q-future/Q-ReAlign-Pro-9B
Q-ReAlign-Pro-9B by q-future is a 9 billion parameter multimodal quality judge built on a Qwen3.5-VL backbone, designed to score the perceptual quality and aesthetic appeal of images and videos. This model utilizes a hybrid linear/full attention text tower and a SigLIP-style vision encoder, achieving an average SRCC of 0.896 on seven QA benchmarks. It excels at no-reference image/video quality assessment, aesthetic scoring, and dataset curation, providing a scalar output between 0 and 1.
Loading preview...
Overview
q-future/Q-ReAlign-Pro-9B is the largest and highest-fidelity variant in the Q-ReAlign family, a suite of lightweight, human-aligned multimodal quality judges. Built upon a Qwen3.5-VL backbone, this 9 billion parameter model is designed to assess the perceptual quality and aesthetic appeal of images and videos. It processes inputs to rate quality on a scale from "excellent" to "bad", collapsing these into a single scalar score between 0 and 1.
Key Capabilities
- Multimodal Quality Assessment: Scores both image quality (IQA), image aesthetics (IAA), and video quality (VQA) within a unified ONE-ALIGN framework.
- High Performance: Achieves an average SRCC of 0.896 across seven QA benchmarks (KonIQ, SPAQ, KADID, AGI, LIVE, AVA, LSVQ), outperforming the original Q-Align's 0.869.
- Efficient Architecture: Utilizes a Qwen3.5-VL backbone with a hybrid linear/full attention text tower and a SigLIP-style vision encoder, trained with full-parameter SFT in bf16.
- Flexible Context: Supports a context length of up to 262144 tokens.
Good For
- No-reference image and video quality assessment.
- Aesthetic scoring of visual media.
- Dataset curation, ranking, and filtering of generated media.
- Generating reward signals for generative AI pipelines.