>>2892014Presuming you're using Automatic1111; based on the image you've shared.
1) You may be generating at too low a resolution for your model to be useful, or at least suboptimal. This can be a hardware limitation too; Anything, like most models of its age, generally works best with portrait aspect ratios like yours, but there is a minimal optimal resolution. If bumping height and width by 20 or 30 percent causes a time out, or is just too inefficient, you're kind of screwed until you find another model.
2) Anything, or at least that derivative of it, is relatively quite old, unless I'm wrong about the version you're using. Websites offering free generation tend to use models less than a year old. That's not bad; my derivative of Yohan Diffusion is more than a year old, and I quite like it (on the other hand, my generations are highly repetitive, and could write pages about the content limitations that require a different model; don't do what I do).
3) You may need to fine tune your prompts, and not just your negative embedding, beyond what you want. There's a fair amount of impressive generations out of Anything, or at least they were a year ago, but no doubt the were the products of highly refined prompting, not to mention probably heavy use of Controlnet. The difference between your first and second images is, to be clear, massive; it's a very crude attempt at NovelAI (I assume) versus something immediately recognizable, if kind of plain (subjective of course) character generation. I'd say it's a far bigger leap than from your second to third generation. But, of course, if it's not "good enough", then I refer you to point one and two above.
Maybe that helps. I might be the wrong person to advise on this because, as noted, a lot of my current work is hardly much more complex than what I was doing a year ago (I just spend a lot less time on it now), and I'm still relatively happy with that, and I largely use an old model.
