diffsynth

MiniMax H3 多模态 LoRA 训练、审计与交付技能：Ref2VA、联合音视频、低显存后端与十步证据链。

Train a Qwen-Image-2.1 LoRA on a 12 GB GPU and 32 GB of RAM — NF4 weights kept in VRAM, no CPU offload, 500 steps in 12 minutes. Unofficial — not affiliated with Alibaba or Tongyi.

Cog Single GPU Quantized Implementation of Step-Video-T2V

Single-GPU full fine-tuning & LoRA of Krea 2 Raw (12.8B) for an anime illustration style: training code, AdafactorSR (stochastic rounding), GGUF quantization, 16 GB ComfyUI tests, technical report