Atomix Anime XL
This is a non-pony anime model.
After using the pony model for a while, I was surprised that it is so good at NSFW content, but I also find that that it is not good at some styles and is not compatible with non-pony Lora, so I merged this non-pony model,which has Lightning acceleration.
Announce: The prompt information of some images in this model is incorrect, which is due to my workflow settings being incorrect, the error prompt is genation by auto-gen nodes，which i used for testing model. If you encounter this situation, please download the images and get the correct prompt from the ComfyUI interface.
Recommended settings
cfg: 1~2
steps: 4~8
sampler_name: dpmpp_sde_gpu
scheduler: sgm_uniformDescription
FAQ
Comments (16)
If I may ask, how do you merge lightning lora with anime models, while retaining the quality ? For me when I try using sdxl lightning model with vanilla pony or animagine the quality degrades a "LOT" ands hands are all messed up, but this model seems quite good.
Any input is appreciated, thanks.
You don't need to merge a Lightning model to obtain Lightning features, just merge this Lora in the final step:sdxl_lightning_8step_lora.safetensors · ByteDance/SDXL-Lightning at main (huggingface.co)
@AlexLai thanks, but I mean, while using lightning lora with vanilla pony, or animagine, the quality degrades by a lot. On this model thr quality seems fine.
Sorry for the confusion.
Using sgm samplers and cfg 1 at 4 to 6 steps, for 4 step lora and 12 steps at 8 step lora. Ots works really well with releastic models, but on anime models, I didn't have good luck.
@newtrialtime101704 It reminds me of a time when I trying to make the Pony model to obtain Lightning acceleration features, the result is similar to your.Perhaps the Pony's underlying technology used is not compatible with Lightning. My model does not include Pony, so using Lightning's Lora is fine.
Or you can try Turbo Lora, maybe it works:LCM&TurboMix LoRA (Only 12MB, 8-step sampling Effect is superior to using LCM or Turbo alone!) - Euler a.fix | Stable Diffusion LoRA | Civitai
@AlexLai thanks for replying
@yuzu77777777502 It seems that you are using WebUI. I have not used WebUI to merge models before, so I'm afraid I won't be able to help you.
But if you use ComfyUI too, the default nodes is enough for merging model:Model Merging Examples | ComfyUI_examples (comfyanonymous.github.io)
@AlexLai thank you very much!!!
@AlexLai most people with Pony already played with Turbo versions, but Lightning might be better... but SDXL anime models and Pony are not necessarily the most compatible due to their divergent base training under the same architecture.
So is this pony related? If not, mentioning it seems confusing
I'm getting noisy result, vae is in automatic. I'm using 1111 no idea whats wrong :/