CosXL is a new base available for research (non-commercial use). There is also an SDXL instruct Pix2pix model we can now test and experiment with. These can both be used with comfyUI.
https://huggingface.co/stabilityai/cosxl
There are two new checkpoints, first is the Base model cosxl.safetensors. This can be used like a normal SDXL checkpoint.
For the Instruct Pix2Pix functions, You will need to place the second checkpoint named cosxl_edit.safetensors in the /models/checkpoints folder. It works just like previous IP2P models, allowing easy image editing with simple prompts. I have included the default workflow with some QOL tweaks.
often SDXL workflows only use Clip G, however using both Clip G+L improve the outputs especially in terms of Lora accuracy, where they may be trained on one either or both text encoders.
v2 contents:
basic workflow for cosXL & cosXL_edit
HD versions included in pack (iterative upscale)
prompt examples:
a winter scene
rembrant style
oil painting
watercolor
Description
FAQ
Comments (5)
cosxl_edit.safetensor in Pix2Pix is returning a grainy green image that barelly resembles the original image. Any hint for me, please?
in one of the videos i showed you can right click on the IP2P node, and "ungroup" the group node.
There are some SDXL lora/models that required a different mode. Aside from this, with the correct model loaded it should work "out the box" with my workflow values, all i can suggest is to update your comfyui and all nodes to be sure.
@driftjohnson Oh! Thanks for your reply. I figured it. The cosxl_edit was corrupted. I doenloaded it again and it worked! Thank you so much!
Does it only change the color of the scene? Is it capable of deleting objects from the scene or changing people's appearance?
Cosine SDXL is very much redundant now, this was improved over SDXL in the true representation of Black, not just very dark blue. It's a very old release now
If you wanted to remove objects, i have other workflows that can do this in SDXL.