search

THUDM / ImageReward

[NeurIPS 2023] ImageReward: Learning and Evaluating Human Preferences for Text-to-image Generation
Apache License 2.0
1.19k stars 65 forks source link

推理阶段图像尺寸 #76

Open jianghonglin30 opened 7 months ago

jianghonglin30 commented 7 months ago

你好，我看到ReFL.py在训练阶段要求图像尺寸在512512，那请问在推理阶段是否支持10241024或者更大的尺寸呢？如果不支持的话，我是否可以直接在训练阶段将图像尺寸修改成1024*1024，这样修改会有什么问题吗？

xujz18 commented 7 months ago

您好，感谢关注！ReFL在所给的代码中之所以图像尺寸是512，是因为我们在代码中是对Stable Diffusion v1.4使用ReFL算法进行微调，而SD v1.4支持的分辨率是512.事实上，支持多大分辨率是由文生图基座模型决定的，不是由微调算法例如ReFL决定的。如果要支持1024，您可以尝试SDXL等支持更大分辨率的更新的文生图模型。

The reason why the resolution is 512 in the given code is because we are fine-tuning Stable Diffusion v1.4 with the ReFL algorithm in the code, and the resolution supported by SD v1.4 is 512. In fact, how much resolution is supported is determined by the text-to-image base model, not by fine-tuning algorithms such as ReFL. If you want to support 1024, you can try a newer text-to-image base model such as SDXL that supports larger resolutions.