Generation (Recommendation)

• Nomal Model Generation (Recommendation)

  • Sampler: dpmpp_2m/Euler/Res-multistep/Er-sde

  • Scheduler: sgm-uniform/beta/beta57/linear-quadratic

  • Resolution: 832x1216, 812x1216, 1152x1536, 1536x1536 (iffy)

  • Steps: 28-50

  • CFG: 3.5-5

• Turbo Model Generation (Recommendation)

  • Sampler: Er-sde/dpmpp_2m_sde_gpu

  • Scheduler: simple/sgm-uniform

  • Resolution: 832x1216, 812x1216, 1152x1536, 1536x1536

  • Steps: 12

  • CFG: 1
    

This model is a merge I created by testing various combinations of LoRAs—a process that involved a lot of experimentation. I don’t recall the exact combinations, but I used the "comfyui-model-diff-to-lora" tool alongside the "Anima 1.0 base 2B" model. I extracted LoRAs by setting the original, unmodified base model as "model 1" and the model containing all the desired LoRA combinations as "model 2." The final output was a 700MB, rank-256 LoRA; given my PC’s specs, the extraction process takes about 15 minutes.

After combining and extracting the LoRAs using this custom node, I used "ComfyUI-Anima-Remap" to adapt the LoRAs from the "Anima 2B" version to the "Anima 2.9B" version. I then saved these adjustments, resulting in customized versions of both the Anima 2B and 2.9B models, available in fp16 and int8 formats.

I also extracted a "turbo" LoRA from the "Anima 1.0 turbo" model.

Using "comfyui-model-diff-to-lora," I set "Anima 1.0 base" as "model 1" and "Anima 1.0 turbo" as "model 2" to perform the extraction.

Note: I did this because I noticed that the "anima-turbo-lora-v0.2.safetensors" file they provide offers lower quality than what is already baked into the turbo models themselves. By using "comfyui-model-diff-to-lora," I was able to isolate and extract the specific differences between the models. I also extracted a LoRA from "anima_aestheticV11," which works perfectly without needing to download a separate version of the model. Note that LoRAs can be extracted at ranks 256, 128, 64, 32, 16, and 8;

the memory requirement is essentially halved with each step down in rank.

https://github.com/shootthesound/comfyui-model-diff-to-lora

https://github.com/shin131002/ComfyUI-Anima-Remap

Specs:

Xeon E5 2650v4, 16GB RAM

RX 7600 8GB

240GB SATA SSD