T-DIG:以文字條件提升圖像生成多樣性
T-DIG: Text-Conditioned Diversity for Image Generation
T-DIG 透過指定圖像生成中需要變化的方向,提升文字生成圖像的多樣性。研究團隊以圖像獎勵的回饋訓練大語言模型生成多樣化提示詞;圖像模型保持凍結,這種方法亦適用於 GPT Image 2 和 Gemini。
Hi!
Ask an image model for “a person in a Halloween costume” and you get the same witch four times? Excited to share our work, T-DIG, that improves diversity in text-to-image generation, conditioned on a user's request!
Demo (please give feedback): https://image-study.org/demo/
Gallery of outputs: https://image-study.org/gallery/

T-DIG lets you pick what should vary (“a food pun”, “low-effort costumes for ML researchers”) and generates images that differ along the axis you specify. The image model stays frozen, so it even works on GPT Image 2 and Gemini, with policies trained on open-sourced models!
We train a LLM to generate diverse axis-conditioned prompts using feedback from image rewards. The reward hacking behaviour was my favorite haha. The policy learns to write the caption "inside" the image to make sure it stays aligned to the prompt while increasing diversity.

More details here: https://image-study.org/t-dig/
來源:HuggingFace Daily Papers(社區熱門論文) · huggingface.co