Request for the prompt upsampler's system prompt
Hi Pruna team,
Could you share the system prompt used by the prompt upsampler (prompt_upsampler,
based on Qwen3.5-9B) in P-H3-Image? I'd like to reproduce its behavior outside the
diffusers pipeline.
If it was trained without a fixed system prompt, a confirmation of that would also
be helpful.
Thanks!
+1 to this. Getting a system prompt would open it up to different pipeline integrations. Thank you PrunaAI
+1
Because I suspect it's censored. When I manually type a “spicy” prompt but then run it through the enhancement model, it just censors everything and all it does is describe the image (at least in image edit mode).
+1
Because I suspect it's censored. When I manually type a “spicy” prompt but then run it through the enhancement model, it just censors everything and all it does is describe the image (at least in image edit mode).
Are you using a Heretic Qwen model for the prompt enhancement or the default Qwen model?
The way I do it is: I have the Heretic model expand my prompt and then feed it to the default model.
+1
Because I suspect it's censored. When I manually type a “spicy” prompt but then run it through the enhancement model, it just censors everything and all it does is describe the image (at least in image edit mode).Are you using a Heretic Qwen model for the prompt enhancement or the default Qwen model?
The way I do it is: I have the Heretic model expand my prompt and then feed it to the default model.
No, I'm using the original. I thought that in the case of p h3 it used a special, fixed model that couldn't be changed. I'll try the heretic variant. Which one are you using exactly? Thanks.