bug: Unable to track image generation calls

[secsilm]

Describe the bug

When using OpenAI's image generation API, Langfuse does not display the trace.

To reproduce

import os
from dotenv import load_dotenv
from langfuse import openai

load_dotenv(".env.dev")

client = openai.OpenAI(
    base_url=os.environ["AI_GATEWAY_API_URL"],
    api_key=os.environ["AI_GATEWAY_API_KEY"],
    max_retries=3,
)


def call_llm(prompt: str, model: str = "gpt-image-1"):
    response = client.images.generate(
        model=model,
        prompt=prompt,
        n=1,
    )
    return response


if __name__ == "__main__":
    call_llm("Generate an image of a futuristic city skyline at sunset")

SDK and container versions

• sdk: 3.2.1
• container: 3.72.1

Additional information

No response

Are you interested to contribute a fix for this bug?

No

[dosubot]

Langfuse doesn't currently support automatic tracing for OpenAI image generation API calls like client.images.generate—this is a known limitation and only chat/completion endpoints are traced out of the box (docs). If you want to track image generation calls, you'll need to manually create traces using the low-level SDK methods. For example, you can use langfuse.generation to create a span around your image generation call, then end it with the output:

from langfuse import get_client

langfuse = get_client()
with langfuse.start_trace(name="image-generation") as trace:
    with langfuse.generation(
        name="openai-image-generate",
        model="gpt-image-1",
        input={"prompt": prompt},
        trace_id=trace.id,
    ) as gen:
        response = client.images.generate(
            model="gpt-image-1",
            prompt=prompt,
            n=1,
        )
        gen.end(output=response)

# Make sure to flush before exit if running in a script
langfuse.flush()

This approach is similar to workarounds used for other unsupported OpenAI endpoints (example). There’s no official example for image generation, but this pattern works for manual instrumentation. Keep an eye on future SDK releases for expanded endpoint support.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[secsilm]

@dosu how to also record usage so that langfuse can calculate cost automatically, or can I caculate the cost then pass it to langfuse?

[dosubot]

Langfuse can't automatically calculate cost for OpenAI image generation calls—you need to pass usage and cost info manually. The standard way is to provide a usage dictionary when creating or ending your generation span. For image generation, usage might look like this:

with langfuse.generation(
    name="openai-image-generate",
    model="dall-e-3",  # or your image model name
    input={"prompt": prompt},
    trace_id=trace.id,
    usage={
        "input": 1,           # number of images requested
        "unit": "IMAGES",     # usage unit
        "input_cost": 0.04,   # USD cost per image (set to your calculated value)
    }
) as gen:
    response = client.images.generate(
        model="dall-e-3",
        prompt=prompt,
        n=1,
    )
    gen.end(output=response)

You can calculate the cost yourself (using OpenAI's pricing for the model/size you use) and pass it as input_cost/output_cost. If you add a custom model definition in Langfuse with the same model name and price per image, and provide the usage count/unit, Langfuse can also compute cost automatically for you. More details and examples are in the Langfuse docs and changelog.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[secsilm]

for anyone interested, here is the final code:

import os

from dotenv import load_dotenv
from langfuse import get_client, observe, openai

load_dotenv(".env.dev")
lf_client = get_client()

client = openai.OpenAI(
    base_url=os.environ["AI_GATEWAY_API_URL"],
    api_key=os.environ["AI_GATEWAY_API_KEY"],
    max_retries=3,
)


@observe(as_type="generation")
def call_llm(prompt: str, model: str = "gpt-image-1"):
    response = client.images.generate(
        model=model,
        prompt=prompt,
        n=1,
    )
    lf_client.update_current_generation(
        model=model,
        usage_details={
            "input_text_tokens": response.usage.input_tokens_details.text_tokens,
            "input_image_tokens": response.usage.input_tokens_details.image_tokens,
            "output_tokens": response.usage.output_tokens,
        },
    )
    return response


if __name__ == "__main__":
    call_llm("Generate an image of a futuristic city skyline at sunset")

then go to your langfuse Settings -> Models -> Add model definition, set like this:

[hassiebp]

Thanks for sharing your solution!

[coletebou]

I've submitted a PR to add OpenAI image generation tracking to the Python SDK:

PR: langfuse/langfuse-python#1511

This adds automatic tracing for:

• images.generate()
• images.edit()
• images.create_variation()

Tracked data includes prompt, model, parameters (size, quality, style), output URLs/base64 images, and image count as usage metrics.

For those using the Vercel AI SDK (TypeScript), I've also submitted a PR to add telemetry support to generateImage:

PR: vercel/ai#12275

This would enable Langfuse to receive image generation spans via OpenTelemetry once merged.