Qwen-Image 2.1 LoRA training plagued by anatomy failures, community hunts for fixes
ResidentFrame4195 · reddit · 2026-09-21
A trainer spent hours fighting Qwen-Image 2.1 character LoRAs in ai-toolkit and reports severe anatomy failures: melted hands, extra limbs, double mouths, waxy skin and grotesque mutations — even at bf16 with zero quantization. Degeneration clusters around the 200-1500-1750 step mark before samples partially recover, then regress. The author suspects damaged base priors and is testing a corrective setup: heavier dataset, high-res buckets, quantized base, guidance loss and blank prompt preservation, with only slight improvement so far. Almost no usable LoRAs have been posted publicly, suggesting the community is all hitting the same wall; the post asks for working rank, lr, dataset size and resolution settings.
More from Multimodal
- Open-source full-duplex speech model caramel-1: 413M params, sounds human but can't hold a conversation — alexcovo_eth · 2026-11-12
- H3-Image touts 16MP generation and realtime inference via 2-step LoRA on ZeroGPU — linoy_tsaban · 2026-11-12
- MiniMax H3 Image renders 4K portrait in half the time Seedream 5.5 takes for 2K — heyronir · 2026-11-12
- Endless screensaver of self-assembling quantum cryostats, made with Gemini 4 Carbon — hapticdata · 2026-11-12
- Apple's ADM 3 Cloud image model ranks #42 in editing and #85 in text-to-image on Artificial Analysis arena — ArtificialAnlys · 2026-11-12
- Adobe Firefly debuts Story Builder tool for streamlined visual storytelling — LudovicCreator · 2026-11-12