photo_judgers/docs/research-q-align.md

Q-Align — Qwen-based Visual Scorer

Model Info

• GitHub: https://github.com/Q-Future/Q-Align
• Paper: ICML 2024 — "Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels"
• HuggingFace: Multiple checkpoints available
• License: Apache 2.0
• Latest: Q-ReAlign (Qwen3.5-VL backbone) — June 2026 update in IQA-PyTorch

Architecture

• Type: Vision-Language Model (VLM) fine-tuned for visual scoring
• Backbone: Qwen LLM + vision encoder
• Output: Discrete score levels (text-defined)
• Innovation: Uses reinforcement learning to align model scoring with human perception

Hardware Requirements

• Mini (0.8B): ~3 GB VRAM
• Lite (4B): ~10 GB VRAM
• Pro (9B): ~20 GB VRAM
• Batch size: 1-4 (VLMs are memory-intensive)

What It's Good At

• Human-aligned scoring: Trained to match human subjective ratings
• Explainable: Can generate text descriptions of why an image scored high/low
• Multi-task: Can do IQA, IAA (Image Aesthetic Assessment), and VQA
• Fine-tunable: Can be adapted to specific domains (e.g., wedding photos)

What It's Bad At

• Slower inference: VLM architecture is significantly slower than CLIP-based models
• Higher VRAM: Even the mini variant needs 3 GB; not ideal for batch processing thousands of images
• Still no wedding-specific training: General-purpose model
• Complex setup: Requires more dependencies and configuration than LAION predictors