TL;DR: This is a DMD2 LoRA = use CFG 1 and 8-12 steps (even if i personally prefer 16 steps, that's still half of what i usually use and i get my pictures in 30s instead of 80s).
This is my "personal" brand of the classic ANIMA Turbo LoRA, based on the DMD2 method.
There is already several "Turbo" LORA for ANIMA:
But interesting enough, someone had the idea to make and test a LoRA using the original Cosmos-Predict2.5 architecture: https://civarchive.com/models/2466415/cosmos-predict25-2b-base-distilled-extracted-dmd2-lora
The main process was: extract (using SVD) a LoRA out of "Distilled Cosmos" minus "Original Cosmos" and test it on ANIMA... and it works! (even if some keys are refused because they changed shape in ANIMA). Below is a comparaison of what those bring to the table (RDBT is very close to the base Turbo and so was excluded).
Now, what is did is "stupid" = Merge those two 😁 (i left RDBT alone as sharing merge is forbidden).
To do so, i re-keyed both LoRA and do a re-rank of the resulting merge (32/16 instead of 64/64+32/32). The ratio was 6/4.
This way, i get more details but with a less saturated result.
And it works perfectly with nArnima V2:
Thanks for reading! 💖
Legal
The CircleStone Model is licensed by CircleStone Labs LLC under the CircleStone Non-Commercial License. Copyright CircleStone Labs LLC. IN NO EVENT SHALL CIRCLESTONE LABS LLC BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.
Description
Posted on Civitai, upload found elsewhere are FORBIDDEN and done without the author permission.
FAQ
Comments (15)
I'm absolutely Shooketh. Took my 60s Hires workflow down to 12s (CFG 1.5/8 steps then CFG 1/8 steps). Literally drag and drop wizardry. It even looks better than before!
Cool, need to try this (when I have time :D). You probably want to fix a typo "Now, what is did is" ;)
Phew i just tamed the base model not to baste my images in butter, this lora just reversed it .. Why...
Probably because most DMD2 LoRA are partial to the dataset used for training and the Cosmos2 part i used probably used pictures that weren't anime but realistic. I also noticed that using this turbo lora help me ger better results on my tests for a realistic version of anima (WIP)
I would like to ask a question. What benefit does a styled-turbo-lora have over just adding the turbo lora to a styled checkpoint?
IMHO, this will just prevent the usage of styles.
Baking the turbo in a checkpoint at least saves VRAM and is slightly faster/efficient.
Would you mind to illuminate me? :)
So, i made this one with quantization in mind. The turbo i have merged is in fact less style oriented than the original one. The main idea of this is for people used to "everything in a single" (like Illus/Pony/SDXL) to test ANIMA and make it fast even on smaller setup. Obviously, more advanced people could and can switch to the base model + LoRA, even if in this version, i have also slightly altered the TE ;-)
@n_Arno I see :)
You trained the TE also?
@Darksidewalker I merged a LoRA i made myself but it is not coming from some training, it's an extract of the abliterated version of the TE.
I don't know why, this turbo lora seems the best to me, it shapes objects much better compared to the other turbo lora, yet the other one sees more support from people. Even if the turbo lora on this page changes style a bit, it just seems best structurally. This is really good.
Thanks for the feedback! 💜💜💜
The best turbo lora that ive tested.
update with the 0.2 official lora? 9w9
Honestly the best Turbo LORA for Anima, but like the original, even v2, detailing and image variation can take a nosedive, different seeds don't vary much, but stability improves. Yet slashing more than twice of generation time makes it worthwhile. Went from 2:34m~ for a 1152x1664 image to 30s~.
Tested using LCM/DDIM at 12 steps and CFG 1 with various models, including AnimaIka, AnimaYume, and WAI-Anima.
Oh, thanks for this comprehensive comment! 💜