Flux.2 Klein with Reference Latent Plus

I have been tweaking this ComfyUI workflow for a while and really like where it's at, so I thought I would share it. It uses Flux.2 Klein's ability to accept multiple image inputs to blend a character (usually just a face) and a scene, and create a photorealistic output, even if the scene reference is an illustration. This is great for placing specific characters you want in scenes you like, even if the reference scene is not in a photorealistic style. It runs just fine on my 3070 Ti with 8GB of VRAM using a 9B GGUF model.

A picture is worth a thousand words, so take a look at the sample images.

This workflow uses ReferenceLatentPlus to enable multiple image inputs:

https://github.com/shootthesound/comfyui-ReferenceLatentPlus

It also uses 1hewNodes to allow you to set up multiple character image and prompts references, then switch between them just by clicking a button:

https://github.com/1hew/ComfyUI-1hewNodes

I'm rocking this uncensored text encoder for Flux.2 Klein, and you should too:

https://huggingface.co/ponpoke/flux2-klein-9b-uncensored-text-encoder