circlestone-labs

/

Anima

like 980

Follow

CircleStone Labs 506

I would be in favor of more photoreal content to improve background and perspective. Booru is 80% white background images. That's a massive waste of space and untapped potential.

I would be in favor of more photoreal content to improve background and perspective. Booru is 80% white background images. That's a massive waste of space and untapped potential.

I never thought of it that way, you may have a point. Interested in hearing more perspectives on this point, Anima is massively improved over illustrious/noobai in every way but backgrounds are the least improved from what i've seen so far. That may be why.

Oh and in lora testing for v0.1, when i forgot to add at least some sort of a background to my dataset, the character may sometimes generate with a purely black or white background. Thought it was interesting how well Anima learns patterns like that.

What about models larger than 0.6B, is there a chance that the release will be like that?

I would be in favor of more photoreal content to improve background and perspective. Booru is 80% white background images. That's a massive waste of space and untapped potential.

For example something from there: https://huggingface.co/datasets/lodestones/pixelprose

I would be in favor of more photoreal content to improve background and perspective. Booru is 80% white background images. That's a massive waste of space and untapped potential.

I feel like it's totally not worth it. By adding photorealistic images to the dataset we only drift further away from the idea of an anime-oriented model and end up with that annoying 2.5D look.

As a user you can easily satisfy that craving yourself without changing the model's original purpose just by training the specific concepts you personally need. For example motimalu made a lora for photorealistic backgrounds which look really great paired with 2d characters. This partially covers the need for detailed backgrounds and the results turn out really interesting.

I feel adding more photorealistic data wouldn't be a problem if they are tagged correctly. Danbooru already has data on it with certain prompts
photo_background = photorealistic backgrounds with anime characters
cosplay_photo = Photos of cosplayers
figure_(medium) = Photos of anime figurines (so real photos)
photo_(medium) = general danbooru tag that has the previous tags in them and more

So, I feel adding more data into some of these would not be bad, they are already in Anima Preview 2, improving them shouldn't affect the overall quality of the model and it might also add variety. Besides don't you want a model that can do things without adding a Lora? IMO the less Loras involved the better.

It’s likely not going to be a problem, but even when images are tagged correctly they still affect the model anyway and if there are too many of them it’s obvious the model will shift in their direction.

It's clear there are photorealistic images in the dataset and in general even now with Anima in its preview state I haven't seen any issues with background variety as you get striking generations even without loras.

I would be in favor of more photoreal content to improve background and perspective. Booru is 80% white background images. That's a massive waste of space and untapped potential.

I feel like it's totally not worth it. By adding photorealistic images to the dataset we only drift further away from the idea of an anime-oriented model and end up with that annoying 2.5D look.

I wonder if this still applies to more modern model architectures, though. Yes, the 2.5D "sickness" of sdxl is well known and I agree. But, for instance, my go to example is Chroma1-HD in combination with a flat art style lora. I get perfectly good 2d results but yet the model is also able to generalize all its photo data and give me perfect backgrounds, angles, perspective, proportions, detailed household objects or whatever while everything stays flat. And that's a model with a very heavy photoreal bias. It should be even more managable with a heavy anime based model.

The question is simple: Do we really want to still struggle with nonsense furniture, nonsense windows, nonsense shadows etc? In this regard Anima is unfortunately as poisened as sdxl.