@yamatazen how is this better than lora in case of training for Anima?

@yamatazen I think I get what you mean, you used to train on SD, right? There really was a difference there. Don't worry about that for Anima; you can train regular Lora models here. The difference is either 0 or ~0, but training should go faster.

@yamatazen I mran what I mean. No sense to train Dora for anima instead of Lora. Only in very specific cases. Anima≠SD.

@yamatazen It all boils down to zero ROI (return on investment).

Here is why DoRA lost its relevance when switching to DiT (like Anima, Flux, Qwen Image, Krea 2):

A) No visual benefit. In older UNet architectures (SD 1.5 / SDXL), decomposing weights into magnitude and direction actually helped capture details better. In Transformer architectures (DiT), this hypothesis no longer holds: the final images are indistinguishable from standard LoRA. Any tiny difference falls well within standard generation and seed noise, making it invisible to the human eye.

B) Speed penalty. Due to constant recalculation of magnitude vectors, DoRA adds extra compute overhead. Training slows down by a couple of percent, accompanied by a slight bump in VRAM usage and processing effort.

C) In the context of DiT, you end up spending more time and resources on extra math for a result that offers zero practical advantage over classic LoRA.

There’s no point in burning extra time and compute on DoRA when standard LoRA on DiT delivers the exact same quality.

@MonsieurBaka DoRA is not a technique meant for UNeTs. If anything, it would work better for DiTs. DoRA has better capacity for the same rank and is more stable than LoRA. The only argument for LoRA is if you're using SingLoRA, but SingDoRA is possible too.

@MonsterAzi 

"DoRA is not a technique meant for UNeTs." Charitable reading of "not meant for": if it means "not exclusive to U-Net, works better on DiT" rather than "doesn't work on U-Net," that's a more defensible framing. But even so - the original DoRA paper itself included U-Net/DreamBooth experiments and reported them as a win, so the strong version of that claim doesn't hold up. And I haven't seen actual comparative data showing DoRA's edge is smaller on U-Net vs DiT - that reads more like an assumption pulled from the paper's framing than something tested.

On capacity: DoRA's advantage over LoRA at equal rank was most visible in the paper's low-rank tests (rank 4-8). That mattered in 2023–early 2024, when VRAM was a real constraint and people trained low-rank out of necessity. That's mostly not the case anymore - 8GB+ (which is enought for a small model like Anima using a few optimization tricks) -- 24GB+ (for a massive models) or cloud rental is standard now, rank 24-64 is trivial, and at that range the visual difference between LoRA and DoRA is basically not there.

So what's actually left: DoRA trains slower (extra decomposition step per iteration), for a benefit that doesn't show up at the ranks people use today. That's the real tradeoff, and it's the reason the choice looks odd for something as narrow as a single clothing concept - not that DoRA "doesn't work," just that it's paying a cost to solve a problem from years ago.

One honest question: have you actually trained two otherwise-identical models: LoRA vs DoRA, or SingLoRA vs SingDoRA, on the same dataset and params to compare? If you've got that side by side, I'd genuinely like to see it. That'd settle this better than either of us quoting the paper (we are using the same Nvidia's website page) at each other.

I don't think one type of learning is worse than the other, like they're different types of pizza. I don't consider one type of model training to be worse or better to another, like different types of pizza. These are just instruments to be used. It would be great if new formats appeared more often. But here, as always, the question will be: speed, results, compatibility, and additional capabilities. For August 2026, the leader is clear.