1038lab
Seeking JoyCaption fine-tuned model recommendations
[JoyCaption] ℹ️ No custom models found, skipping user-defined HF models.
[JoyCaption GGUF] ℹ️ No custom models found, skipping user-defined GGUF models.
I noticed your custom JoyCaption node supports alternative fine-tuned JoyCaption models. Do you have any recommended ones? The original model often responds with self-posed questions.
The message output in your log is normal — it just means no custom models were found. If your concern is that JoyCaption sometimes produces self-questioning lines like “What is this?” or “Could it be...”, you don’t need another model for that.
you can try using the JoyCaption Adv node instead. It allows you to set a custom prompt so you can filter or guide the output style (for example, remove uncertain or self-reflective phrases).
The message output in your log is normal — it just means no custom models were found. If your concern is that JoyCaption sometimes produces self-questioning lines like “What is this?” or “Could it be...”, you don’t need another model for that.
you can try using the JoyCaption Adv node instead. It allows you to set a custom prompt so you can filter or guide the output style (for example, remove uncertain or self-reflective phrases).
For comma separated tags as outputs the "ASSISTANT" content showing up at all is the main issue. Even if what follows after "ASSISSTANT:" is non-deterministically removing uncertain or self-reflective phrases part of the time.
remove uncertain or self-reflective phrases
It would be great to see GGUF quantization support added to 1038lab/ComfyUI-QwenVL. The self-questioning descriptive output from joycaption beta one is quite persistent and hard to suppress.