Hi! This is my unofficial Anima Turbo v1.0 LoRA. I made this so I could use it in my own merges. It also looks good on its own.
This LoRA was extracted from the new Anima Turbo v1.0 base model
Not much to say about this one. Good for merges, or if you dont want to use the base model and still get the Turbo v1.0 style and speed.
CFG: 1.0
Steps: 8-12
Strength: ~1.0
v1.0-Resize/DoRA
This is more of a technical update. I pulled a fixed 64 rank with dora. Then resized using budgeting/svd 8-64. Result? Quality improves, overall better balance/health, file size is nearly half the original 64 and just over the original 32. If you dont want the DoRA, strip/convert the keys before merging or exclude them during.
v1
Original extract with multiple variants 32 / 48 / 64
I do this for fun and to learn. Likes, comments and reactions are nice to see if my efforts are worth while and posting. I dont really care about the buzz or whatever but drop me a comment and like if you enjoy what Ive done. <3
Description
Initial Release
FAQ
Comments (9)
Just curious, do you think circlestone labs did a better job of distillation than the existing combinations of cosmos and 0.2?
On this new turbo v1.0? The stats ive pulled so far says its solid. Its been holding up well in testing. The 0.2 was rough. Ive checked out some of the other merged loras and their not bad. v1.0 would be a healthier base to use from what ive seen. Sorry if I'm missing your point though, it's been a long day
@SpicyKatsu After some testing of Anima Turbo 1.0 - i noticed that ther are some artifacts and messy images, some noise etc, but i guess 8 steps not enough.
Well, I wish you a good rest, thanks for the loras!
@BlueberryTrain My baseline is normally 12 steps :)
A "not so quick" follow up. It is for sure an interesting take. The v0.2 had multiple issues, esp the "sweating", but the style was good. Merges of v0.2 can end up being great. v1.0 while it resolved many issues and did overall improve on the health of the weights, it has a dramatic shift in style. The strong weights can backfire in merges, and needs to be toned down a lot, else it overpowers.
I did multiple iterations of my new merge using v1.0 and compared against my current Anima Turbo Refined merge. Its a whole other beast. I still think the new v1.0 is a much healthier base to use and doesnt hit you with random issues, but its also personal preference on what types of styles you enjoy. Both of them can be pushed in many directions to get variance out of them esp in merges.
Personally, I would love to see a middle ground here. Style closer to v0.2 but without the terrible lora health and sweating. Unfortunately, its not fair to A/B compare v1.0 and v0.2 as they are simply too different.
@SpicyKatsu I also noticed that the extracted lora has a baked-in style of its own, a bit washed-out and very blurry. This applies even to artists trained in the model, of course, I know that distillation itself destroys the original style, but 0.2 had the most accurate (by Turbo Lora standards) styles (and, of course, the sweating issue, yes). I spent some time playing with the weights and merging the new lore with 0.2, even with Cosmos with different coefficients – the results weren't particularly good. I was hoping that Cosmos would act as a stabilizer – since it's very stable despite the complete lack of artistic style – but unfortunately, I don't have the knowledge to work with it.
The new turbo model also has checkerboard noise on leaves, eyes, and other complex objects because it requires 12 steps, while 0.2 — and especially its mix with Сosmos — allowed it to be generated even at 6 steps without noticable drop of quality and noise, while maintaining the artists' styles at an acceptable level.
I guess the circlestone labs did a whole new distillation.
@BlueberryTrain did you test the combined cosmos/anima as well. Was that any better? The detailer I added in those seems to go a long way to improving specific aspects and backgrounds. I Havent really stress tested much yet and i tend to stick with 1536 and 12 step as my default since those work well for me. Mixing around samplers/scheduler for fun. So far from what ive seen cosmos mixing in does help overall, depending on the amount mixed and how you merge them. It can also outright ruin the merge if too much cosmos gets in.
I did notice some checker boarding in certain images on some leaves but it was not obvious at first... I ran those thru adore and seems to clear it up well.
Ive been putting a lot of work into my app to make all of this possible in my own ways. If you want you can give it a try for merging or extractions. Might lower the bar for entry. If so feedback on it would be great too :)
Im still experimenting with many things as I only got into all this image generation very recently.
@SpicyKatsu I tested all the turbo Loras on the site so
This - is based on 0.2 and Cosmos - can generate on 8 steps without artifacts, but sweating.
https://civitai.red/models/2709753/expressive-accurate-turbo-for-anima?modelVersionId=3043931
Yours Remix is very nice - but it need weight around 1.40 to push styles.
I also found this
https://huggingface.co/sorryhyun/anima-turbo-2step/tree/main
I don't know how author managed to do it - but it generates very well at 8 steps and preserves styles, i guess description about 2 steps is a joke.
Also - Euler/simple is more prompt adherent - but ER_SDE is more interesting in shading - and colors are more accurate, for example (on Euler) instead of black shorts it's generate dark green and so on (on pretty much all turbo loras out there).
Well thanks man, this ought cut some time with the old models.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.