For years, LoRA has been the default way creators fine-tune image and language models without renting a data center. A new Hugging Face benchmark published June 18, 2026 asks a pointed question: is the most popular technique actually the best one? For image generation, the answer is no.

What Happened

Hugging Face engineers Benjamin Bossan, Sayak Paul, Marian, and Kashif Rasul ran a head-to-head test of LoRA against a roster of newer parameter-efficient fine-tuning (PEFT) methods: BEFT, Lily, LoRA-FA, OFT, DoRA, rs-LoRA, AdaLoRA, LoHa, GraLoRA, and Cartridges. The motivation is the sheer dominance of one method. Of the model cards on the Hub that mention a single PEFT technique, 20,509 out of 20,834 name LoRA, a 98.4 percent share. Among image-generation checkpoints, 7,111 of 10,000 were LoRAs. The original LoRA paper is now four years old, and the field has moved well past it.

Why It Matters

For creators training custom styles or characters, the headline result is that OFT (Orthogonal Fine-Tuning) beat LoRA on image generation. OFT reached a similarity score of 0.708 versus LoRA's 0.697, and it did so using less memory: 9.01 GB against 9.97 GB. That combination of higher fidelity and a smaller footprint is exactly what matters when you are training on a single consumer GPU, the same constraint behind tools like LoRA studios that fit on 16GB cards. On math reasoning, Lily edged LoRA on accuracy (54.9 versus 53.2 percent), while BEFT traded accuracy for a much lighter memory budget.

Key Details

The benchmark is not a blanket "switch away from LoRA" verdict. LoRA stayed competitive and easy to use across tasks, and no single method won everywhere. The practical unlock is that the PEFT library now lets you swap techniques by changing only the configuration object, so testing OFT against LoRA is a one-line change rather than a rewrite. PEFT also added the ability to convert other adapters into LoRA format for inference tools like vLLM, with the team reporting minimal quality loss in the conversion (similarity moved from 0.702 to 0.694). That keeps your deployment pipeline intact even if you train with something other than LoRA.

What to Do Next

If you train image LoRAs for Flux, SDXL, or video models such as NVIDIA Cosmos Predict 2.5, run OFT as a side-by-side on your next dataset and compare similarity and VRAM before committing. Start from the PEFT documentation to check which methods your base model supports. The takeaway is simple: the default is not automatically the best fit, and the cost of checking is now one config line.