NVIDIA's NeMo Automodel Now Fine-Tunes Diffusers' Video And Image Models
NeMo Automodel covers FLUX.1, Qwen-Image, Wan and HunyuanVideo, running full and LoRA fine-tunes straight from Hugging Face checkpoints, no conversion required.
- FLUX.1-dev (12B) hits 35.51 images per second in full fine-tuning on 8 H100 GPUs.
- Qwen-Image, a 20B model, reaches 41.21 images per second under the same full fine-tune.
- FLUX.2-dev at 32B and Wan 2.2's 27B video model both get LoRA recipes at rank 64.
Why it matters: Fine-tuning a 32 billion parameter model still means renting Nvidia's own hardware to do it.
Hugging Face blog (NVIDIA/Hugging Face teams) ↗ · Jul 17, 20267/17/26 · ✓ Checked✓ Check