Im getting just messy images.

[AbyszOne]

Idk what im doing wrong. Im using protogen fp16 with fp16 yaml. Then tried img2img, but all i got is this. Prompt "Pretty girl" or something. The most coherent thing i got is a very deformed main picture at 0.1 denoising, wich doesn`t make sense.

[comfyanonymous]

Can I see your workflow?

[AbyszOne]

Can I see your workflow?

I copied it verbatim from img2img example, with only yaml and euler A changes. I checked few times, but there is nothing different. https://comfyanonymous.github.io/ComfyUI_examples/img2img/
I got the cuda error and then fixed it as is explained. Not much less. Thats my first try. This is denoising 0.1:

[comfyanonymous]

I got your workflow from the images you posted. The issue is that your ProtoGen_X3.4 does not contain any CLIP model weights.

Use another checkpoint or you can download the SD1.5 CLIP model weights from here: https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/text_encoder/model.safetensors

Put them in the models/clip folder and use the CLIPLoader node to load them and connect it to your CLIPTextEncode nodes. Then it should work.

[AbyszOne]

Thank you. I'll try it later. My interest is to be able to experiment with multiple controlnets, that fulfill specific functions. As well as the possibility of using multiple img2img. I don't know if it's possible, but your gui seems flexible enough to try. When there is a native batch I will surely experiment with it.

[comfyanonymous]

Yes you can use multiple controlnets and run as many img2img passes as you want with different settings/models.

[AbyszOne]

It got better with your advice.
Regarding multiple simultaneous img2img, I have not found the mechanics. My interest is to achieve the same thing as controlnet + img2img, where one influences the other during the same inference. Maybe it would be useful to know how to achieve that first, since it's a process that actually works in the gui.

[comfyanonymous]

Take the controlnet example here: https://comfyanonymous.github.io/ComfyUI_examples/controlnet/

And instead of feeding the sampler an EmptyLatentImage feed it an image like in this example: https://comfyanonymous.github.io/ComfyUI_examples/img2img/

Is this what you want to do?

[AbyszOne]

I think that's the mechanics. I confirm it later. Now I was trying a way of "almost" double img2img, and got an image of what I wanted while respecting the light and color composition of the other. i`ll keep trying.