Kontext Chat Guide to Editing Images Through Conversation

Kontext Chat is Replicate's browser interface for editing images with Black Forest Labs' FLUX.1 Kontext pro. Open the hosted demo, provide a Replicate API token and upload an image to begin. App repository Frontend

What are the key facts about Kontext Chat?

What kinds of image edits can you make with Kontext Chat?

Upload a reference or select a starter image, then describe the desired edit. Frontend

Replicate's launch roundup describes object changes, background changes, text replacement and style transfer as Kontext use cases. Those are useful starting points for evaluating an editing session. Announcement

Think of a session as a sequence of proposed image revisions. After each response, decide whether the current image is suitable as the starting point for another edit, rather than accepting the sequence automatically.

For example, start with a product photograph and request a new background. Once you approve the product's appearance, try a different surface or lighting treatment. Keep a saved copy of the original for comparison.

This workflow can help a collaborator describe a visual change in ordinary language. Agree on the required result beforehand: “place the cup on a wooden table” is easier to review than “make this look more professional.”

Use small requests during the first session. You want to learn which changes the tool handles well for your images before asking it to alter the setting, materials and composition at the same time.

How does Kontext Chat handle references, uploads and tokens?

The frontend selects the latest image in the conversation and sends it with the current prompt. The backend passes those fields to black-forest-labs/flux-kontext-pro; this is the implementation's image-editing mechanism. Frontend Backend

Do not rely on a long conversation to substitute for a visible reference. If an earlier requirement still matters, include it in your current request or restart from an approved image that visibly satisfies it.

The frontend labels its upload limit 10 MB and enforces 10 × 1024 × 1024 bytes (10 MiB). It reduces uploads above one megapixel to approximately that pixel count, keeps smaller dimensions unchanged, and encodes the upload as JPEG. This app-specific processing matters when your source contains fine lettering or small product details. Frontend

The current backend requires a token in the request header and stores generated images through Cloudflare Images. Running a local copy of the interface does not move model inference onto your own GPU. Backend

The README omits image-storage bindings used by the current backend. Inspect that implementation when configuring a deployment. Repository Backend

How do you edit an image with Kontext Chat?

Begin with the hosted app

1. Open the official Kontext Chat demo linked from Replicate's repository. Obtain a Replicate API token from your account and provide it through the app's token form.
2. Upload a suitable image or choose a starter image. Begin with a source whose most important details are clearly visible.
3. Describe one change and name the details you need preserved. Submit the request and inspect the returned image.
4. Save an approved result before continuing. Start a fresh sequence from a saved reference when you need to explore another direction.

Kontext pro runs against the supplied Replicate account and is billed through that hosted service. Set a session budget before experimenting, and inspect failed requests before repeatedly submitting the same change. Backend Replicate billing

An example first request is: “Change the wall behind the chair to pale blue. Preserve the chair's shape, fabric and position.” Inspect both the changed wall and the supposedly unchanged chair before proceeding.

If the chair changes shape, return to the source image and simplify the instruction. If the edit succeeds, save it with a description such as “approved blue wall” so later comparisons have an unambiguous reference.

Use the underlying model directly

For a scripted edit, use Replicate's Python client with REPLICATE_API_TOKEN configured. The model identifier and input fields match the app backend. Backend Python client

import replicate

with open("reference.png", "rb") as reference:
    output = replicate.run(
        "black-forest-labs/flux-kontext-pro",
        input={
            "input_image": reference,
            "prompt": "Change the wall to pale blue. Preserve "
                      "the chair's shape, fabric and position.",
        },
    )
print(output)

Install the replicate package in your Python environment before running the example. Save the returned image, review it, and explicitly choose the source for the next call; this script does not reproduce the app's conversation interface.

Make a small review checklist for the task. For furniture imagery, compare silhouette, legs, seams and fabric pattern. For packaging, compare shape and every required label before assessing whether the new background looks suitable.

Record the model identifier and instruction with each accepted output. If you later switch the interface or model, the same source image and requested change give you a useful comparison case.

For medium and texture changes, use the sibling Kontext style-transfer guide. For BFL's own browser access, see Kontext online.

How does Kontext Chat compare with BFL Playground and Kontext dev?

For a local graph workflow, consult the ComfyUI complete guide. Choose the environment based on how you need to manage references and retain results.

What else should you know before using Kontext Chat?

Which image does Kontext Chat edit next?

Kontext Chat selects the latest image in the conversation and sends it with the current edit prompt to Kontext pro. Re-upload an earlier approved image when you want that image to be the starting point. Frontend Backend

Do I need a Replicate API token?

Yes, Kontext Chat requests a Replicate API token in its frontend and checks for it in its backend. Use your own account's access and billing settings when planning a session. Frontend Backend

Can I run the image model on my laptop?

Kontext Chat uses hosted Kontext pro even if you run its interface locally. Kontext dev is a different downloadable checkpoint with a separate model license. Repository Dev card

Sources

• Replicate Kontext Chat repository
• Replicate launch roundup
• Published frontend source
• Published backend source
• Replicate model API reference
• Replicate billing documentation
• Replicate Python client documentation
• BFL Kontext overview
• BFL image-editing documentation
• Kontext dev model card
• Official hosted demo

Related guides on PromptZone

• Local LLMs 2026: Run Llama, Mistral, Qwen on Your Hardware