Could you give me the eval scripts on visual conditional controls

[zc1023]

Hi there. I am reproducing these results on ADE20K. However, the results are not as great as described.
My prompt is like ", ". May be something wrong with my scripts? Could you give me the eval scripts?

[staoxiao]

@zc1023 , the prompt format we used to generate images for ADE20K is:
Segmentation to Image\n {text_prompt} <img><|image_1|></img>.
We use the script from https://github.com/liming-ai/ControlNet_Plus_Plus to compute metrics, and use the same text prompt with ControlNet_Plus_Plus: https://huggingface.co/datasets/limingcv/Captioned_ADE20K
Besides, this is the results for https://huggingface.co/Shitao/OmniGen-v0.5. If you use OmniGen-v1, the results should be 40: https://arxiv.org/pdf/2409.11340

[zc1023]

Thank you for your timely reply. @staoxiao I am a little confused that you said that 'use <|image_1|> remove suit, rather than remove suit <|image_1|>.' in inference.ipynb. However, in your eval scripts, you put the textual prompt before the input image. Which one is better?
Also, could you give me the resolution of the input image in your eval scripts?
Thank you!

[staoxiao]

@zc1023 , for edit task, we found <img><|image_1|></img> remove suit is better than remove suit <img><|image_1|></img> . For the visual conditional task, we haven't compared the performance of the two methods, just simply placed it at the end of the prompt. There might be better ways, but I haven't tried them.
The input image size is 512*512.