Due to the absence of ONNX models in Civitai I upload this one for everyone that wants to test this setup.
The main model is a merge, that I do not remember where it came from...but it's based on SD1.5 models.
All images have been generated without later editing, only Conan & Red Sonja got additional steps for its creation: character independent creation, sum of latents and restore its faces..) . No Loras, No textual inversion or other adds & tools.
I've uploaded three versions:
Not Optimized: consumes more VRAM memory but it load the model faster(10s), recommended for testing. i.e.: 640x640 in 4Gb card.
Mem Optimized: less VRAM consumption, but higher initial loading time (up to 1min per initial model load). i.e: with 4gb you might reach 1024x768 with no issue
Latent Consistency : needs low steps (6-12) and each step is 40% to 50% faster than a normal steps, consumes a bit more memory than a Not Optimized Model. This model is a fp16 replica of "SimianLuo/LCM_Dreamshaper_v7" and not an own merge.
Also, if you want to test them, you might check at my UI:
https://github.com/NeusZimmer/ONNX-ModularUI-StableDiffusion/
Feedback will be much appreciated.
**LCM: a model following the full description as its papers, those with LCM-LORAS weights inside were already working
As SD in ONNX requires a directory structure to work, I've uploaded the main models as zip files, but Civitai only allows me to mark them as training data: Those are the main models.
PD: Working for SD1.x and SD2.x models, but not SDXL as I'm not able to convert SDXL models currently to onnx-fp16.
I've Uploaded:
3xMain Base modelx
VAE encoder & Decoder
TextEncoder (clip-slip1)
Pending to upload: (but available on huggingface) ControlNet base models and Danbooru tagging, TextEncoder for clip-slip 2,3 and 4