3 ms·
I'm curious how many people have actually tried these networks out. I've dropped a bunch of demos below so you can actually try them and I encourage everyone to
by godelski 4y ago
I'm curious how many people have actually tried these networks out. I've dropped a bunch of demos below so you can actually try them and I encourage everyone to do so before they form strong opinions here.
As someone that works in generative modeling I think it is important to note that we are hand selecting results and not showing the average ones. There's been such a hype machine that if we show our average results that we'll get killed in review (publishing in a hyped space unfortunately requires a lot of salesmanship). Not a thing I'm particularly happy about, but I think this is important to note because if you're just looking at the images on Twitter, Blogs, and other social media then you're seeing the top 1% results. Images that take quite a bit of time and effort to produce. Likely not as long as an actual artist to make them, but still a lot of skill is involved.
I do agree that I think it'll be a big boon for artists. I've even been using these to get better at drawing. The results are bad even with prompt engineering, but hey, my human mind can fill in the gaps and it is good for expanding creativity. We'll get better at these models for sure but I'm not quite convinced it'll kill a lot of art jobs. But maybe I don't know what those jobs are that are being killed. But then again, I've even seen artists do much better at these models than me even though I understand the code and math going on. There's also cool things like walking through the latent spaces, but you won't be able to do that with the hugging face code. But that is art that you won't be able to do elsewhere. It is entirely new.
https://huggingface.co/spaces/akhaliq/Ghibli-Diffusion https://huggingface.co/spaces/akhaliq/Ghibli-Diffusion
https://huggingface.co/lambdalabs/sd-image-variations-diffusers https://huggingface.co/lambdalabs/sd-image-variations-diffus...
https://huggingface.co/spaces/shi-labs/Versatile-Diffusion https://huggingface.co/spaces/shi-labs/Versatile-Diffusion
https://huggingface.co/spaces/akhaliq/openjourney https://huggingface.co/spaces/akhaliq/openjourney
https://huggingface.co/spaces/anzorq/finetuned_diffusion https://huggingface.co/spaces/anzorq/finetuned_diffusion
https://huggingface.co/lambdalabs/sd-pokemon-diffusers https://huggingface.co/lambdalabs/sd-pokemon-diffusers
https://huggingface.co/spaces/hakurei/waifu-diffusion-demo https://huggingface.co/spaces/hakurei/waifu-diffusion-demo
- woolion 4y agoI tried Ghibli diffusion to get some decent cartoon styled results with SD, but it's really bad. Results were all hallucinations from the training data, no interesting new combinations, no novel prompts that would work. I stopped after a few dozens. I'd still like to get some good cartoony results with SD, but it's clearly not possible yet.
- godelski 4y agoIt is still possible. I've gotten some really good quality stuff. The thing is you still need a lot of work and you're often going to have to play with parameters hugging face may not give you access to. We also generally generate images in batches and select the better ones lol. But that's why you see DALLE give you several images of the same prompt.
- woolion 4y agoAny model, prompts, etc you recommend? I even tried to use DreamBooth to train on my own inputs and the results are not really usable, and that was with SD1.4 and AnythingV3 as base models.
- godelski 4y agoI've been more messing with DALLE and in those specifying style is really useful. For SD I think the best thing is to follow some of the Twitter accounts that are posting prompts with their images (though they aren't posting other settings usually). Twitter on my computer is not working right now so I'll try to remember to come back later. But suggestions I'll give is that you need to add detail to the prompts. SD also has negative prompts which really help. So use a lot of words if you can (twitter accounts will help you find magic phrases like "bismuth neon chrome copper patina verdigris"[0]). But you may even see here that the magic words with SD aren't as simple as with DALLE I do think you're understanding, though, what my post is about and why I think a lot of people are talking from a naive perspective. If you only show the wins you have some extreme biases. [0] https://twitter.com/sureailabs/status/1596211001120153601 https://twitter.com/sureailabs/status/1596211001120153601
- woolion 4y agoInteresting magic word. My prompts were often full of 'masterpiece' or 'octane render' (which improves 3D consistency). With regard to your main point, I generated more than a thousand images trough various models, and yes good results such as these posted on social media are very rare. And people definitely spend lots of time to adjust things with inpainting and other tricks, in a way that I would find infuriating. For the effort it takes I would rather just redraw it. It depends on the styles; with brushstrokes you can hide a lot, with cartoon lineart is crucial so fixing requires more work.