H3 Character LoRA Training: Likeness Far Behind Wan 2.2, Tips and Pitfalls Shared
Tiny-Highlight-9180 · reddit · 2026-09-09
A Reddit user shares experience training an H3 character LoRA with disappointing results. Using an RTX 5090, musubi-tuner, 60 images, 3000 steps (5 hours). Key hyperparameters: networkdim/alpha=16, LR 1e-6, sigmoid timestep sampling, EMA decay 0.999, blockstoswap=38. Testing 1500 vs 3000 steps shows 1500 is insufficient; EMA vs non-EMA made little difference. Other findings: fl2va and ref2va are separate, LoRAs don't transfer; enabling usepinnedmemoryforblockswap cut iteration time from 11.6s to 6.1s. The author notes the same dataset produced much better results on Wan 2.2 and seeks advice.
More from Multimodal
- AI filmmaker PJaccetturo discusses building his NEXUS universe with Dreamina — DavidmComfort · 2026-10-27
- Runway launches App Studio: build no-code creative apps that generate image, video and audio — runwayml · 2026-10-27
- MiniMax H3 runs Maestro locally for free; Kling 4.0 vs Seedance 2.5 video editing compared — cocktailpeanut · 2026-10-27
- AI-generated movie-grade clip goes viral, blurring the line with real cinema — cneuralnetwork · 2026-10-27
- 10 practical ways to run Astra computer use on Premiere Pro for video editing — beechinour · 2026-10-27
- MiniMax H3 runs fully on a Galaxy S26: 5-second video with audio in ~7 minutes — BidFunny4862 · 2026-10-27