Maybe you can convert your 3d images to 2d with img2img and then try to train a model?

@Shurik @RalFinger Thank you for the replies and explanation. I actually don't want 3D-looking generations, so Ral's idea of using img2img on them to make 2D versions would be a very good one if, like you said, we could get consistent results. I've done countless experiments with trying to get posed 3D figure renders to translate into anime or illustrated 2D generations, but unless I'm missing something, there just isn't a way to use both a low denoising strength to preserve the 3D render's composition and details, and one high enough to change the overall style. Based on your latest experiment though, it does sound like it's possible to get anime results from realistic training data.

If you need 3D renders for any testing in the future, let me know and I'll see what I can do. Thanks again for the information.

@SnowSultan I'm totally with you here. If you get a great pose or form, then use I2I with a low strength, it won't change the style. But if you increase the strength enough to change the style, you're likely to end up with a different pose. Very frustrating. However I just saw this video, and it may be of help/interest with this particular issue. https://www.youtube.com/watch?v=YJebdQ30UZQ

@SteveWarner Thank you, I thought I might be missing something obvious. I will watch that video shortly, thanks for that as well. I wonder if small but specific custom loras or textural inversions for specific poses (like this silly example) or even body part positions could aid in either retaining poses or giving more control in text-to-image generations. More precise control over posing is the one thing I want most from AI and it would change the entire industry (for both good and bad) if it can ever be done well.

edit: watched most of the video - oh man, that skeleton 'rig' version is what I've been imagining (and wishing for) since I first used AI. Can't wait until these get an Automatic1111 release of some sort.

@SnowSultan Glad you liked the video. I'll be downloading that tool this week and seeing if I can run it through the paces. Being able to work through the stages of art production (from a studio manager's standpoint) is important, as that's what clients expect. No one wants you to show them final art in your first meeting. You start with sketches, then go to intermediates, then final art. And in that, you lock down the framing, pose, etc. Being able to lock each of these elements in so they don't change between renders is really critical. I'm hoping these tools move us closer to that reality.

@Shurik Well speaking from some experience, as long as the 3d model is toon shaded/flat shaded, or unshaded, and not soft shaded then it won't have as drastic of a 3D effect, for instance I worked on a vTuber, whom is very small, and they gave me a toon shaded model, and a bunch of high quality photos from VRChat in many angles, poses, lighting conditions, and backgrounds, and the final result of the LoRA looked very much 2D in nature. So you can easily introduce 3D Models if any ethical concerns arise, or it's a niche character, concept, etc,

@Shurik did you see the new git called "ControlNet" which has the option to give your images outlines? I guess that could work perfectly to render out 2d from 3d images, give it a try :)

@RalFinger I've been doing almost nothing but doing tests with it since yesterday. It's awesome for applying poses from a second image, but I haven't really had luck transferring only a style to a 3D render yet. Will keep experimenting. :)

@Shurik Oh, that would be wonderful. Take all the time you need.

any news?