clarification for Forgedit: text guided image editing via learning and forgetting

[witcherofresearch]

First of all, congratulations for such an awesome and complete image editing survey and special thanks for including our paper Forgedit: text guided image editing via learning and forgetting in the survey. I am the first author of this paper and I think there might be some misunderstandings of our method in Table 1 of this survey and mistakes for editing results presented Figure 13.

First, our Forgedit was designed to tackle general text-guided image editings so in fact Forgedit is capable of conducting most tasks in Table 1, which I will show in the following examples.

Second, the editing results with Forgedit in Figure 13 is incorrect. Recently I have refined the Forgedit code to make it easier to reproduce our results. Next, I will show the results of all editing examples from editeval-v1 in Figure 13 from your paper and list the hyperparameters to reproduce our results. For fair comparison, all results are obtained with Stable Diffusion 1.4. The success rate and editing quality could be improved if the base model is switched to realistic vision series. Only target prompt and input image of editeval-v1 are used since Forgedit could use BLIP to generate source prompt.

Editing type: action change
Input Image, target prompt='A polar bear raising its hand'

forgedit command:
accelerate launch src/sample_forgedit_batch_textencoder.py --train=True --edit=True --save=True --forget='encoderattn' --interpolation=vs --gammastart=11 --gammaend=15 --numtest=7

editing type: object addition
Input Image, target prompt='A glass of milk next to a stack of cookies on a wooden board with a gray background'

forgedit command
accelerate launch src/sample_forgedit_batch_textencoder.py --train=True --edit=True --save=True --forget='donotforget' --interpolation=vs --gammastart=13 --gammaend=15 --numtest=7

editing type: object removal
input image, target prompt='A mountain lake'

forgedit command:
accelerate launch src/sample_forgedit_batch_textencoder.py --train=True --edit=True --save=True --forget='donotforget' --interpolation=vs --gammastart=15 --gammaend=18 --numtest=4

editing type: object replacement
input image, target prompt='A floor lamp standing next to a potted plant in a cozy room'

forgedit command
accelerate launch src/sample_forgedit_batch_textencoder.py --train=True --edit=True --save=True --forget='donotforget' --interpolation=vs --gammastart=11 --gammaend=15 --numtest=7

editing type: background change
input image, target prompt='A silver car parked at a dense jungle'

forgedti command
accelerate launch src/sample_forgedit_batch_textencoder.py --train=True --edit=True --save=True --forget='encoderattn+encoder1' --interpolation=vs --gammastart=12 --gammaend=15 --numtest=4

editing type: style change
input image, target prompt='A Van Gogh style painting of a light house sitting on a cliff next to the ocean'

forgedit command
accelerate launch src/sample_forgedit_batch_textencoder.py --train=True --edit=True --save=True --forget='donotforget' --interpolation=vs --gammastart=13 --gammaend=15 --numtest=7

editing type: texture change
input image, target prompt='A statue of a horse running in a field'

forgedit command
accelerate launch src/sample_forgedit_batch_textencoder.py --train=True --edit=True --save=True --forget='decoderattn+decoder2' --interpolation=vs --gammastart=13 --gammaend=17 --numtest=4

For emotion expression editing, Forgedit could tackle it too.
Input image, target prompt='a smiling man and a smiling woman'

Here I switch to use realistic vision for human editing, yet I think Stable Diffusion 1.4 should be working too.
forgedit command
accelerate launch src/sample_forgedit_batch_textencoder.py --train=True --edit=True --save=True --forget='donotforget' --interpolation=vs --targeth=768 --targetw=768 --gammastart=8 --gammaend=11

for object move and object size change, there are multiple cases in TEdBench, another text-guided image editing benchmark from Google. Our forgedit could tackle these cases too. The results could be found in Forgedit TEdBench.

Finally, if you have any difficulties reproducing Forgedit's results on editeval, feel free to leave a comment or contact me via email. It would be great if the editing results of Forgedit could be corrected in the next version of this survey. Thanks again.

[MingfuYAN]

Thanks for your information! We will update Table 1 and the experimental results in the next version of the paper.