FrozenBurning / Text2Light

[SIGGRAPH Asia 2022] Text2Light: Zero-Shot Text-Driven HDR Panorama Generation
https://frozenburning.github.io/projects/text2light/
Other
577 stars 46 forks source link

Trying to generate 8K hdr images #7

Closed Junrongh closed 1 year ago

Junrongh commented 1 year ago

Dear author,

Thanks for releasing these fantastic code. I tried to generate a 8K hdr image with your repo but found some issue. I changed the input h=512 w=1024 into h=1024 w=2048 in this file https://github.com/FrozenBurning/Text2Light/blob/master/text2light.py#L74

The generated holistic image is good, since these parameters will not affect this file. holistic_ Sunset California beach However, the [2048x1024] ldr file has some problem(I resized it manually for uploading) ldr_ Sunset California beach

It seems that the spe coordinate should be changed with the input width and height. I'm not so sure what shoud be adjusted. Could you please look into it?

The command line I used is python text2light.py -rg logs/global_sampler_clip -rl logs/local_sampler_outdoor --outdir ./generated_panorama --text "Beijing afternoon 4 pm" --clip clip_emb.npy --sritmo ./logs/sritmo.pth --sr_factor 4

Thank you!

Best wishes

FrozenBurning commented 1 year ago

To generate 8K HDR, instead of modifying h, w, you need to change the input parameter as --sr_factor 8, which directly upscale $512\times1024$ to $4096\times8192$.

Junrongh commented 1 year ago

So the model you trained is fixed to 512x1024?