Quality extremely poor, control image poorly respected as well
It seems to me that this model is vastly undertrained or uses poor data.
Using the following control image with the same parameters as in the example, with the prompt "a landscape":
the result is
Using a stronger conditioning scale of 0.9 the quality is insanely poor:
All images are 1024x1024
It is recommended that this weight be used solely for the initialization of pre-training. We sincerely apologize for the lack of sufficient computational resources and data to run the most optimized open-source models. To support real-world business applications or to achieve state-of-the-art (SOTA) results, developers are strongly advised to fine-tune the model using their own data.
I've had great success with this controlnet model actually, and I use relatively low control strength (<0.4) and high range (0-0.7) and it gives amazingly good result. This is the only controlnet model so far that can "enforce" some of the anatomical elements like legs and hands, much better than the pose model. This is now a standard part of my img2img workflow for SD3.