How about the results with human images as inputs?

[LaLaLailalai]

Hi @haoosz , Thanks for your fantastic work, I‘m curious about the results with human images as input. Could you show more results of human images?

[haoosz]

Thank you for being interested in our work!
I train the model on the images of Gal Gadot and here are some results:

"A cyberpunk photo of S*"

"A photo of S* sitting in a movie theater"

"A photo of S* sitting in the kitchen"

I also upload the training images and pretrained weights and you can try it.

[haoosz]

I have updated the training code. You can train the model on your desired human images now. Thanks!

[Landroval2]

Hello! I am trying to replicate some of the results shown here, but i am not getting good results. Is S* the special token for all the checkpoints provided? Thanks!

[markrmiller]

Not getting good results with a trained model either. Something seems off. The images generated during training look okay, but then those generated by vico_txt2img.py are not even close...

[haoosz]

That's weird. What are your results directly using the pretrained weights?

Put all the pretrained weights under logs/gal_gadot/checkpoints and the training images under images/gal_gadot.
Use the following command to test:

python scripts/vico_txt2img.py --ddim_eta 0.0  --n_samples 4  --n_iter 2  --scale 7.5  --ddim_steps 50  --ckpt_path models/ldm/stable-diffusion-v1/sd-v1-4.ckpt  --image_path images/gal_gadot/1.jpg --ft_path logs/gal_gadot --load_step 399 --prompt "a cyberpunk photo of *" --outdir outputs/gal_gadot

It is supposed to produce similar results as my run above. Please try it out and put your outputs here. I may locate the problem based on that. Thank you.

[haoosz]

Hi @Landroval2,

The pretrained weight file embeddings_gs-STEP.pt is the trained S. It varies among different steps (300, 350, 400). You need to ensure you use the weights at the same step for the image attention module and the S.

[Landroval2]

Hi @haoosz, thanks for your answer! I have been testing this again and getting good images with the gal gadot model. However, the results with the batman model are not great, almost no variability. Could you share some prompts/steps that you used in that case? Thanks again!

[haoosz]

The images of the batman toy are casually self-collected. The results with the batman indeed show low variability using some prompts.

You can try the following prompts (I use the default time step = 400):

• A photo of a S* in the jungle
• A photo of a S* on top of a dirt road
• A photo of a S* among the skyscrapers
• A photo of a S* on top of a wooden floor
• A photo of a S* with a city in the background
• A photo of a S* with a wheat field in the background
• A photo of a S* with the Eiffel Tower in the background
• A photo of a S* on top of green grass with sunflowers around it
• A photo of a S* with Japanese modern city street in the background

Thanks!

[Landroval2]

Thanks for your answer! I will be trying those prompts to see what happens.

[okaris]

@Landroval2 were you able to get better results? Can you share some insights please?

[okaris]

@haoosz I was finally able to get the same results with gal_gadot for inferences. Could you share the training parameters and command for that particular run please?

[markrmiller]

The issue I had was simply not using an identifierKeith . I think I was trying a 3 letter token. When I changed to the same type token in the config currently, everything worked as expected. It almost felt like prompt influence is even worse than TI, but otherwise results were stellar.

[LiGe-In]

Hi @haoosz , Thanks for your great work, I got similar results using the images of Gal Gadot, but got bad results on my own datasets. Is there anything I need to pay attention to when making a data set? Are there any requirements?

[haoosz]

You can try to adjust the training step, the random seed, and the initial word. Besides, the quality of the training data is also important. I have tried on my own images and got reasonable results.