RDBT [Anima]
This page contains the "original LoRA files" for advanced users.
See the main page for more info. (ckpt version, usage, faq, int8 ckpt...)
See this page for update log and version info.
Sharing merges using this model is not allowed. This "restriction" won't affect anyone. It's only aimed at those who steal others' models to sell.
Description
FAQ
Comments (16)
Nice, I like the lora versions better, glad you're updating them.
Not an issue of this version specifically, but the model seems very prone to mosaic censors, and removing it by cfg does not yield great results.
Put "uncensored" in positive.
@NanahiraFan Huh, that does look better. Kind of annoying, but that will have to do for now. Thanks for the tip!
Thanks for the lora version boss. 🙏
I'm guessing someone's already asked this, but I haven't seen it, so I'll ask myself. Can you make a LoRA without CFG distillation and DMD2? All that stuff is cool and all, but I'd love to see a clean LoRA without them. They definitely speed up generation, but you have to sacrifice quality for it. I'd rather wait a bit longer and get a higher quality result. Thanks for your work!
You can mostly achieve that using a negative weight for v0.6_cfg_distilled_only roughly matching the weight of a dmd2 LoRA
_> you have to sacrifice quality for it
This is a common misconception about distilled model. No, it doesn't sacrifice quality. It sacrifices diversity. Distilled model has less perplexity, results in higher quality and less steps.
Just like z-image-turbo and z-image-base. Nobody uses z-image-base, because the quality is lower than turbo.
I don't release the "base" finetuned model because I don't need to, same reason.
@VeerGeer I recommend just lower the LoRA strength and enable cfg and use more steps.
@reakaakasky I get that distillation reduces perplexity and pushes the average aesthetic higher by default. For quick standalone generations, the tradeoff is worth it.
The issue is flexibility. Distilled weights force a very narrow inference workflow. If I want to stack your LoRA with other character or style LoRAs, the distilled weights often clash and create artifacts. It also heavily limits how much I can push the CFG for strict prompt adherence before the image burns, even if I use a "normalizer" node.
A base, undistilled LoRA leaves the original distribution intact. It lets users control the generation through their own preferred samplers, steps, and prompt weighting without constantly fighting the baked-in distillation target. The turbo version is excellent for rapid generation. A base version gives people room to actually build with it.
amazing lora saved my goofy ah gpu from burning ❤️
whats ur gpu waltuh
Do you plan to update this LoRA to be on par with your checkpoint version?
[AI Translation] I swear this is an absolutely amazing LoRA model! I’ve fallen head over heels for it and can never do without it now!
Anyone else just getting blank generations?
Any plans for a Lora based on Anima 1.0? I prefer using the Lora over the checkpoint as it stacks better with other Loras.
I am currently still using p3 v0.24f dmd2 b with Anima base and it works very well though.
That version really was the best, I agree.
I'm guessing it isn't because of dmd2.
Is it because v0.24 won't change style? It does not have aesthetic finetuning like recent versions. (which equals: overfitting the model with a small dataset)