Hacker News Reader

Comment by naillo | Hacker News Reader

Not really relevant bc it's kinda nitpicky but amazingly it's not even 4gb, it's more like 2gb if you use the float16 version (which has no quality degredations). Quite amazing that so many images fit in that small a package.

View on HN

hmm -- I tried converting to float16 just using a naive model.half() call and saw some quality degradation in my images compared to just autocasting parts of the model to float16 while leaving others at float32. Curious if anyone else has had the same experience.

Might be that there's some degredation but I think it's pretty close. Anyway I'm using their 'official' fp16 version which they might be doing some extra magic on idk. I.e. via

  StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4",
    revision="fp16", torch_dtype=torch.float16, ...)

But you just said "no degradation".

Try it yourself ;) I haven't spotted any but I can't dispute the parents observations.

My best guess is that I just did something wrong lol. Autocast seems to convert most of the model to fp16 anyways and that works great, so I'll just keep using that!

Can I download ckpt file for fp16?

You may want to try bfloat16 instead.

That's how I felt about the original ipod.

The images do not. The output of the trained / weighted correlations do.

Exchanging compute for space in the space-compute tradeoff curve.

Reply on news.ycombinator.com