Generated image quality about gill HOT 3 CLOSED

Right now I don't have these planned for the near future, but in principle it should be quite easy to swap in stronger LLMs, like some of the instruction or dialogue fine-tuned one. From some conversations with others about the work, I think it's possible that some people are already working on this, so hopefully we might see something public soon!

from gill.

Hi, thanks for your interest! I think this level of output is somewhat expected. Your first image is the exact same as the one from our notebook, so the model should be correct.

Note that the image generator being used here is Stable Diffusion v1.5 (which is worse than SDXL and models like DeepFloyd), so the image realism/quality (not semantic match, as explained in our paper) is somewhat upper bounded by that model. Hope that helps!

from gill.

Hi @kohjingyu , thanks for your quick reply, will the code support stronger diffusion models and LLM models in the future?

from gill.