circlestone-labs

/

Anima

like 1.99k

Follow

CircleStone Labs 922

From my reading of the description, it seems to have image generation capabilities.

However, the bigger issue is that after stripping away the cumbersome parts for comparison, although the overall generative part is smaller than Anima, there is not much difference in essence because it simply offloads the load to the inference side.

Furthermore, Wan2.2's VAE is actually quite lacking in visual reconstruction and trainability.

It falls short compared to modern VAEs like Flux 2 VAE or continuously improved series like Qwen Image 2.0 VAE.

Its reconstruction quality is also inferior to the Qwen Image VAE currently used by Anima, which is on par with Flux 1 VAE.

Wan2.2's VAE is undoubtedly a tier below, and it is arguably even worse than Wan2.1's VAE.

At the same time, looking at the situation, the model seems to lack high-resolution images, so a lot will need to be supplemented during subsequent training and fine-tuning.

On top of that, since the generation component packs in too many things, its effective capacity might not be very large.

Cosmos3 Edge may not be a suitable choice for a next-generation model.

What about that new Microsoft model? How flexible is it with full scale finetunes? It really is far from SOTA, but seems like a upgrade from Cosmos and it's small and has a good license

Still I think native MMDiT can be an upgrade for consumer-GPU developers since various image-to-image methods require multimodal attention architecture.

I don't work with CircleStone Labs but I can definitely talk about these:

1. I noticed that COSMOS3 EDGE 4B doesn't have text-to-image capabilities at all — it's purely a reasoning model. Does this perhaps mean that the community's expectations for COSMOS3 have been completely dashed?

It does, and in fact it's sort of an "omni-modal" model so it supports text/image/video/action inputs and outputting text/image/video/action. Sadly, they stripped audio capabilities from Edge 4B, so you'd have to use Nano 16B or Super 64B for audio inputs/outputs. The models are just 50% Qwen3 2B, 8B, or 32B LLM (the "Reasoner" tower), and 50% DiT (the "Generator" tower).
The obvious issue is that Cosmos 3 is trained to be robotics-first, and very VERY few people, not even most ML researchers, are able to work with such a broadly capable architecture effectively. While the pre-trainin is there, it will be a huge task to post-train tens of millions of images and videos including 15 million in the current booru dataset. But I think it could (theoretically) be genuinely incredible.

2.Will the next iteration of ANIMA continue to focus on aesthetic refinement built on top of the existing COSMOS2 foundation, or will it seek opportunities on other base models?

There are actually Cosmos-Predict2-0.6B-Text2Image and Cosmos-Predict2-14B-Text2Image, alongside the 2B that the current Anima 1.0 was trained on. I could see those being the easiest targets for additional Anima models, but they are over a year old as of now. Another potential base is Krea 2 Turbo (which is CFG distilled and only requires 8 steps), however it is 12B so it would take a lot of VRAM (would run about the same speed as Flux.1).

I genuinely don't understand what the hub-bub about "HEY GUYS LET'S BURN TENS OF THOUSANDS OF DOLLARS DOING THE SAME TRAINING ON NEW MODEL" expecting a different, much better, result.

You could instead spend a few hundred dollars giving someone a job to implement Normalized Attention Guidance for this architecture, enjoy the various high-diversity low-stepcount distillations, mayhaps even invest into implementing both of these things into krita-ai-diffusion as that is the only competent image-diffusion + image-editor mix I've seen.

Then you'd have a great compositional model, which prints images very quickly even on piddly hardware, and has working positive and negative prompting that doesn't break with controlnets for further pipeline expansions (e.g. using OPii Rig03 in Blender)

but that's like, just my opinion, man
and that requires like, giving someone with programming skills a job, interacting with people...

This model already works shockingly good, with a few toolchain expansions, it would work fantastic as the bedrock of low-hardware 2D art production - why throw all that away?

