TURBO for ANIMA (less steps)

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). Versions Comparison (V1 to V4, 8, 12 and 16 steps): - No other LoRA: - With a character LoRA: - With a style LoRA: CFG vs LoRA strength (V4, 12 Steps): Quick clarification: - V1 is using turbo 0.1 - V2 is using turbo 0.2 - V1.5 is using a merge of 0.1 and 0.2 - V3 is based on anima-turbo-v1.0 - V2.9 is based on anima-turbo-v1.0 and turbo 0.2 - V4 is based on anima-turbo-v1.1 - All of them has been merged with the revised cosmos2.5 DMD2 LoRA - Yes, you can use CFG 1.2, 1.4, etc... you may need to add more steps then (that's why i do 16). V4! By popular request, this version is based on my own extract of anima-turbo-v1.1 (128/64) + the cosmos LoRA (ratio 0.65/0.35) and resized to 64/32. PS: V4 has been re-uploaded: i removed the diff layers because ComfyUI don't like those with ANIMA. Correct sha256sum is aa198376f2f31159ba3f60dce19924b27411cdd93e7252543e5c45249ca89048 V2.9 With more extensive tests and help from heavy user of the previous version, it was clear that V3 was killing LoRA when used. So, here is V2.9, a step back to previous turbo for the sake of a better compatibility with LoRA. Recipe (merge done at Rank 128/64): - Cosmos2.5 DMD2 LoRA: 0.3 - Extract Turbo V1.0 LoRA: 0.2 - Turbo v0.2 LoRA: 0.4 Then, all this went through my (MAXDIM 64, CUTOFF 98) to get this slightly heavier LoRA, but providing from my first tests better results. 🥰 V3 With officially out, i knew some people would ask about an update. So here it is! I switched my approach here as this is a full model we are talking about. First, i merged base and turbo at 50/50 to get "half-turbo" in which i merged the Cosmos2 LoRA at 0.5 to get "custom-turbo". V3 is a LoRA extraction of "custom-turbo" minus "base". V1.5 Borrowing a book from and for the people not satisfied by 0.1 and 0.2 and wanted something in the middle, i did a merge of official 0.1 and 0.2 before doing the merge with the cosmos2.5 distilled LoRA. This end being a version between V1 and V2 = V1.5. This will probably be the last version except if a 0.3 version of the official LoRA comes out (and/or if a full distilled version of ANIMA base is released). NB: this version feels more sensitive to watermarks, you can use CFG 1.2-1.4 to be able to add a negative. V2 By popular demand, here is a V2 of this LoRA based on the V0.2 of the source turbo LoRA. 😘 V1 This is my "personal" brand of the classic ANIMA Turbo LoRA, based on the method. There is already several "Turbo" LORA for ANIMA: - - But interesting enough, had the idea to make and test a LoRA using the original Cosmos-Predict2.5 architecture: The main process was: extract (using SVD) a LoRA out of "" minus "" 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 : 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.