Flux Kontext Character Consistency in ComfyUI: Full Workflow Guide

Getting the same character to show up correctly across ten different images used to mean training a LoRA on that character first. FLUX.1 Kontext Dev changes this — hand it one reference photo and a text instruction, and it generates a new pose or scene while keeping the face, outfit, and hairstyle close to the source. No training required.

This guide walks through the exact workflow: which nodes to connect, what each setting does, and how to write instructions that keep identity intact instead of drifting into a different-looking person by the third generation.

If you're working with Qwen Image Edit or Krea 2 instead — including multi-character scenes with two or more people in one frame — see the general ComfyUI character consistency guide or the Krea 2 character consistency guide. This page is specific to the Kontext node chain, which is different from both.

Tested on: ComfyUI v0.3.42, RTX 4090 (24 GB VRAM).
Minimum VRAM: 11 GB (fp8 model).

How Identity Is Preserved

Kontext carries your reference photo into the sampler through a single path — and every node in that path has to be connected correctly, or identity breaks:

• FluxKontextImageScale — resizes and pads your reference photo to one of the resolution buckets Kontext was trained on. Feeding it an arbitrary resolution without this node produces warped or cropped identity in the output.
• VAEEncode — converts that scaled reference image into a latent.
• ReferenceLatent — merges the VAE-encoded reference latent into your positive conditioning, right before it reaches the sampler.

Your instruction (encoded through a standard CLIPTextEncode, using the T5-XXL + CLIP-L pair) still matters — it tells Kontext what to keep and what to change — but the latent path is what actually anchors the face.

Example: One Reference, Three Scenes

Below is one reference photo and three separate generations from that same reference — swap in your own screenshots here. Click any image to zoom.

🔍 Click to zoom

🛋️ Same reference · 1024×1024 · steps 20 · guidance 2.5

🔍 Click to zoom

New Pose + Setting — Elegant Garden

Copy Prompt▶View Instruction

☕ Same reference · 1024×1024 · steps 20 · guidance 2.5

🔍 Click to zoom

New Pose + Setting — Luxury Hotel Lounge

Copy Prompt▶View Instruction

🌃 Same reference · 1024×1024 · steps 20 · guidance 2.5

🔍 Click to zoom

New Pose + Setting — Beach Sunset

Copy Prompt▶View Instruction

What You Need

Kontext ships as separate model, text-encoder, and VAE files — ComfyUI does not bundle them into one checkpoint. Download these before you start.

Where to Place the Files

Each file type goes in a different subfolder under ComfyUI's models/ directory. A file in the wrong folder means the dropdown in that node stays empty.

Download the Workflow

How to Load the Workflow JSON into ComfyUI

1. Download the .json file to your computer.
2. Open ComfyUI in your browser — usually at http://127.0.0.1:8188.
3. Drag the .json file directly onto the ComfyUI canvas. Drop it anywhere on the empty grey area.
4. The workflow loads automatically. You should see the Model Loaders, Size, and Sampling groups appear on the canvas.
5. If you see only red error blocks, jump to the Troubleshooting section below.

🔍 Click to zoom

Workflow Structure: What Each Node Does

Here's what each node in the chain controls, left to right:

Step-by-Step Setup

1. In DualCLIPLoader, select t5xxl_fp16.safetensors and clip_l.safetensors, and confirm type reads flux.
2. In VAELoader, confirm ae.safetensors is selected.
3. In UNETLoader, select your Kontext model file. Set weight_dtype to fp8_e4m3fn here if you're on a card under 16 GB.
4. Click LoadImage and upload your reference photo. Use a clear, well-lit, front-facing shot — side profiles and heavy shadows make identity lock harder.
5. Open the CLIPTextEncode node and write your instruction — see the next section for the exact structure.
6. Confirm the width/height PrimitiveInt nodes feed both EmptySD3LatentImage and ModelSamplingFlux — this workflow defaults to 1024×1024.
7. Click the orange Queue Prompt button. A progress bar appears below it. On an RTX 4090 at 20 steps, generation takes roughly 15–30 seconds.

How to Write the Instruction

Structure every instruction in two parts: what to preserve, then what to change. Lead with the preservation language — it carries more weight at the start of the prompt than buried at the end.

KEEP the same [subject] and change the pose and setting: [new pose, new location, new framing — be specific about posture, hands, expression, camera angle]. Preserve exact facial identity, makeup, hairstyle, body proportions, clothing, accessories, and all fine details from the reference image.

Worked example from this workflow:

KEEP the same woman and change the pose and setting: she is now sitting gracefully on a luxurious cream-colored sofa inside an elegant five-star hotel lounge with tall marble columns, crystal chandeliers, and warm golden lighting. She sits with one leg elegantly crossed over the other, her left arm resting comfortably along the back of the sofa while her right hand gently holds a porcelain coffee cup on her lap. Calm, confident expression with a soft smile, looking slightly toward the camera. Preserve exact facial identity, makeup, hairstyle, body proportions, clothing, accessories, and all fine details from the reference image.

There's no separate negative prompt in this workflow — Kontext's conditioning runs through ReferenceLatent alone, so unwanted patterns are usually fixed by adjusting the instruction itself rather than rejecting them in a negative field.

Settings That Matter

🔍 Click to zoom

FLUX.1 Kontext Dev is a full, non-distilled model, so it needs more steps than a Turbo-class model like Krea 2. The max_shift and base_shift values in ModelSamplingFlux are wired to your width/height so the sampling schedule stays matched to your output resolution automatically — leave them unless you know you need to tune them.

Going Further: Stacking a Style LoRA

Once identity holds up reliably, you can stack a style LoRA on top of the Kontext model to push the output toward a specific look — cinematic grading, an illustration style, a particular photography aesthetic — without touching the reference-latent chain that holds the face.

Need multiple characters in the same frame instead of one at a time? That's a different node setup — see the multi-reference section of the general ComfyUI character consistency guide.

"This node type does not exist" (red node in workflow)

The workflow uses a custom or newer core node you haven't installed. FluxKontextImageScale, ReferenceLatent, and ModelSamplingFlux are relatively recent core additions.

1. Click Manager in the top menu.
2. Click Install Missing Custom Nodes.
3. Restart ComfyUI completely.

If nodes still show red, update ComfyUI itself.

The character's face looks different in the output

This usually means the reference-latent chain isn't reaching the sampler correctly, or the reference photo is a poor angle for the model to lock onto.

1. Confirm FluxKontextImageScale → VAEEncode → ReferenceLatent is fully connected, not left disconnected anywhere in the chain.
2. Confirm ReferenceLatent receives both the positive conditioning and the VAE-encoded latent.
3. Try a clearer, more front-facing reference photo.
4. Make sure the KEEP instruction comes first in your prompt, not after the scene description.

Out of memory / CUDA error during generation

The full-precision model is roughly 24 GB and can exceed VRAM on cards under 24 GB.

1. Switch to flux1-dev-kontext_fp8_scaled.safetensors, or a GGUF quant if you're under 12 GB.
2. Set weight_dtype on the UNETLoader to fp8_e4m3fn or fp8_e5m2.
3. Lower your output resolution (try 768×768 instead of 1024×1024).
4. Close other GPU-heavy applications before queuing.

Frequently Asked Questions

Published: 2026-07-14 · Last updated: 2026-07-14 · Tested on RTX 4090 (24 GB VRAM) · ComfyUI v0.3.42 · Model: flux1-kontext-dev.safetensors