Open sidgairo18 opened 1 month ago
@sharonal10 a response to this would be highly appreciated. Many thanks in advance :)
Hi Siddharta,
We did not release a command for concept compositions. As we use text templates e.g., 'a photo of a
Hope this helps!
Best, Sharon
From: Siddhartha Gairola @.> Sent: Monday, September 23, 2024 4:59 AM To: sharonal10/langint @.> Cc: Sharon Lee @.>; Mention @.> Subject: Re: [sharonal10/langint] Generating New Images (Issue #1)
@sharonal10https://github.com/sharonal10 a response to this would be highly appreciated. Many thanks in advance :)
— Reply to this email directly, view it on GitHubhttps://github.com/sharonal10/langint/issues/1#issuecomment-2368004876, or unsubscribehttps://github.com/notifications/unsubscribe-auth/A3R6FZOG7RCILEOLEQCCDCTZX77CHAVCNFSM6AAAAABNWQDIH2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGNRYGAYDIOBXGY. You are receiving this because you were mentioned.Message ID: @.***>
Hi Sharon,
Thanks a lot for releasing the code.
I was trying to visualize the results in the paper and also generate new samples using the extracted concepts. But currently the readme / code doesn't seem to point obviously to how this can be done.
In the readme it shows how we can extract the 3 concept embeddings / axes from the given image - "python scripts/train_clip_inversion.py -c configs/train_deepfloyd_inversion.yaml -t furniture-three-extract -d 'chair-plastic-gray' training.optimizers.embeddings.kwargs.lr=0.001 shared_tokens=1 gt_init=0 num_placeholder_groups=3 num_placeholder_words=72 real_image='real_imgs/chair-plastic-gray.jpg'"
Is there a way to run the code to generate (i.e. mix and max new concepts / images) as shown in the paper Fig1. and Fig 4. for the task of "Concept Recomposition".
Will be very grateful for your response. Looking forward!
Thanks and best regards, Siddhartha