[Bug]: Azure AI FLUX.2-pro image edit silently ignores reference image — wrong JSON key "image" instead of "input_image"

[marcelgaweda-scieneers]

What happened?

When using azure_ai/flux.2-pro for image editing via /v1/images/edits, the reference image is silently ignored. FLUX generates a completely new image from the prompt instead of editing the provided reference image.

The root cause is in litellm/llms/azure_ai/image_edit/flux2_transformation.py. The transform_image_edit_request method sends the base64-encoded image under the JSON key "image", but the Black Forest Labs API on Azure AI Foundry expects "input_image".

The API does not return an error. It simply ignores the unknown "image" field and generates a new image from scratch, making the bug hard to detect.

I verified this by patching line 113 from "image" to "input_image" in a running LiteLLM proxy container. With the fix, the reference image is correctly preserved and edited. Without it, a completely different image is generated.

References:

• Azure Foundry FLUX Guide
• BFL Image Editing Docs

Steps to Reproduce

1. Configure azure_ai/flux.2-pro in config.yaml:

- model_name: flux.2-pro
  litellm_params:
    model: azure_ai/flux.2-pro
    api_base: "os.environ/AZURE_AI_API_BASE"
    api_key: "os.environ/AZURE_AI_API_KEY"
    api_version: "preview"
  model_info:
    mode: "image_generation"

2. Send an image edit request:

import os
import litellm

response = litellm.image_edit(
    model="azure_ai/flux.2-pro",
    image=open("car.png", "rb"),
    prompt="Make the car red.",
    api_base=os.environ["AZURE_AI_API_BASE"],
    api_key=os.environ["AZURE_AI_API_KEY"],
    api_version="preview",
)

3. The returned image is a completely new image (different car, different scene) instead of the reference image with the color changed.

4. Patch flux2_transformation.py:113 from "image" to "input_image", restart the proxy, and repeat. The reference image is now correctly used.

The fix (one-line change in flux2_transformation.py:113):

- "image": image_b64,
+ "input_image": image_b64,

Secondary issue: _convert_image_to_base64 discards all but the first image from a list (line 126-129). The BFL API supports up to 8 reference images via input_image, input_image_2, ..., input_image_8.

Relevant log output

No errors logged. The API returns HTTP 200 with a valid but wrong image.

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on?

v1.82.3-stable (also verified on current main branch, same code, same bug)

Twitter / LinkedIn details

No response