026-08-09
- MFU as an AI Compute Metric Was Partly Marketing, HFU Faded Due to Awkward Acronym — cHHillee · 2026-08-09
- Utah's 16GW Data Center Sparks Concerns Over Severe Heat and Water Impact — aakashgupta · 2026-08-09
- Floci UI: Open-Source Unified Web Console for Local Multi-Cloud Environments — tom_doerr · 2026-08-09