3 ms·
Played with it for a bit in DreamStudio so I could control more of the settings. So far everything it generates is "high quality", but the AI seems to lack the
by fpgaminer 4y ago
Played with it for a bit in DreamStudio so I could control more of the settings. So far everything it generates is "high quality", but the AI seems to lack the creativity and breadth of understanding that DALL-E 2 has. OpenAI's model is better at taking wildly differing concepts and figuring out creative ways to glue them together, even if the end result isn't perfect. Stable Diffusion is very resistant to that, and errs towards the goal of making a high quality image. If it doesn't understand the prompt, it'll pick and choose what parts of the prompt are easiest for it and generate fantastic looking results for those. Which is both good and bad.
For example, I asked it in various ways for a bison dressed as an astronaut. The results varied from just photos of astronauts, to bisons on earth, to bisons on the moon. The bison was always drawn hyper realistically, which is cool, but none of them were dressed as an astronaut. DALLE on the other hand will try all kinds of different ways that a bison might be portrayed as an astronaut. Some realistic, some more imaginative. All of them generally trying to fulfill the prompt. But many results will be crude and imperfect.
I personally find DALLE to be more satisfying to play with right now, because of that creativity. I'm not necessarily looking for the highest quality results. I just want interesting results that follow my prompt. (And no, SD's Scale knob didn't seem to help me). But there's also a place for SD's style if you just want really great looking, but generic stuff.
That said, the current version of SD was explicitly finetuned on an "aesthetically" ranked dataset. So these results aren't really surprising. I'm sure the next generations of SD will start knocking DALLE out of the park in both metrics. And, of course, massive massive props to Stability.ai for releasing this incredible work as open source. Imagine all the tinkering and evolving people are going to do on top of this work. It's going to be incredible.
- GenericPoster 4y agoInteresting. I took a stab at your prompt and SD really struggles. It just completely ignores part of the prompt. Even craiyon puts in an effort to at least complete the entire prompt. The bison is very realistic at least. So maybe the future is different models that have different specialties. Edit: managed to get this one after a few more tries https://imgur.com/a/3061n5d https://imgur.com/a/3061n5d
- IanCal 4y agoYou can take an image and feed it into SD along with the prompt (look for img2img in the readme), I think you can use that to create the idea in craiyon and then move to SD for a quality finish.
- ralfd 4y agoI read that you have to give longer prompts to Stable diffusion. This is my bison astronaut: https://i.imgur.com/ohIuG6F.png https://i.imgur.com/ohIuG6F.png The prompt was: "A bison as an astronaut, tone mapped, shiny, intricate, cinematic lighting, highly detailed, digital painting, artstation, concept art, smooth, sharp focus, illustration, art by terry moore and greg rutkowski and alphonse mucha"
- naillo 4y agoImg2img and inpainting is usually a lot easier to get cool results (with a lot of control on your end) with in my view.
- raylad 4y agoNot sure how you got that from stable diffusion. This is the kind of thing I get - just a bison. https://i.postimg.cc/bw1R10gB/A-bison-as-an-astronaut-tone-mapped-shiny-intricate-cinematic-lighting-highly-detailed-digital.png https://i.postimg.cc/bw1R10gB/A-bison-as-an-astronaut-tone-m...
- Rebelgecko 4y agoI've found that sometimes you need a lot of attempts to get good results. Here's a run I just did, I'd say the top left result is OK, the bottom 2 are wrong/misinterpretations, and the top right result is fantastic: https://i.imgur.com/l8BsWvI.png https://i.imgur.com/l8BsWvI.png
- raylad 4y agoIt doesn't seem to understand things that DALL-E seems to (usually) understand. For example, it doesn't know that there are 5 fingers on most hands. It cuts objects in half randomly, etc. I haven't found many of the images it produces to be really usable. Linked is an example from Dream Studio showing some of these issues: https://i.postimg.cc/qrDKGSVJ/Screen-Shot-2022-08-22-at-7-41-33-PM.png https://i.postimg.cc/qrDKGSVJ/Screen-Shot-2022-08-22-at-7-41...
- Valakas_ 4y agoAnd then you have Midjourney which would most likely be the best one at doing that specific request. The most crazily creative artistic of the 3.
- danielbln 4y agoMJ has announced a new beta mode, that uses SD under the hood.