No official trainer? We open-sourced a working fine-tuning pipeline for H3
We got supervised fine-tuning working on H3 π ~150-line trainer + latent caching on top of the official Diffusers integration. Wrote up the conventions that silently break training (timestep = 1βΟ, data-ward velocity, dual shifted schedules). Feedback welcome:
https://github.com/IAmIronMan42/MiniMax-H3-FineTuning
When fine-tuning a distilled model, can we achieve the expected results? Dose it degrade the performance of the distilled model? And will image blurriness occur after training?
Yes, fine-tuning the distilled model works β we've validated heads-only and LoRA runs, and a ~30s-clip LoRA run (real audio in the loss) is going smoothly right now.
Two things matter: get the conventions right (timestep = 1βΟ, target = xββΞ΅; see FIXES.md β with the wrong sign the loss rises and the model corrupts fast), and don't over-train: long runs slowly wash out the baked-in guidance, so we keep lr β€ 1e-5, cap steps, and pick checkpoints by visual A/B. No blurriness observed at our scale β it only shows up if you over-train or feed low-res/upscaled targets.
Details in the README; we also released h3-ir-enhancer for the inference side.