February 12 — Xiaohongshu (RED) has open-sourced its new picture enhancing basis mannequin, FireRed-Picture-Edit, releasing code, a technical report, and demo pages on GitHub and Hugging Face. Mannequin weights are anticipated to comply with within the coming days.
The mannequin has achieved state-of-the-art (SOTA) outcomes on a number of main picture enhancing benchmarks, together with ImgEdit and GEdit.
The workforce additionally launched RedEdit Bench, a proprietary analysis framework protecting 15 sub-tasks akin to object insertion/elimination, portrait enhancement, and low-quality picture restoration. The benchmark may also be open-sourced.
Technically, FireRed-Picture-Edit adopts a three-stage coaching technique:
- Pre-training: Multi-condition perceptual bucket sampling and dynamic instruction augmentation to enhance generalization.
- Positive-tuning: Excessive-quality curated knowledge to refine enhancing efficiency.
- Reinforcement studying: A novel Format-Conscious OCR-based Reward mechanism that penalizes typos, misaligned characters, irregular font scaling, and structure distortions—considerably enhance text-editing accuracy and stylistic consistency.
Core capabilities embrace sturdy instruction following, exact textual content enhancing, fashion switch, multi-reference picture fusion, outdated picture restoration, and high-fidelity picture enhancement.
Xiaohongshu stated future updates will additional strengthen portrait retouching, textual content enhancing precision, and consistency preservation, with further open-source releases—together with text-to-image basis fashions—deliberate within the coming months.
Supply: QbitAI
Elevate your perspective with NextTech Information, the place innovation meets perception.
Uncover the most recent breakthroughs, get unique updates, and join with a world community of future-focused thinkers.
Unlock tomorrow’s traits as we speak: learn extra, subscribe to our e-newsletter, and turn into a part of the NextTech group at NextTech-news.com
