austinxy / giraffehd Goto Github PK

Thank you for releasing the code. Why do I get an error when I set -- CONTROL_i to 10 and 11?
File "eval.py", line 120, in eval img_rep[args.control_i] = p IndexError: list assignment index out of range

Hi! Thank you for the awesome work and providing codes.

I have a somewhat simple question about measuring FID score on FFHQ 256x256.

Which one of the two processes below do you recommend to get a great FID score on FFHQ 256x256?

1. Generate 1024x1024 resolution images and resize them to 256x256 and calculate the resized 256x256 fake images on FID score.
2. Generate images directly to 256x256 resolution and calculate the 256x256 fake images on FID score.

Thank you!

Hello,
Thanks for the repo!

I'm trying to train GIRAFFE HD on a version of CompCar with the removed background at 256x256 resolution. Based on what I read in this other issue, I trained the model for 50k iterations on 8 GPUs with batch size 32, which should be equivalent in terms of total images. However, my generated images are not looking good. These are some examples of my last checkpoint

Do you think this is a problem with the number of iterations and I should continue training for a longer time or is something else in the training config as well?

Thanks in advance!

Hi,

Nice work and congrats!

I'm trying to use your work as a baseline to my own dataset. However I find the training extremely slow.

Here's my training script:

python -m torch.distributed.launch --nproc_per_node=6 --master_port=$RANDOM train.py --wandb --batch 32 --dataset hairset --size 128 --datasize 128

With 6 V100, it is showing more than 900 hrs estimated time:

Is it normal? What's your training strategy (epoch, batch size, etc.) for FFHQ-like dataset? Thank you.

Hi!
Thank you for the awesome work and the code.

Where can I find TwoStageGenerator_Refine class in model.py?
I was reading the codes in calc_mbs.py

I recently went through the GIRAFFEHD research paper and I'm impressed. Thanks for providing the code.

I am having few questions, you can help me with.

• Can you share with me the minimum number of images that need to be used for GIRAFFEHD model training to get a good result?

• For training on the custom dataset, do the images need to be present in different variations of a single object as present in the CompCar dataset? I mean an object image taken from different angles, with different colors, or at different times. Or a custom dataset containing random images of a single object, which may not be present with different variations, can also be used to get the same results as the product using the CompCar dataset?