WeGenEditBench
An atomic checklist-driven benchmark for image editing models, covering single-image editing, text-dense editing, and multi-reference image editing scenarios, using pre-generated checklists to decompose requests and an evaluation process that deducts points first and aggregates afterward.
Timeline (last 90 days)
WeGenEditBench contains 2,740 evaluation instances and 148 task types, covering three major scenarios: single-image editing, text-intensive editing, and multi-reference-image editing.
WeGenEditBench uses a pre-generated checklist shared across models, decomposing each editing request into atomic-level requirements, to ensure different models are examined under the same standard and to reduce noise from evaluation-standard drift.
WeGenEditBench's evaluation pipeline first establishes item-level deductions and visual evidence, then aggregates them into scores across four dimensions: editing execution, content preservation, visual quality, and task-specific correctness.
Among the 15 evaluated configurations, GPT-Image-2 leads single-image editing with an Overall score of 4.076, and Seedream-5.0-Pro ranks first in text-intensive editing with 4.291.
The combination of WeGenEditBench and WeGenEvoStudio provides an interpretable evaluation path, where every score is backed by atomic requirements and visual evidence, and every failure point has a corresponding data construction and adaptation mechanism.
The framework's coverage is still limited to 148 task types, and the pre-generation method of the checklist may introduce the designers' prior preferences.
All Facts (6)
WeGenEditBench contains 2,740 evaluation instances and 148 task types, covering three major scenarios: single-image editing, text-intensive editing, and multi-reference-image editing.
50%UnverifiedWeGenEditBench uses a pre-generated checklist shared across models, decomposing each editing request into atomic-level requirements, to ensure different models are examined under the same standard and to reduce noise from evaluation-standard drift.
50%UnverifiedWeGenEditBench's evaluation pipeline first establishes item-level deductions and visual evidence, then aggregates them into scores across four dimensions: editing execution, content preservation, visual quality, and task-specific correctness.
50%UnverifiedAmong the 15 evaluated configurations, GPT-Image-2 leads single-image editing with an Overall score of 4.076, and Seedream-5.0-Pro ranks first in text-intensive editing with 4.291.
50%UnverifiedThe combination of WeGenEditBench and WeGenEvoStudio provides an interpretable evaluation path, where every score is backed by atomic requirements and visual evidence, and every failure point has a corresponding data construction and adaptation mechanism.
50%UnverifiedThe framework's coverage is still limited to 148 task types, and the pre-generation method of the checklist may introduce the designers' prior preferences.
50%