this model already baked a base model and text encoders and VAE, and several LoRAs.
put the model into ComfyUI/models/diffusion_models
use this custom node for loading model as a normal Checkpoint
https://github.com/easygoing0114/ComfyUI-easygoing-nodes
more info:
base model ... FP8 format
text encoder ... FP8 format
VAE ... BF16 format
so, total file size is only 17GB. compact and small data losses.
fun stuff:
includes some LoRAs for NSFW stuff :-)
slightly better from the base model, but it's not perfect.
personally, I'm so satisfied with the quality because thoroughly tested it again and again.
license:
QI2.1 is non-comercial license, so be careful about it, if you are planning buisiness.
Description
1st version
FAQ
Comments (7)
强大，无需多言！
想起了2511的那个aio 后续 。如果后续融合加入可靠稳定的加速lora就完美了。加油！
turbo lora, nice idea! but Spectrum node is also very useful and speedy, isn't it? 謝謝大家
你加了lora，那我还能叠加lora使用吗
yes, you can. baked loras is 0.1 strength each, very weak. so adding nomal loras will works fine.
Bro, you're hero
For some reason, the higher the resolution, the more things fall apart—extra fingers appear, and so on. Connecting a separate text encoder improves the images, but not enough to achieve a truly "photorealistic" result.