A pretty simple workflow with a few common features. There are some important (more or less) notes inside the workflow. It is designed for 'legacy' mode. For now, I still think that Nodes 2.0 are mostly buggy and slower garbage, even though they have a few cool things.
I tried Fun-Control ControlNet. For some reason, I always get a black image with any model and any settings, even in the official template.
Description
Replaced "Load image with metadata" node in my own version. Crystool extension is no longer needed for this workflow;
Replaced the text concat node in my own version. The Was-ns extension is no longer needed for this workflow;
Added bookmarks;
Edited Model links note
FAQ
Comments (4)
The black image issue is from sage-attention, disable it and it should work. For me, I'm using comfy in browser and running a .bat w/ sage attn to start up, so i made a second .bat for disabling sage for QWEN gens. the bat file looks like this:
python main.py --use-quad-cross-attention
cmd /k
this multiple ways of doing this, but thats a workaround that solved it for me.
I don't use SageAttention because my video card doesn't support it, so I don't have those modules installed. I tried the argument anyway, and got a black screen as a result. But thanks for letting me know.
No it isn't. As he said, he doesn't use SageAttention and neither do I on a Mac and I get black images using ControlNet as well. The main issue isn't SageAttention, it's Qwen itself. Adding in ControlNet causes Qwen to struggle with competing guidance constraints, the way to compensate for this is to adjust the strength parameter on the ControlNet lowering it until you get a proper image and not a black one. 2.2is way too high for Qwen, it needs to be down at 0.5 or even lower. You also need to adjust the step count, too. Having a low step count doesn't allow the checkpoint room to do what is needed, it may be fast with the lightning LoRa but that really needs to be put on the back burner, so to speak if you're using ControlNet. Make it 8 - 12 steps at least. You'll need to play around and see what works.
@mrmrswiggly612 I tried this before, but let's try it again. Steps don't matter; if even the first image is black, there's no reason to keep going. I tried different models and different ControlNet unions. With and without LoRA, different CFG, different ControlNet power settings and preprocessors. Nothing worked for me. Probably my card is too old for it; even other models work pretty fine. Actually, I stopped using Qwen as well as Qwen Edit. In most cases, Z-Image and Flux2 FLain 9B give me better results and are much faster.