Subject LoRAs on Minimax H3 Image learn poor likenesses while identical datasets excel on Qwen 2.1
Ad_Organic · reddit · 2026-11-16
A Reddit user reports systematic failures training Subject LoRAs on Minimax H3 Image:
- Tested ranks 16–64, optimizers from AdamW to Automagic3/4, both AI Toolkit and Fizgig, and datasets of all sizes — the model consistently learns an approximation rather than the actual subject, and more epochs don't help.
- Key control: the exact same datasets yield excellent LoRAs on Qwen 2.1, Krea 2 and Moss; even the Minimax H3 video model learns subjects better from still images.
- H3 also fails to learn motion, whether trained on stills or video clips.
- The author asks whether anyone has achieved high-fidelity likenesses, and what settings or dataset strategies make a difference.
For image-LoRA practitioners, this is first-hand negative evidence pointing to a possible systemic issue in H3 Image's architecture or training pipeline.
More from Multimodal
- Testing Qwen Image 2.1 and Minimax H3 on anime-to-photo conversion and film photo restoration — Ant_6431 · 2026-11-16
- Second take of a 'Cubee run' video generated with Minimax H3 — JahJedi · 2026-11-16
- Non-artist builds 14-chapter hand-painted storybook game with Fable 5.2 in 4 hours — jradoff · 2026-11-16
- Seedance-generated anime train scene stuns Reddit with cinematic quality — Ok-Solution-7889 · 2026-11-16
- AI-generated 'Men in Black' reimagined in 1977 goes viral — techhalla · 2026-11-16
- Hinglish, spoken by 350M, is the real test: Sarvam models power full voice agent pipeline — itsOmSarraf_ · 2026-11-16