Hi, thanks for the great work!

I tried to run the code by following the instructions in the readme file
however, I faced the error below, which seems that 'timm' package is required.

Traceback (most recent call last):
  File "/root/miniconda3/envs/tada/lib/python3.9/runpy.py", line 197, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "/root/miniconda3/envs/tada/lib/python3.9/runpy.py", line 87, in _run_code
    exec(code, run_globals)
  File "/source/sihun/TADA/apps/run.py", line 7, in <module>
    from lib.trainer import *
  File "/source/sihun/TADA/lib/trainer.py", line 21, in <module>
    from lib.dpt import DepthNormalEstimation
  File "/source/sihun/TADA/lib/dpt.py", line 8, in <module>
    import timm
ModuleNotFoundError: No module named 'timm'

I was able to solve it and run the code by simply installing timm by pip
$ pip install timm
Although this is a simple task, I thought it would be good to add it to the requirements.txt

This is an amazing work! I have played with it for a while but I cannot generate skirt or dress as shown in the paper. Could you please give me some hints? Thanks in advance!

I tried to find how to implement the interpolated latent code of Equation(6) in paper, the z = alpha z^I + (1-alpha)z^N.
I have read the code, it seems that you just use image latent code and normal latent code for loss separately? Maybe it's a Equivalent operation?

    def train_step(self, data, is_full_body):
           ......
           loss = self.guidance.train_step(dir_text_z, image_annel).mean()
            if not self.dpt: 
                # normal sds
                loss += self.guidance.train_step(dir_text_z, normal).mean()
                # latent mean sds
                # loss += self.guidance.train_step(dir_text_z, torch.cat([normal, image.detach()])).mean() * 0.1
            else:
                if p_iter < 0.3 or random.random() < 0.5: #  
                    # normal sds
                    loss += self.guidance.train_step(dir_text_z, normal).mean() # use normal map directly
                elif self.dpt is not None :
                    # normal image loss
                    dpt_normal = self.dpt(image) # estimate normal map
                    dpt_normal = (1 - dpt_normal) * alpha + (1 - alpha)

                    lambda_normal = self.opt.lambda_normal * min(1, self.global_step / self.opt.iters)
                    loss += lambda_normal * (1 - F.cosine_similarity(normal, dpt_normal).mean())

Waiting for your reply, thanks a lot!

run 'python -m apps.run --config configs/configs/tada_w_dpt.yaml.yaml --text "Aladdin in Aladdin" '
AttributeError: 'Trainer' object has no attribute 'log_ptr'

Thank for your wonderful work!
I have a question about how to animate the result? The final training result in this work seems to be mesh. But i dont know how to use motion-diffusion-model or priorMDM to animate/drive the result. Could you please give me some advice ?
Thank you so much～～～

您好，我现在希望在您的代码的基础上修改一下，让他只生成一个人脸的模型（不要身体的部分）
请问可以通过修改这里得到吗：

Thanks for your amazing job!
I would like to know how many GPUs are needed to reproduce the result.
Thank you!

Hi，what is the correct project structure？Can you offer a picture of this repo？I followed your installment instruction，but i can run it.

How to extract the smpl rotations to driven mixamo etc characters?

It seems that directly optimizing the texture map may cause some noise. Why not optimize MLPTexure3D to get mesh texture? Have you done some experiments about that since you have implemented the choice of MLPTexture3D.

Hi tingting,

Thanks for your excellent work. I wonder to know where I can find generated mesh examples.

Can I get a mesh with color and texture?

Trying to execute the code on my RTX 3060 and unfortunately I get an error of insufficient memory (OOM). Is it possible to optimize the generation so that everything works as it should?
Or maybe it is possible to run on CPU?

Thank you for your great work! Could you please upload one sample file like sample00_rep00_smpl_params.npy? I have some difficulties when generating motion using MDM and I hope to inference your anime.py first. Thank you very much.

Thank you for your work! @YuliangXiu , I am trying to run the Animation demo. May I know where to download the files for ckpt_file below? I think it's the path to the pre-trained characters. I have downloaded TADA Extra Data and unzip it as ./data, but there is no checkpoints in it. And the code breaks at assert os.path.exists(ckpt_file) (L404 of anime.py).

Also, I have download TADA 100 Characters and there is also no checkpoints, but there is a params.pt in each character's folder. However the key in these .pt files dict_keys(['v_offsets', 'betas', 'expression']) do not match the ones in the code. There should be a 'raw_albedo' key in addition to these, according to your code.

As the paper said, TADA densify the smplx by Adaptive Upsampling and Interpolating Skinning Weights. But in the code part( class DLMesh), I only find the subdivision of smplx 3d point. May I ask is there any part of code includes densifying the smplx? Or which part of code does the work of Adaptive Upsampling and Interpolating Skinning Weights?
Thank you.

Using your method, I have played around well with your model.
I have one question.
Is it absolutely necessary to go through the training process to create new character?
Is it possible to create new character's 3d mesh without undergoing the training process?

I attempted to integrate the stable Diffusion XL model into the TADA pipeline to evaluate its performance and compare results with sd 2.1. However, upon implementation, I encountered significant challenges. The training process with sd-XL appears to converge very slowly, if at all, unlike the smoother convergence observed with sd 1.5 or 2.1.

I'm reaching out to inquire if there are specific reasons why TADA may not be performing well with sd-XL. Could there be particular components or configurations within the pipeline that are not optimized for sd-XL, leading to this discrepancy in convergence rates? Any insights or suggestions on how to address this issue would be greatly appreciated.