Stability AI releases a new model, two in fact. It is called Cosine-Continuous Stable Diffusion XL (CosXL) and strangely there is not too much ado about this.
But these models are actually pretty interesting.
First, if you remember the Pix2Instruct SD model, CosXL uses a similar approach. You are using an image as an addition to the positive and negative prompts.
Second, it performs some tonal magic approximating HDR in a way, that allows the creation of very dark and/or light images.
Pros:
It creates beautiful tonality and colors
It is fast, compared to SDXL with several Controlnets applied to achieve similar results
The CosXL Edit model is a beast
Great rendering of light conditions
Cons:
It has anatomy and proportions issues (in both models) that you need to fight with
Wonky hands included
I am not sure about the performance on lower VRAM (let me know in the comments if I am mistaken)
Especially the 'Edit' model can be a good tool for retouching even in its current state. I have tested the 'Base' model a lot, and it tends to generate a painterly look—which can be a plus, depending on what you are trying to achieve. The way CosXL works with lights and tones can help artistic and abstract compositions and styles.
I have created some workflows to test both models (because why not to), link to my original article on running CosXL on StableSwarmUI or Comfy> sandner.art | Cosine-Continuous Stable Diffusion XL (CosXL) on StableSwarmUI
So, try it if you dare delve into the ComfyUI:
