Generate only 1 image

[luchoster]

Would it be possible to generate only 1 image? similar to what thispersondoesnotexist.com does it?

I'm still training atm, but I figured the end result of the generation of images, is 8x8 collage of images, am I right?

[lucidrains]

@luchoster yea! i had no idea somebody would get this far! I can offer another flag? --num-image-tiles ? Or how do you think it should be named?

[luchoster]

I think that flag name makes sense, that'd be awesome!!!

I'm 47% into the training, so it'd be awesome to have that option. Also, how long does it take to generate an/the images? I would love to build a clone of thispersondoesnotexist :)

[lucidrains]

@luchoster if you have a GPU, a single image would be less than 100ms. It'll be very fast! What are you training it on? Is it learning?

[luchoster]

ooo that's goood!
I have a windows machine for this.

Here are some pictures of the training.

[luchoster]

the last 2 are the most recent ones, 46 and 47 %. As you can see, it started to generate nice faces and in the last 2, started to get weird again haha

[lucidrains]

@luchoster very cool! when the discriminator loss hits 0 consistently, D: {num} in the logs, the training is done, and you usually have to pick the generator just before it explodes

[luchoster]

oh that's good to know. I'll keep an eye on that.

[lucidrains]

@luchoster ok, I just updated the package with the new flag! just set it to 1 and you are all set

[luchoster]

you are awesome!!!!
Thank you! I'll keep you posted on the results :)

[lucidrains]

@luchoster please do!

[tjukanovt]

@lucidrains I am also trying to create only a single image as in the original issue, but can't seem to get this working. I am running command stylegan2_pytorch --data /mydata/ --name mapproject --results_dir /myresults/ --models_dir /mymodels/ --num_image_tiles 1

also tried with --num_image_tiles 1 and modified all the num_image_tiles = 8 values to 1 in the Python script, but nothing seems to change the outputs.

Still I am not getting a single image, but a mosaic instead.

[HeleenE]

Oeh I'm super interested in this new flag as well! Been training the software on a very small set of drawings, but with surprisingly good results. It would be nice to create some higher res versions. Now running the training again with the image-size 256 flag, going to take some time 😅

I'm running it on Google Colab, but it hasn't been super stable. Sometimes it crashes right before the 10000 save. If I lower that number in the code would it also create more models?
SAVE_EVERY = 10000

EDIT: I'll create a new issue for that last question in case someone else goes looking for it in the future