Using image generator for vision test

Using image generator for vision test | Generated by AI

Home 2026.07

Question: Why does my ww vision-model test always get finish_reason=stop with empty response?

Answer:

The main issue is likely model type mismatch.

You tested:

--model black-forest-labs/flux.2-pro

FLUX.2-pro is an image generation model, not a vision-language model (VLM).

Your test sends:

image + text prompt
        ↓
expecting:
text response describing/analyzing image

But FLUX expects:

text prompt
        ↓
generated image

So the API can legally return:

finish_reason=stop
content=""

because there is no textual answer channel.

Use a real vision-language model instead:

Examples:

google/gemini-2.5-flash
google/gemini-2.5-pro
openai/gpt-5
anthropic/claude-sonnet-4
qwen/qwen2.5-vl-72b-instruct

For OpenRouter style testing:

ww vision-model test \
  --model google/gemini-2.5-flash \
  --max-tokens 1000

Your test image is also extremely small:

64x64 blue square

A VLM will probably answer something like:

“The image is a solid blue square.”

Also check your request handling. Many vision APIs require image content in this format:

{
  "role": "user",
  "content": [
    {
      "type": "text",
      "text": "Describe this image"
    },
    {
      "type": "image_url",
      "image_url": {
        "url": "data:image/png;base64,..."
      }
    }
  ]
}

Not:

{
  "prompt": "Describe image",
  "image": "..."
}

Your warning is unrelated:

RequestsDependencyWarning:
urllib3 ... doesn't match supported version

It only means your Python dependency versions are unusual. It won’t cause empty model output.

A good debugging sequence:

# 1. Test text model
ww vision-model test --model google/gemini-2.5-flash

# 2. Test image understanding model
ww vision-model test --model qwen/qwen2.5-vl-72b-instruct

# 3. Test image generation separately
ww image-model test --model black-forest-labs/flux.2-pro

So the short version:

FLUX is the wrong category. It generates images; it does not look at images and describe them. Use Gemini/Qwen-VL/Claude/OpenAI vision models for vision-model.

Back Donate