Regarding Texture
Deepak2405 opened this issue · comments
How can I improve texture quality for custom images. I did retrain them, but texture quality does not improve much.
Also, the sleeve size is not according to the cloth image, but the sleeve gets adjusted to user input image. Could you suggest how I can make results better?
@Deepak2405 Can you share with me how you managed to execute the dci-vton script? It will be very helpful.
I have tried following the readme file but it's a headache.
Actually, I just followed HR-VITON (ReadMe instructions) and applied it here.
@Deepak2405 How many images did you use to train the DCI-VTON with your own data?
and How many GPUs did you use? I met a problem "CUDA out of memory" although I used 4, then 5 GPUs. Did you meet this problem?
Hope you can respond me.
Thank you so much.
@hoangtdhdhcn we used 2 GPUs (each GPU has 40 GB memory). I did not face any issues as such.
@Deepak2405 yes. Can I ask one more question?
How many images did you train?
Around 2000 I guess. But we started training from VITON-HD weights.
Around 2000 I guess. But we started training from VITON-HD weights.
OK, thank you so much
@Deepak2405 In our subsequent experiments, we found that inpainting mask will have a great impact on the results, so we tried to use HR-VTON's mask selection strategy or a more aggressive strategy, which may help you improve performance. You can refer to #21 (comment)
Thank you. I will check it.
@Deepak2405 Here are a few insights from my experiences that might help enhance the current outcomes:
- Consider altering the masking strategy. The one used in this repository is quite reserved, with fewer masked regions. You might want to look into more assertive approaches, similar to those found in VITON-HD, HR-VITON, and TryOnDiffusion. However, be aware that TryOnDiffusion hasn't published its code, so you would need to develop an implementation on your own.
- Prompt the model by fine-tuning it with your dataset. The released weights of DCI-VTON were trained on a dataset (VTON-HD) comprising just 10k pairs for training, which means there's considerable potential to expand and adapt it to a wider variety of scenarios.
Hello, I have a question I'd like to ask. If training on my own dataset, how should I preprocess the data? For example, with DensePose images, should I retain the IUV format? However, I noticed that in the provided dataset, DensePose images seem to have only the I values. Thank you very much for your response
@Deepak2405 how you managed to execute the dci-vton script? i try several attempts but not success full