3 ms·

I just tried the realtime-text2img demo (uses npm for frontend which i think is too much for this). Modified it to produce only 1 image instead of 16. Works wel

I just tried the realtime-text2img demo (uses npm for frontend which i think is too much for this). Modified it to produce only 1 image instead of 16. Works well on a laptop with RTX-3080. It's probably 2 images / sec. EDIT: The `examples\screen` demo almost feels realtime. Says 4 fps on the window but don't what it represents. EDIT: Denoising in img2img is very low though which means thee returned image is only slightly different from base image.

godelski 3y ago
How's the actual quality, diversity, and alignment though. I'm away from my GPU for a few days. It's always hard to judge generative papers without getting hands on because you write to the reviewers which means you gotta cherry pick (I think this is bad, but it's where we're at). They're using tiny autoencoder? Artspew did that too and was getting higher FPS (but weren't using TensorRT but were using triton) but the quality was garbage (still cool). Regardless, these are impressive even if the quality isn't anywhere near whats shown, but it's hard to tell.
- smusamashah 3y ago
  Not good which is very understandable. You are not supposed to use the output as final image. Find a good propt/seed by iterating quickly. Then go for higher step count to render a higher quality image.
  - foxhop 3y ago
    I use automatic1111 hires fix with SDXL-turbo on a RTX 4090 and it looks the best by far for high quality images once you find a good prompt/seed (but heres the thing turning it on makes all the prompts/seeds that much better...)
  - godelski 3y ago
    I mean there's nothing wrong with low diversity of outputs, it just has different context and its uses are different. The problem is just having all of this properly communicated so that we can evaluate properly.
- numpad0 3y ago
  If you mean not disciplined to caricaturize and stigmatize racial stereotypes a la Disney films by diversity and alignment, that won't be coming from anywhere Asia... Especially the "Asian" face makeup prevalent in NA, sometimes derogatorily called "Pocahontas" face. That one is American special.

godelski 3y ago

How's the actual quality, diversity, and alignment though. I'm away from my GPU for a few days. It's always hard to judge generative papers without getting hands on because you write to the reviewers which means you gotta cherry pick (I think this is bad, but it's where we're at). They're using tiny autoencoder? Artspew did that too and was getting higher FPS (but weren't using TensorRT but were using triton) but the quality was garbage (still cool). Regardless, these are impressive even if the quality isn't anywhere near whats shown, but it's hard to tell.

smusamashah 3y ago
Not good which is very understandable. You are not supposed to use the output as final image. Find a good propt/seed by iterating quickly. Then go for higher step count to render a higher quality image.
- foxhop 3y ago
  I use automatic1111 hires fix with SDXL-turbo on a RTX 4090 and it looks the best by far for high quality images once you find a good prompt/seed (but heres the thing turning it on makes all the prompts/seeds that much better...)
- godelski 3y ago
  I mean there's nothing wrong with low diversity of outputs, it just has different context and its uses are different. The problem is just having all of this properly communicated so that we can evaluate properly.
numpad0 3y ago
If you mean not disciplined to caricaturize and stigmatize racial stereotypes a la Disney films by diversity and alignment, that won't be coming from anywhere Asia... Especially the "Asian" face makeup prevalent in NA, sometimes derogatorily called "Pocahontas" face. That one is American special.