Researchers cut the training data for personalized image AI by over 90%

The most expensive part of teaching an image generator to recognize your dog, your face, or your product isn't the training. It's the work that happens before training starts.

A new method called CRAFT, posted to arXiv on 14 August 2026 by a team led by researchers at South Korea's DGIST, replaces the multi-stage curation pipeline that prior subject-personalization systems rely on with an attention-level routing trick. The result: state-of-the-art performance on XVerseBench, a public benchmark for placing multiple specific subjects into AI-generated scenes, using roughly 10,000 reference images, where prior generalized methods needed 150,000 to over 2,000,000 composed-target pairs (arXiv 2608.14403; CRAFT project page).

That's a 15× to 200× reduction in the data the model never has to see, and the paper reports the same recipe transfers to other reference-aware backbones without a rewrite.

What changed

Subject personalization, the task of teaching an image generator to reproduce a specific person, pet, or object in new scenes, has historically been a data-curation problem as much as a modeling problem. To train a model to put your cat into a novel setting, you don't just need pictures of your cat. You need pictures of your cat already placed in many different settings, with the cat correctly labeled in each one.

Building those pairs required what the paper calls a "multi-stage curation pipeline": an LLM generates prompts, a text-to-image model synthesizes candidate scenes, a subject extraction model crops the subject out, a vision-language model filters for quality, and a correspondence labeler ties each image back to the right identity. The output is hundreds of thousands to millions of (reference, target) pairs, and for any new subject, that dataset has to be rebuilt from scratch.

CRAFT skips the pipeline. It works at the level of attention, the mechanism inside the diffusion transformer that decides which parts of a prompt and which parts of the training image should influence which pixels of the output. The method aligns noise-token and phrase-token attention with the correct reference subject during training, then uses per-subject attention masks to gate a pixel-level identity reward so the image-space supervision stays consistent with the attention routing the model is learning (arXiv HTML).

The paper calls this the "Where to look" principle. In practice, the model learns which tokens and which image regions to attend to for each subject, and the identity reward is applied only where the model is already looking at the right thing.

What the team did

The authors fine-tuned FLUX.2-klein-9B, a specific open image-generation model, using low-rank adapters (LoRA), small, cheap fine-tuning modules that sit on top of a larger model without retraining it from scratch. The training signal comes from reward feedback learning (ReFL), a training approach that uses a reward signal to steer the model rather than imitating labeled images. The reported result on XVerseBench is state-of-the-art performance with no composed-target supervision, and the same recipe transfers to other reference-aware backbones with consistent improvements.

What it shows, and what it doesn't

CRAFT is a research result, not a product. The paper has not been peer-reviewed. The result rests on XVerseBench alone, and the authors do not test it at higher resolutions, on larger counts of subjects per scene, or on subjects outside the kinds XVerseBench tests. The 10,000-image training set is the method's reference bank; the reduction is in upstream curation cost, not necessarily in the inference cost a user would see.

The data bottleneck in personalized image AI was curation, not raw data, and attention routing, applied at training time, dissolves that bottleneck without paying the data tax. Whether 10,000 reference images holds at higher resolutions, with more subjects per scene, or on subjects outside XVerseBench's test set is the question the next round of follow-up work will have to answer.

The full paper is at arXiv 2608.14403, with supplementary material on the authors' project page.