@EKKIVOK Thanks for the Feedback - of course I did "install missing nodes" several times, but that did not solve the problem. And I of course (two days... :) ) found the fact that there is an alpha-two version in the meantime. But I will try the source you mentioned and I hope you are right and it will solve my problem. :)

@EKKIVOK Noops. I cloned the source into my models folder, but even if I rename the folder "joy-caption-alpha-two" to Joy_caption there ist no clip_model.pt in it - which is the file your script/workflow is searching for. Perhaps I am too silly - then please help me. :)

@EKKIVOK Thanks for your support, but I already edited the path, that was not the problem. I noticed that the *.pt files (Pretrained models) were not in the folder maybe because I did not "git clone" but "install by URL" in the Comfy manager. Interesting. I downloaded the files manually. But no success - the workflow crashes coming to the "Load JoyCaption2 Model" node. Error message is: "rope_scaling must be a dictionary with two fields, type and factor". This seems to be a problem which is caused by a newer version of the LLama LLM https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct/discussions/15 and the solution should be an upgrade of transformers and vllm. I did that, but no success. I did even edit the configuration_llama.py of the JoyCaption2 model - no success. After I had updated pytorch and transformers as suggested, my Comfy installation was completely dead - the modules "Comfy-Easy-Use" and "ComfyUI-essentials" - which are required by several other workflows - did not worked anymore. I noticed then that they are obviously not compatible with the current PyTorch Version 2.5.0 and I had to downgrade again. And the really disgusting fact is that JoyCaption2 or lets say the LLM model perhaps needs an actual version of transformers. And some shitty Comfy modules (ComfyRollStudio seems to be one suspect...) downgrade the current actual transformer version while starting Comfy - and then JoyCaption2 crashes. I think you can be very lucky, if your installation and your workflow works. The whole Python system is wonderful on the one side; as you can combine several modules and have great workflows and software. And it is an awful, glibberish shit of incompatible dependencies and traps to waste your lifetime.

By the way - the other question is - why the heck should I trust these glibberish and intransperent modules and install them on my production system? The reason must be the brain-washing possibilities of StableDiffusion. :)

@EKKIVOK Today I had some time again and I tried a new approach. I bypassed the Joycaption2-Module and provided the text-string manually at https://huggingface.co/spaces/fancyfeast/joy-caption-alpha-two and with a simple prompt input node into your node #336. The the workflow "works" and is busy for over 20 minutes in node #325 "KSampler Adv. Fooocus". There is a picture with inpainted face (I tried this with the mask in the input picture...) in node #59 and #236, but then there is (after 20 minutes+++...) a new error message: "xformers wasn't build with cuda support" (of course it was, but perhaps a shitty comfy node downgraded the installation...) And the whole "Final Touch" section does not come active. I will give up now - that's too crazy.

@EKKIVOK I see. Thank you for your perfect answer. I have an additional question. When I mask and use Lora to replace the face, the light, shadow and skin color on the face will be a little inconsistent with the character's limbs. Do you have a workflow for this problem?