Evaluation metrics(FID) #124
Comments
|
Hi @Zhu1116 , May I check the inference code you used? Thanks! |
for sent_key,sent_info in tqdm(content):
imgid=str(sent_key).zfill(12)
img_prompt_path = os.path.join(condition_image_folder_path, f"{imgid}.jpg")
image = Image.open(img_prompt_path).convert("RGB")
w, h, min_dim = image.size + (min(image.size),)
image = image.crop(
((w - min_dim) // 2, (h - min_dim) // 2, (w + min_dim) // 2, (h + min_dim) // 2)
).resize((512, 512))
condition = Condition("canny", image)
result_img = generate(
pipe,
prompt=sent_info[0],
conditions=[condition],
).images[0]
result_img.save(os.path.join(save_dir_gen, f"{imgid}.jpg"))
image.save(os.path.join(save_dir_resized, f"{imgid}.jpg"))
condition.condition.save(os.path.join(save_dir_canny, f"{imgid}.jpg"))like this，prompts and images are from coco val dataset |
|
Almostly copy from examples/spatial.ipynb |
|
Hi @Zhu1116 , The inference code looks good and the only thing different from our implementation is that there is an extra central croping. But I dont think it will cause such a performance drop.
|
|
oh, use the original coco images? I use the central cropped 512x512 images |
|
No. Actually you should use the central cropped images. |
|
That's strange. The quality of the generated images is good. May I have a copy of your FID evaluation code? |
|
Could you please your email for further discussion? |
|
OK, OK. Thank you very much for your careful reply. |
|
Sure, I'd be more than happy to. 246939556@qq.com |
|
sorry, email is wrong. 2469395556@qq.com |
Great work! I have a question here. When I was replicating the paper, under the conditional control of Canny, I generated 1000 images. Then I used the following code to calculate the FID metric and got a result of around 70, which is quite different from the result of more than twenty in the paper. May I ask what the reason for this is?
The text was updated successfully, but these errors were encountered: