Some inference questions

[luzhoutao]

Hi, thanks for open-sourcing this interesting project!

1. During inference https://github.com/csuhan/Tar/blob/main/t2i_inference.py which component used as the diffusion-based SANA image decoder? It looks like the inference code only use LLamaGen decoder?
2. For the vector quantised SigLIP, how do you evaluate reconstruction quality when there’s no real image to generate?

[csuhan]

Hi @luzhoutao ,

1. We do not release SANA image decoder now. t2i_inference.py only supports AR-based decoder. We will clean and release the training & inference code of AR decoder and diffusion decoder later.
2. We evaluate its reconstruction quality by an image decoder, such as AR-decoder. So the whole pipeline is:
   image->siglip->vq->AR/diffusion decoder->reconstructed image

[libai-lab]

Does it support image editing and multi image editing