FLUX.2 ComfyUI Reference Image Guide

Use a FLUX.2 dev-compatible workflow and matching model components. Give each reference a clear role in the prompt, keep the input order consistent, and begin with a small controlled test.

Watch on YouTube · Video published 2025-11-26 · Companion reviewed 2026-09-23

Which FLUX.2 does this cover?

The November 2025 video demonstrates the original FLUX.2 dev workflow. It is not a setup guide for every later FLUX.2 variant. Check the exact model revision before borrowing its encoder, VAE, guidance or step count. The video’s ten reference slots belong to the supplied graph; they are not a claim that every FLUX implementation exposes the same controls.

Prepare the model files

models/diffusion_models/flux2_dev_fp8mixed.safetensors
models/text_encoders/mistral_3_small_flux2_bf16.safetensors
models/vae/flux2-vae.safetensors

Use the official FLUX.2 dev setup guide for the matching downloads and current template. Read each model’s license before choosing a client use case. The workflow download does not include model weights.

Generate a simple test

Load the workflow and select the matching diffusion model, text encoder and VAE. Resolve missing nodes before loading all your references.
Start with text-to-image or one reference. Use batch size 1, a fixed seed and a resolution your GPU can support.
The video shows 50 steps and a 1920 × 1088 canvas, with guidance around 3–4 discussed as a starting point. These are historical demonstration settings, not a requirement or a low-memory promise.
Run once, inspect the main subject and composition, then save the working graph. Add reference images one at a time.

Assign a job to each reference

Use a short input note beside your files: reference 1 = person, reference 2 = location, reference 3 = clothing, reference 4 = pose. Match that order to the connected graph. If you remove or bypass an input, update the prompt so its numbering still makes sense.

Suggested multi-reference prompt

Create a realistic editorial photograph using the person in reference image 1, the room in reference image 2, and the jacket in reference image 3. Use the seated pose from reference image 4. Keep the person in the foreground, facing the camera, with soft window light from the left. Preserve the room layout and avoid adding extra people.

This is a practice prompt, not a promise of exact identity or pixel-perfect copying. A reference can influence appearance, layout or motion cues, but the output still needs review. Use images you are entitled to process.

Improve the prompt before adding complexity

Write the scene as connected sentences. Explicit relationships are often more useful than an unrelated pile of quality words: “the red bag is on the chair behind the person” tells the model where objects belong. If two references conflict, decide which one controls each feature and say so. Keep essential constraints near the relevant object.

Do not assume Stable Diffusion weighting syntax or negative-prompt habits transfer unchanged. Check the nodes in this FLUX graph and test any syntax you rely on. Longer text can introduce competing instructions, so remove contradictions before increasing guidance.

Troubleshooting

Out of memory: reduce output dimensions, reference count and competing GPU activity. Quantized alternatives require their own compatible loaders.
References seem ignored: check that the image branch is connected and enabled, then simplify to one reference.
Identity or clothing drifts: inspect the source quality, prompt role assignment and conflicting visual cues. A fixed seed does not guarantee identity.
Old graph fails after an update: save a backup, compare with the official current template, and record the model and node revisions used.

FLUX.2 in ComfyUI: setup and prompting with reference images