Flux Kontext: Change Clothes in ComfyUI – Swap Any Outfit by Prompt (2026)

Drop a photo into ChatGPT and ask for a red leather jacket – you get a new person back. Different face, different pose, different background. Fine for a quick sketch, useless for anything else.

Flux Kontext does the opposite: it changes only what you name. The white t-shirt becomes a leather jacket – face, posture, jeans and background stay pixel-for-pixel identical. That's the whole difference between "a new image" and "editing this image".

This guide walks the complete path – tested locally on an RTX 3060 with 12 GB VRAM, entirely free, no cloud required.

As of July 2026 · ComfyUI (current) · Flux.1 Kontext Dev as GGUF · custom node "ComfyUI-GGUF"

What you'll get:

• Why Kontext preserves the subject – while ChatGPT reinvents it
• The 4 model files and exactly where they go
• The ComfyUI workflow in 4 steps – with the GGUF trick for 12 GB cards
• The right prompt wording for swapping clothes
• Fixes for when the outfit won't change

Kontext vs. "ChatGPT could've done that too"

The key concept is image conditioning. Kontext receives your original image not as a vague description but as a fixed reference point (the ReferenceLatent node in the workflow). The model holds everything you don't mention constant and computes only the change you name.

In practice:

• One photo, one prompt, one change – no redrawing of the person.
• Face, posture, lighting and background are preserved.
• You can edit the same image several times in sequence (first the jacket, then the hair, then the background).

That's the creative lever: not "AI paints something", but precise image direction. Product shots, character variants, model looks without a second shoot – all from one source image.

Requirements

Empfohlenes Tool · Cloud-GPU

Keine starke Grafikkarte? Führe ComfyUI in der Cloud aus.

Flux & SDXL sind speicherhungrig. Statt ~1.800 € für eine eigene Karte mietest du bei RunPod eine 24-GB-GPU ab ca. 0,50 €/Stunde – und zahlst nur, was du nutzt.

GPU mieten & Flux starten →Affiliate-Link · für dich ohne Mehrkosten

Step 1: Install the "ComfyUI-GGUF" custom node

The GGUF loader isn't a stock node – it comes via the Manager:

1. Launch ComfyUI → click Manager on the right.
2. Open the Custom Nodes Manager → search for ComfyUI-GGUF → Install.
3. Fully restart ComfyUI.

Skip this and the "Unet Loader (GGUF)" node is missing later and the workflow stays red.

Step 2: Download the four models and place them

Flux Kontext has four building blocks: the main model, two text encoders and a VAE.

The folder structure ends up like this:

📁 ComfyUI/
└── 📁 models/
    ├── 📁 unet/
    │   └── flux1-kontext-dev-Q4_K_M.gguf
    ├── 📁 text_encoders/
    │   └── t5xxl_fp8_e4m3fn.safetensors
    ├── 📁 clip/
    │   └── clip_l.safetensors
    └── 📁 vae/
        └── ae.safetensors

Step 3: Load the Kontext workflow – with the GGUF trick

You build nothing by hand. ComfyUI ships the workflow as a template:

1. Click Workflow → Browse Templates at the top.
2. Pick category Flux → "Flux.1 Kontext Dev (Basic)". The full graph loads automatically.
3. Now the one important move for 12 GB cards: delete the "Load Diffusion Model" node. Right-click the canvas → Add Node → bootleg → Unet Loader (GGUF) → select flux1-kontext-dev-Q4_K_M.gguf → reconnect the MODEL output to the KSampler.

Quickly check that the remaining loaders are correct:

• DualCLIPLoader → t5xxl_fp8_e4m3fn + clip_l, type: flux
• VAELoader → ae.safetensors

Step 4: Load the image, write the prompt, generate

1. Load the image. Upload your source photo in the Load Image node.

2. Write the prompt. Enter the change in the positive prompt – English works most reliably with Flux:

change the white t-shirt into a red leather jacket

3. Set the parameters. The tested values for Kontext:

4. Generate. Click Queue Prompt.

Result: the t-shirt is now a red leather jacket – face, pose, jeans and background unchanged.

The right wording: how to swap clothes cleanly

Kontext is precise when you are. Three rules:

• Name the garment + material + color. "a red leather jacket" beats "nice jacket" by a mile.
• One change per run. To change the jacket and the pants, do two runs – each step stays controlled.
• Pin it explicitly on drift: if the model clings to the subject, add keep the same face, pose and background.

A few proven patterns:

Outfit won't change? Quick fixes

No 12 GB card? Use the cloud

If your GPU can't keep up or lacks VRAM: run the same workflow on a cloud GPU – an RTX 4090 (24 GB) from ~$0.59/h renders Kontext in seconds and with no quant compromise. The full setup is in our hub:

👉 Set up ComfyUI on RunPod (Network Volume)

Empfohlenes Tool · Cloud-GPU

Keine starke Grafikkarte? Führe ComfyUI in der Cloud aus.

Flux & SDXL sind speicherhungrig. Statt ~1.800 € für eine eigene Karte mietest du bei RunPod eine 24-GB-GPU ab ca. 0,50 €/Stunde – und zahlst nur, was du nutzt.

GPU mieten & Flux starten →Affiliate-Link · für dich ohne Mehrkosten

Conclusion & next steps

Flux Kontext isn't image generation, it's image direction: you keep your subject and swap individual elements on purpose. Changing clothes is the easiest entry point – the same technique carries hairstyles, backgrounds and product variants.

• 📥 Download: The finished workflow JSON is at the end of the article – created in ComfyUI via Workflow → Export, ready to run.
• 🔗 Missing the basics? → Install Flux.1 Dev with ComfyUI
• ⏭️ Next in the series: "Flux Kontext: change hairstyle" and "Flux Kontext: swap the background".

One photo, one sentence, a new outfit – and everything else stays exactly as it was.