Multi-Reference with ReferenceLatent in the Flux.2

This workflow in ComfyUI is an advanced setup for image generation with the Flux model.2 Dev, focused on precise control through multiple reference images.
Works as follows:

Up to three reference images are loaded and resized to around 1 megapixel (together keeping the proportion).
Each image is encoded in latent using the official Flux VAE.2.
The resulting latents are applied sequentially via chained ReferenceLatent nodes, creating a cumulative effect: the influence of each reference is added to the previous conditioning.
O prompt textual (positivo) is applied first, with adjustable guidance (aqui em 4.0), and then receives the references in the order of the chain — the last reference in the chain tends to have the strongest influence on the main subject.
The generation occurs in a latent void of 720×1280 (formato retrato), com 20 steps, sampler Euler e seed fixo (pode ser alterado).
Includes automatic preview: the three resized reference images are concatenated into a grid and saved along with the final result.

It is ideal for those who want to combine pose/environment of an image, rosto/character from another and specific elements (like objects or posters) de uma terceira, all in a controlled way and without depending only on the textual prompt. The reference images and the prompt are fully customizable — each user can exchange everything according to the desired project.