New CRAFT framework drastically cuts data needs for personalized image generation

Researchers have developed CRAFT, a new framework for subject-driven image personalization that significantly reduces the need for extensive training data. Unlike previous methods that require millions of paired reference and composed-target images, CRAFT utilizes a compact set of only 10,000 reference images and subject masks. This approach fine-tunes a reference-aware multimodal diffusion transformer (MMDiT) using LoRA adapters and attention-level rewards, aligning attention with the correct reference subject without requiring composed-target supervision. Applied to the FLUX.2-klein-9B model, CRAFT achieves state-of-the-art results on the XVerseBench benchmark, demonstrating its effectiveness and efficiency. AI

IMPACT Reduces data requirements for personalized image generation, potentially lowering barriers for content creation tools.

RANK_REASON Academic paper introducing a new method and framework for image personalization. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• alphaXiv
• arXiv
• CatalyzeX
• Connected Papers
• DagsHub
• FLUX.2-klein-9B
• Gotit.pub
• Hugging Face
• Litmaps
• LoRA
• MMDiT
• ScienceCast
• scite Smart Citations
• XVerseBench

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →