search

Karine-Huang / T2I-CompBench

[Neurips 2023] T2I-CompBench: A Comprehensive Benchmark for Open-world Compositional Text-to-image Generation
https://arxiv.org/pdf/2307.06350.pdf
MIT License
168 stars 5 forks source link

Data Details When Finetune #16

Open zhuangshaobin opened 2 months ago

zhuangshaobin commented 2 months ago

Thanks for your excellent work. I encountered a small problem. When finetune, do I need to retrain a lora for each test class such as color, or should I mix all the fake images together to train a lora?

Karine-Huang commented 2 weeks ago

Thank you for your kind words! In our work, we fine-tune a separate LoRA for each specific category, such as color, shape, etc. Of course you may mix all the generated images to train a lora.