EvoQuality¶

Development line: project:evoquality · thread bytedance-evoquality-development
Last event: 2026-06-11 · 1 dated since 2026-06-11 · Researched: 2026-09-05 · confidence: high

What it is¶

EvoQuality is a no-reference image-quality VLM for researchers and image-pipeline teams. - Scoring: scores one image on perceptual quality. - Comparison: compares two images side by side. - Ranking: produces a voting-based ranking signal for filtering or data cleaning.

The released BF16 checkpoint is 8B parameters and represents one self-evolution round. Use it as a perceptual proxy with spot checks, not as an objective or high-stakes decision-maker.

Development line¶

• 2026-06-11 — ByteDance published an EvoQuality resource on Hugging Face. ByteDance posted a public link to EvoQuality on Hugging Face on 2026-06-11. It marks the public release point for the project. The post omits artifact type, version details, capabilities, and announcement notes.

What changed¶

2026-06-11 — EvoQuality weights became available at ByteDance/EvoQuality. The release exposes an 8B BF16 checkpoint trained from Qwen2.5-VL-7B through self-supervised pairwise voting and GRPO.

2025-09-30 — The paper appeared on arXiv. It describes the label-free voting-and-ranking method, reporting a 31.8% PLCC improvement over its base VLM in zero-shot IQA evaluation.

How to use this¶

Use the ByteDance/EvoQuality Hugging Face resource from 2026-06-11 as the dated reference point. Confirm artifact type and usage details before adoption.

1. Load ByteDance/EvoQuality with Transformers as an image-text-to-text model, submit an image plus the model-card scoring prompt, and parse the final boxed numeric score. — https://huggingface.co/ByteDance/EvoQuality
2. For a local ranking batch, run the repository's pairwise-voting inference script with your checkpoint, image directory, pair CSV, and an explicit voting count. — https://github.com/bytedance/EvoQuality
3. For serving, start the model with vLLM or SGLang and send image-plus-text chat-completions requests. — https://huggingface.co/ByteDance/EvoQuality

Best practices¶

• Use it for ranking, filtering, data cleaning, and pre-production assessment. Combine the score with business signals and manual spot checks. — https://huggingface.co/ByteDance/EvoQuality
• For pairwise inference, randomize image order to reduce positional bias. Keep prompts, sampling count, and decoding settings fixed when comparing runs. — https://huggingface.co/ByteDance/EvoQuality
• Do not use the output as the sole criterion for moderation, medical imaging, legal evidence, or other high-stakes decisions. Domain shift and pseudo-label bias remain stated limitations. — https://huggingface.co/ByteDance/EvoQuality

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The Hugging Face card documents the checkpoint and use paths but lacks a versioned release-note history. We found no later official EvoQuality checkpoint in reviewed sources.
• The repository's installation snippet uses a placeholder clone URL (your-organization/evoquality). Use the public repository URL above when cloning.

Sources¶

Agent brief¶

• Subject: project:evoquality, thread bytedance-evoquality-development, 1 dated events 2026-06-11 → 2026-06-11.
• Practical note: From 2026-06-11, practitioners should use the ByteDance/EvoQuality Hugging Face resource as the dated public reference point for evaluating project artifacts, while independently confirming the exact artifact type and usage details before adoption.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.