comfyanonymous / ComfyUI

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface.
https://www.comfy.org/
GNU General Public License v3.0
55.17k stars 5.83k forks source link

Unlike SDXL's results, SD1.5's results are not clean. #3194

Open poisonzafina opened 7 months ago

poisonzafina commented 7 months ago

스크린샷 2024-04-03 211730

My ultimate goal is to use animatediff as comfyui.

However, the SDXL model is the only one that produces stable results with comfyi.

Anyway, I don't get good results with the model of SD1.5.

Untitled-1

Is this problem only happening to me?

Is there a problem with my workflow or doesn't the original comfyui produce good results?

Is there a solution that works for me?

Using a translator, the description can be awkward. Help me, please.

NeedsMoar commented 6 months ago

Your negative embeddings need to be written with "embedding:" in front of them, which could be causing a lot of difference since you're using 3 negative embeddings.

Also, most LoRAs need an activation prompt in the text somewhere. A1111 sort of hijacks this I think by allowing you to insert the lora in the prompt, but many of them don't do anything at all in Comfy without their keywords in the prompt.

NeedsMoar commented 6 months ago

Also if you need to turn off LoRAs, right click them and click bypass rather than setting them to 0.0. If you use the arrows to set them you might actually be ending up with 0.00000000001. If I accidentally put an SDXL lora in a chain with a 1.5 model it'll spew a bunch of tensor size mismatch errors to the console but still affect the results. I'm not sure why this is, since size mismatches literally anywhere else error and stop the workflow on that node. Anyway, disabling it is the better idea here. When you need to turn it back on it isn't exactly intuitive but you just right click and click bypass in the menu again and it turns back on, there's not a re-enable or anything like that.