I genuinely don't understand what the hub-bub about "HEY GUYS LET'S BURN TENS OF THOUSANDS OF DOLLARS DOING THE SAME TRAINING ON NEW MODEL" expecting a different, much better, result.

You could instead spend a few hundred dollars giving someone a job to implement Normalized Attention Guidance for this architecture, enjoy the various high-diversity low-stepcount distillations, mayhaps even invest into implementing both of these things into krita-ai-diffusion as that is the only competent image-diffusion + image-editor mix I've seen.

Then you'd have a great compositional model, which prints images very quickly even on piddly hardware, and has working positive and negative prompting that doesn't break with controlnets for further pipeline expansions (e.g. using OPii Rig03 in Blender)

but that's like, just my opinion, man
and that requires like, giving someone with programming skills a job, interacting with people...

This model already works shockingly good, with a few toolchain expansions, it would work fantastic as the bedrock of low-hardware 2D art production - why throw all that away?

Keep using SDXL at this point? Why is each lab retraining every few months, do they want to waste money?

We need a better architecture, because finetuners are lamenting catastrophic forgetting, at least. Model will not grow as sdxl like this.

You can try and strawman it into SDXL worship (sd1.5 was better until noobxl, due to reloading with a different LoRA stack having such a negligible time cost with sd1.5, btw btw), but that is besides the point.

"each lab" is "retraining every few months" because that's how you keep the gravy train going. They want to waste money, because they have a cozy job they'd like to keep. Yes.

'Catastrophic forgetting' is not an architectural problem, it is the same problem as every model has always had; dataset fragmentation.
Models train better, if you have data in there that has absolutely nothing to do with what you're trying to teach it, and absolutely everything to do with what the model already knows.

If the people solving problems back when I was a young'un were to solve this problem, they'd likely end up with something like The Pirate Bay for training data; decentralized database of Stuff that is labelled according to a common standard.

Except now, there is usually no standard. Nobody wants to share their dataset, because their dataset is pilfered and stolen.
Which was never a problem for TPB, as 'making cashmoney' was never on the menu; making life better, was.

What do you think happens, if you finetune an existing model, on data that is similar-looking to what it has already seen, yet labelled in a different order, using slightly different wording, missing tags, missing descriptions?
That's right; loose connections break apart as Noise is introduced. It forgets.

yet instead of standardizing on language, labelling, order of labels and description, alongside What's In The Dataset - it is instead a sisyphean goose chase of The Next Big Thing.
it is as funny and hysterical, as it is pathetic and sad

speaking of funny; Anima is actually much better in this regard than almost any other model, as it at least makes an attempt at sharing what its input dataset looks like.
it's still not quite "common corpus of Stuffs that the model knows with image and text label, a small one for small LoRAs, a bigger one for bigger finetunes, so as to stabilize things" - alongside a full pipeline to label New Things exactly as everything else was labelled - but it is certainly better than The Things Before.

"each lab" is "retraining every few months" because that's how you keep the gravy train going. They want to waste money, because they have a cozy job they'd like to keep. Yes.

Sure, clearly we aren't getting a better arch every few months that makes training worth it.

'Catastrophic forgetting' is not an architectural problem, it is the same problem as every model has always had; dataset fragmentation.

Are we sure? Why aren't we experiencing it that bad with Krea2 and other models? Anima is the worst offender.
Also, no, I doubt the problem is some captioning issue, considering we are using LLMs to create textual embedding now and we have the captioning guidelines.

Regardless, I want something with flux2vae, that won't forget stuff extremely easily, it's pointless to train on it otherwise.

I really don't get the catastrophic forgetting problem people claim Anima has, what am I missing exactly? That said, I still think something based on new model arch is always welcome, perhaps right now nothing gives a meaningful upgrade though, if a model drops that is on par with Krea 2 and responds well to large scale fine-tune e.g. not your LoRA slop with a few concepts at most, but large scale like Anima where there is a fuck ton of new concepts, characters, and artists not just a few then it's 100000% worth it