AI BriefingNVIDIA-HuggingFace Unveil Integrated Diffusion Model Training
Key point
NVIDIA and HuggingFace have unveiled a NeMo Automodel integration for fine-tuning diffusion models.
Details
NVIDIA NeMo Automodel is an open-source training library based on PyTorch DTensor, fully integrated with HuggingFace's Diffusers ecosystem to support large-scale distributed diffusion model training.
Key Features:
- No checkpoint conversion needed: Can be applied directly to any Diffusers model on the HuggingFace Hub
- Scalability: FSDP2, tensor parallelism, pipeline parallelism, etc. can be selected via YAML configuration alone
- Memory efficiency: Supports latent-space VAE caching, multi-resolution bucketing, and activation checkpointing
- LoRA support: Enables both full fine-tuning and parameter-efficient fine-tuning
Supported models: FLUX.1-dev (12B), FLUX.2-dev (32B), Wan 2.1 (1.3B–14B), HunyuanVideo 1.5 (13B), Qwen-Image (20B)
Performance: On 8x H100, full fine-tuning of FLUX.1-dev achieves 0.9 seconds per step, reaching 35.5 images/second. LoRA-based training uses memory in the range of 23–67GB.
Provided under the open-source Apache 2.0 license, with SLURM support and Kubernetes support planned for the future.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Qwen3.5-based Small Decision Model 'Kev' Released with TypeSafe API Compatibility and Performance Metrics