I get the exact same results as clip-L
Running the same prompts with the same seed and parameters gives the exact same results wether I'm using the original Clip-L, or the "ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF" encoder (with Flux schnell). And I'm not just saying that the difference isn't easily noticable, even the hashes of the two images match (when stripping off metadata). The images are absolutely identical.
Can you share the prompts you used to get the example images on the model Card?
Edit: nevermid, I tried with ComfyUI, and it does change the output. Maybe there's a bug with the other inference backend I was using...
Thanks for the heads up about that! Might consider opening an issue with (Forge, if it was that you've initially used - if whatever you used happens to have a github). Might indeed be a bug, in that case. Cheers!
I was using stable-diffusion.cpp, I reported the bug there