Image GPT Support

[appvoid]

https://github.com/openai/image-gpt

i don't know how hard would be to implement this model. It seems to have the same architecture as gpt-2.

What is fantastic though it's the possibilities around this model. Just imagine having a stable diffusion like model in your raspberry pi. And even though the 32x32 size is pretty limited, there is already things like realesrgan which amplifies the resolution upto 4 times!

128x128 high-quality images without any dependencies, please considered at least!

[appvoid]

I'm pretty bad at these things (I'm just a python guy) but here's a high level overview if anyone wants to try:

• This involves reshaping the images into a 1D sequence and applying the transformer decoder to predict the pixels.
• They define a "mask" that randomly hides certain elements (pixels) in an image using BERT. I think we have it here: bert.cpp
• Parameter counts are s:76M, m:455M, l:1362M (Don't know yet if they published xl)

To give an idea on how good this could be:

Test image (64x64)

Native 4x upscaling (256x256)

I used an ncnn implementation of realesrgan to upscale it. Meaning that this could be a good oportunity of doing a pretty cool spinoff with the ncnn community hence proving that open source projects yet similar, can cohexist and work together somehow.

[appvoid]

Closing this as it seems like there is not any interest in it in general. It appears to be just an image completion tool and since projects like this already exist, it doesn't worth it trying at least for now.