Skip to main content

Profile Image Editing with AIPerf

First use Build from Source and Build a Bundle. This example assumes the existing Qwen Image Edit 2511 profile has been built as qwen-image-edit-2511.bundle and passed its family's correctness checks. Build for the example reference image's 640×382 input; the requested output is 1024×1024, 20 steps, CFG 4, seed 42. The Qwen editing builder uses its image dimensions for the conditioning input.

Run both terminals in your prepared GPU environment, from the repository root. trtmc-server must be on PATH.

1. Install AIPerf​

python -m pip install -e '.[serve]' -C py-only=true
python -m venv .venv-aiperf
.venv-aiperf/bin/pip install aiperf==0.13.0

2. Start the built model (terminal 1)​

trtmc-server ./qwen-image-edit-2511.bundle \
  --model-name qwen-image-edit-2511 \
  --startup-timeout 1800 --request-timeout 1800 \
  --records artifacts/aiperf-edits-server.jsonl

Wait for Uvicorn running on http://127.0.0.1:8000. Loading a large bundle can take several minutes. Add --runtime-root when native libraries are in a separate directory.

3. Benchmark it (terminal 2)​

Prepare the existing example photo as a lossless PNG. Its pixels and dimensions stay unchanged; this avoids a second lossy JPEG encoding in the client.

python - <<'PY'
import json
from pathlib import Path
from PIL import Image
Image.open("families/qwen_image/tests/data/test_img.jpeg").save("/tmp/edit-reference.png")
row = {"text": "Turn the image into a watercolor illustration while preserving the main subject.",
       "image": "/tmp/edit-reference.png", "extra": {"negative_prompt": " "}}
Path("/tmp/image-edits.jsonl").write_text(json.dumps(row) + "\n")
PY
.venv-aiperf/bin/aiperf profile \
  --url http://127.0.0.1:8000 --model qwen-image-edit-2511 \
  --endpoint-type image_edit --tokenizer builtin \
  --input-file /tmp/image-edits.jsonl --custom-dataset-type single_turn \
  --extra-inputs size:1024x1024 num_inference_steps:20 guidance_scale:4 cfg_scale:4 seed:42 \
  --concurrency 1 --warmup-request-count 1 --request-count 5 \
  --export-level raw --artifact-dir artifacts/aiperf-edits

AIPerf uploads the image and instruction using multipart POST /v1/images/edits. Each response contains a complete edited PNG in data[0].b64_json, usable after the server removes its temporary files. Raw exports retain the response. Server records bind the uploaded reference hash and dimensions, instruction hash, controls and output PNG hash to the same request ID.

Read client latency and images/sec in profile_export_aiperf.json. Server model_call_ms measures public_task_call_wall, including the family's image conditioning and generation. Uploads, image-container decoding, PNG encoding, HTTP transfer, build, loading, startup and warmup are outside steady-state Task timing. Keep client latency and Task time separate.

Use another built editing model or reference​

Select a bundle declaring image_edit or SDK images_text_to_image_edit and consuming one reference image. Replace the bundle, API name, reference, instruction and explicit controls with those supported by its existing full profile. Follow the owning family's input/build geometry requirements; input and output dimensions can differ. A plain image generator does not become an editing model by receiving an image.

The server accepts one RGB PNG/JPEG upload, size, seed, num_inference_steps, guidance_scale, cfg_scale and negative_prompt. It rejects masks, multiple references, remote reference URLs, output batching, streaming and unsupported options. For matched-noise checks, initial_latents_b64 uses the owning family's existing raw float32 layout; increase --max-body-bytes when the replay upload exceeds the default limit.

For automated input/output hash checks, request/metric joins and saved edited images, use image-edit-profile. It reuses the image-generation qualifier and records build geometry separately from actual input and requested output geometry. Owning-family correctness on the served bundle remains a separate prerequisite.