Text-to-Image

Text-to-image generation runs a vLLM-Omni worker with --output-modalities image.

Tested Models

To run a non-default model, pass --model to the launch script:

bash examples/backends/vllm/launch/agg_omni_image.sh --model AIDC-AI/Ovis-Image-7B

Launch

Launch using the provided script with Qwen/Qwen-Image:

bash examples/backends/vllm/launch/agg_omni_image.sh

Generate an Image

Image generation is available on two endpoints: /v1/chat/completions returns base64-encoded images inline in choices[].delta.content[].image_url, while /v1/images/generations returns a URL or base64 depending on response_format.

For /v1/images/generations and /v1/images/edits, Dynamo’s vLLM-Omni image handler accepts output_format values png (default), jpeg, and webp. It reports the encoded format in output_format; use jpeg for JPEG output. It reports size when every output has the same valid dimensions and returns null when dimensions are mixed or unavailable. This handler accepts response_format values url and b64_json; if omitted, the response uses base64. JPEG output composites transparent pixels onto white. This handler does not apply quality or background; requests that set either field are rejected with HTTP 400 rather than echoing values that were not used.

curl -s http://localhost:8000/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen-Image",
    "prompt": "A cat sitting on a windowsill",
    "size": "1024x1024",
    "response_format": "url"
  }'

Create a Variation with Image Edits

To generate a new image guided by an existing image, send a reference image and a prompt to /v1/images/edits. Use a model that supports image conditioning; text-to-image support alone does not establish editing support. This example uses black-forest-labs/FLUX.2-klein-4B with vLLM 0.31.0 and vLLM-Omni 0.31.0rc1.

Launch the reference-capable model from the repository root:

bash examples/backends/vllm/launch/agg_omni_image.sh \
  --model black-forest-labs/FLUX.2-klein-4B \
  --enforce-eager

In a second terminal, encode a local PNG named reference.png into a JSON request. The file is read on the client; the server receives a base64 data URL in input_reference.

python3 - <<'PY' > edit-request.json
import base64
import json
from pathlib import Path
reference = base64.b64encode(Path("reference.png").read_bytes()).decode("ascii")
print(json.dumps({
    "model": "black-forest-labs/FLUX.2-klein-4B",
    "prompt": "Create a watercolor variation of this image, preserving its main subject.",
    "input_reference": f"data:image/png;base64,{reference}",
    "size": "512x512",
    "n": 1,
    "response_format": "b64_json",
    "nvext": {"num_inference_steps": 4, "guidance_scale": 1.0, "seed": 42},
}))
PY
curl --fail-with-body -sS http://localhost:8000/v1/images/edits \
  -H 'Content-Type: application/json' \
  --data-binary @edit-request.json > edit-response.json

Decode the returned PNG:

python3 - <<'PY'
import base64
import json
from pathlib import Path
response = json.loads(Path("edit-response.json").read_text())
Path("edited.png").write_bytes(base64.b64decode(response["data"][0]["b64_json"]))
PY

To use a hosted reference instead, set input_reference to an HTTP(S) image URL that the worker can fetch. Do not use curl -F or the native vLLM-Omni image field through the Dynamo frontend; this endpoint currently requires the JSON input_reference contract.

See Also

• Text-to-Image with SGLang
• Text-to-Image with TensorRT-LLM
• Disaggregated Serving — GLM-Image (2-stage text-to-image)
• vLLM-Omni Configuration reference