Using SD3 ControlNet Models in ComfyUI: Comparison with SD1.5 and SDXL, Controllable but Unstable 

InstantX, the author of InstantID, has released three ControlNet models for SD3, and ComfyUI quickly updated to support them. In this video, I’ll show you how to use these models, how to set the parameters, and what the results look like, especially in comparison to the mature SD1.5 and SDXL.
To summarize:
Need to upgrade ComfyUI, but no need to upgrade any Python libraries.
Keep the weights low. Using single model should be at most 0.7, multiple models should be at most 0.5, and the end time should be earlier.
Anatomical accuracy has improved, but not significantly.
Less adherent to reference images, especially compared to the SDXL and SD1.5 models.
It will lower image quality, so it’s recommended to use it for initial drafts, followed by upscaling and adding details.
Model download link:
InstantX/SD3-Controlnet-Canny: huggingface.co...
InstantX/SD3-Controlnet-Tile: huggingface.co...
InstantX/SD3-Controlnet-Pose: huggingface.co...