NVIDIA NeMo Automodel Fine-Tunes Any Diffusers Model

The Conversion Tax on Diffusion Fine-Tuning Just Got Cheaper

Anyone who has tried to fine-tune an image or video model at scale knows the tax: before you train anything, you convert checkpoints, rewrite the model definition to suit your training framework, and then convert the result back so it will load in a normal inference pipeline. On July 17, 2026, NVIDIA and Hugging Face shipped an integration that removes that step. NeMo Automodel can now fine-tune any Diffusers-format model straight from the Hub — no checkpoint conversion, no model rewrites — and the resulting checkpoint loads directly into a standard DiffusionPipeline.

• Supports FLUX.1-dev (12B), FLUX.2-dev (32B), Qwen-Image (20B), Wan 2.1 T2V (1.3B–14B) and HunyuanVideo 1.5 (13B)
• Both full fine-tuning and LoRA are supported, using FSDP2, tensor parallelism and multiresolution bucketing
• NVIDIA's published throughput on 8× H100 80GB at 512×512: FLUX.1-dev LoRA at 53.73 images/sec, full fine-tune at 35.51 images/sec
• Video throughput at 512×512×49 frames: Wan 2.1 1.3B at 8.50 clips/sec, HunyuanVideo 1.5 at 1.35 clips/sec

What Do the Benchmark Numbers Actually Say?

NVIDIA's published figures, measured on eight H100 80GB GPUs at 512×512 resolution, put FLUX.1-dev full fine-tuning at 35.51 images per second — 0.902 seconds per step — with the LoRA path reaching 53.73 images per second. Qwen-Image full fine-tuning lands at 41.21 images per second. On the video side, Wan 2.1 at 1.3B parameters manages 8.50 clips per second, the 14B LoRA configuration drops to 2.11, and HunyuanVideo 1.5 sits at 1.35.

A fair caveat: these numbers come from NVIDIA and Hugging Face's own writeup rather than independent benchmarking, and 512×512 is a friendly resolution. Treat them as a vendor-published baseline, not a verdict. What they do establish is the shape of the thing — full fine-tunes and LoRA runs on 12B-to-32B image models are now an eight-GPU job with published, reproducible configurations rather than a bespoke engineering project.

Why Removing Conversion Steps Matters More Than It Sounds

Every format conversion in a machine learning pipeline is a place where fidelity quietly degrades and where reproducibility goes to die. When the training framework and the inference format are the same, the loop tightens: fine-tune, load, evaluate, iterate. The multiresolution bucketing support is the other practical win, since real training datasets are almost never uniformly sized and padding everything to a square wastes compute.

The combination of FSDP2 and tensor parallelism is what makes the larger models tractable. FLUX.2-dev at 32B parameters is not something you shard casually, and having a maintained, open configuration for it is worth considerably more than another point on a benchmark chart.

An Open-Weights Ecosystem That Keeps Getting More Usable

This lands in a month that has been unusually good for people who actually build on open models — see our coverage of Kimi K3, the largest open-weight model released so far and PyTorch 2.13's FlexAttention work on Apple Silicon. The pattern across all three is the same: the frontier is not the only thing moving. The tooling underneath it is getting steadily less painful, and that is what determines whether open weights translate into open practice.

For teams doing custom image and video generation, the practical takeaway is narrow but real. The path from a Hub checkpoint to a fine-tuned model you can serve is now shorter by several steps, and those steps were the annoying ones. More on model tooling in our AI section.

Sources: Hugging Face Blog — NVIDIA — July 17, 2026; NVIDIA NeMo Automodel on GitHub; NVIDIA NeMo Automodel documentation.