xiaokuixk/PortraitCraft-Track1-sky

PortraitCraft Track-1 — Qwen3.5-4B-VL Fine-tuned

A 4B vision-language model fine-tuned for portrait composition aesthetic assessment, submitted to the CVPR 2026 Workshop PortraitCraft Challenge Track-1.

The model jointly predicts:

• 13 fine-grained aesthetic criteria (each as a continuous 0-10 score)
• An overall aesthetic score (integer 1-100)
• A 4-way multiple-choice VQA answer

All three are emitted in a single strict-JSON output.

Quick start

pip install -r inference/requirements.txt

# Run inference on the official Track-1 test set:
bash run_inference.sh \
    /path/to/track_1_test.json \
    /path/to/PortraitCraft   # directory containing images_00/ images_01/ ...

This produces submission.json and submission.zip in the repo root.

Inference

The pipeline at inference time uses 2-pass test-time augmentation:

1. Standard resolution (max_pixels=1003520) + original image
2. Standard resolution + horizontally flipped
3. High resolution (max_pixels=2007040) + original
4. High resolution + flipped

The continuous criterion scores from all four passes are averaged, and only then mapped to discrete levels by fixed thresholds (<5→A, 5-7→B, ≥7→C). The VQA answer is taken from the standard-resolution original pass.

Output schema

{
  "image_path": "...",
  "criteria": {
    "Color Harmony": {"level": "A|B|C"},
    "Visual Style Consistency": {"level": "A|B|C"},
    "Sharpness": {"level": "A|B|C"},
    "Light and Shadow Modeling": {"level": "A|B|C"},
    "Creativity and Originality": {"level": "A|B|C"},
    "Exposure Control": {"level": "A|B|C"},
    "Application of Classical Composition Principles": {"level": "A|B|C"},
    "Depth of Field and Layering": {"level": "A|B|C"},
    "Visual Center Stability": {"level": "A|B|C"},
    "Visual Flow Guidance": {"level": "A|B|C"},
    "Structural Support Stability": {"level": "A|B|C"},
    "Appropriateness of Negative Space": {"level": "A|B|C"},
    "Subject Integrity": {"level": "A|B|C"}
  },
  "total_score": 65,
  "question": "...",
  "options": {"A": "...", "B": "...", "C": "...", "D": "..."},
  "answer": "A|B|C|D"
}

Environment

Pinned versions for reproducibility (see inference/requirements.txt):

Package	Version
vllm	0.19.1
transformers	5.5.4
torch	2.10.0 (CUDA 12.x)
Pillow	11.3.0

For best reproduction we recommend running on NVIDIA H20 GPUs (matching the training/inference setup).

License

Apache 2.0 (inherited from the Qwen3.5-4B-VL base model).

Citation

If you use this model, please cite the PortraitCraft challenge and the Qwen3.5 base model.