Spaces:

CVPR

/

Text2Human

like 45

Runtime error

Hi @radames

Many thanks for implementing the function of drawing the segmentation mask by hand! That's so cool!

In the UI demo, we have the function of manipulating the segmentation mask by hand. However, in the Hugging Face demo, there lacks such a function currently. You can definitely publish the drawing component as a Hugging Face Space. This will make the Hugging Face demo more complete and more interesting to use!

For the issue of publishing the drawing component, may I ask about the format? Will you merge the drawing component into this Hugging Face space or you just open a new space?

Looking forward to your reply! Thanks again for your contribution to making this Text2Human demo more interactive and interesting!

@radames Wow, it's so cool. Nice work!

Hi @yumingj , thanks for your prompt response

Unfortunately, we can't integrate a custom component to Gradio yet. A plugin like is in the roadmap. For now, I'll publish the drawing component under a new static Space and add a description and link to the complete experience here. What do you think?

Still, it will be using the Gradio API running on this Space or maybe on @hysts . I could also clone this Space if we get too many hits and longer queues here. I don't want to overwhelm your demo.

One question: considering that textures descriptions are limited to the five textures, and you mention the format as
<upper clothing texture>, <lower clothing texture>, <outer clothing texture>

Could I create three select lists with the five possible textures, or are there more natural language possibilities for the texture description?

Thanks again

Hi @radames

Sure. You can publish the drawing component under a new static and add the descriptions and links here.

As for which Gradio API to use, either this Space or @hysts is okay.

For now, we only support five textures. And currently, we only support template-based languages or languages containing some specific texture keywords. So I think we can simply use select lists for textures descriptions.

Thanks!

hi @yumingj , thanks for the update.

Here is the final space, https://huggingface.co./spaces/huggingface-projects/drawings-to-human please let me know what do you think.
I'll probably publish on social media on Monday.

Thanks

Hi @radames

Sorry for the late response! The demo is fantastic.

I have one minor suggestion: Could you add a sliding bar or another control mode to allow for the change of sample steps?

Btw, what is the default sample steps value now?

Thanks!

hi @yumingj thanks for the feedback! yes absolutely I'll add the sample steps range. I've been using the lowest 10 to make inference fast, but it's a great idea to have it accessible.

hi @yumingj , I've added the steps range. I'll do some social sharing next week, are you on Twitter? One question, would you expect the model to always generated textures for the giving segmented areas ? can it generalize ? for instance if you draw a tie on a dress like person,

Hi, @radames Thanks for your help! The current version looks good!

My Twitter account is @Jiang_Yuming

For the example you showed, I think it would be little hard for models to generate a tie for a person wearing a dress. As in the dataset, there are few such cases. It would be hard for the model to create a tie.

Thanks!

hi @yumingj , thanks for the update! yes it makes sense, I was wondering if the model could interpolate in those weird combinations.