H3-Image LoRA training fails with blocky artifacts under W4A8 quantization on 16GB VRAM
IDonLikeAds · reddit · 2026-11-15
An experienced author (hundreds of LoRAs trained) fails to train a MiniMax H3-Image LoRA in AI-Toolkit: constrained to 16GB VRAM and 64GB RAM, they must use W4A8 quantization to avoid OOM, but the result is full of large blocky artifacts unlike anything they've seen.
Trained with Sigmoid it breaks; they're now trying the default Shift activation and suspect A4A8 is simply too degraded for training. Asking the community for similar experiences and fixes.
More from Multimodal
- Early Seedream 5.5 tests impress: fashion-grade portrait generations with shared prompts — LudovicCreator · 2026-11-15
- Kimodo runs text-to-3D-motion locally on iPad at 30fps, clips up to 10 seconds — Swaggsz · 2026-11-15
- speech.cpp v0.11.0 lands Haru-4B, a full-duplex voice chat model running locally in C++ — alexcovo_eth · 2026-11-15
- Everyone uses Fable 5.2 to code, almost nobody for UGC ads — _jaydeepkarale · 2026-11-15
- Meituan open-sources LongCat-Video-Avatar 1.6 for 10+ minute talking-avatar videos — pmttyji · 2026-11-15
- Change one thing at a time: why iterating AI images isn't design improvement — alifcoder · 2026-11-15