Bonus: llm-vision only (disabled reference latents for the DiT):

no LoRA	style_reference LoRA

Maybe this could be an option? (what would be the best UI for that?)

... and now I find this one, which seems to use different setup:
model: https://huggingface.co/conradlocke/krea2-identity-edit
implentation: https://github.com/lbouaraba/comfyui-krea2edit

• refs are concatenated before the target image rather than after, but since non-causal attention permutation invarient, I'm pretty sure this doesn't matter as long as the RoPE ids are the same
• Uses a slightly different prompt format (probably not that important)
• Does not force timestep=0 for the refs

Just disabling the separate timestep for the refs seems to be enough to make it work pretty well

Thank you for your contribution.