Open-Source Image Text Rendering for Chart Footnotes

Which open-source model is a candidate for chart footnotes?

MOSS-Image leads the open-weight models evaluated on LongLongTextBench with 96.3% mean text coverage. The published longest-group results are 97.1% for English and 92.4% for Chinese. For an infographic carrying source notes and qualifiers, evaluate MOSS-Image-SFT as a visual drafting tool and validate all text before release. Official model results · Technical report

What the public evaluation tells an infographic editor

The official page illustrates multiline text inside a scene and reports design comparisons that include infographics. Chart footnotes add a separate problem: a missing qualifier can change how readers interpret otherwise accurate values. We would judge small print after the image is reduced to its delivery size, when cramped line spacing and missing symbols are easier to miss. Published examples and evaluation

For a proposed newsroom exercise, use an approved chart with a short source line and a longer bilingual methodology note. Freeze the underlying data and labels before generation. Compare a generated note against an editable text overlay. Have the reviewer read each qualifier aloud and match its marker to the chart. LLTB coverage describes text presence in the benchmark. Exact copy, marker binding, and legibility are additional checks.

Use the documented layout inputs to reserve note space, then keep a separate accessible text description with the final graphic. Project repository · SFT card

See technical diagram labels, agenda text, and general text rendering.

English SEO keywords

open-source image text rendering, chart footnotes, LongLongTextBench.