Pasted Arial beats every image model at translating manga sound effects

What the paper is, and what it claims

Takara Taniguchi and Hideki Nakayama of the University of Tokyo, with Wataru Shimoda and Kota Yamaguchi of CyberAgent, posted arXiv:2609.21199 on 18 September 2026, a SIGGRAPH Posters '26 paper. The task is visual onomatopoeia translation: take a manga panel where a Japanese sound effect is drawn into the art, remove it cleanly, and draw an English one in the same visual style.

This is a real production problem. A publisher redrawing sound effects pays for lettering and touch-up on every panel; a publisher leaving them in Japanese with a small gloss pays nothing and loses the effect. OnomatoBridge is a pipeline for doing the first option automatically, evaluated against three image-editing models on the Manga109 onomatopoeia dataset.

Tokyo Side of Life (@TokyoSideofLife), Are Mangas Properly Translated? Bilingual Looks at Onomatopoeias in Mangas, 25 February 2025. A bilingual reader walking through published English editions and showing which sound effects were redrawn, which were left in Japanese with a gloss, and which were dropped. It supports the claim that the three options in this article are the three options publishers already choose between, and that readers notice.

Table 1, with the control left in

NED is normalised edit distance between the rendered English text and the intended English text, higher is better. JPDet is the fraction of outputs where a vision-language model still detects Japanese text, lower is better. Style, Residual and Readability are 1-to-5 means from a five-person study. Source: arXiv:2609.21199, Table 1. Bold marks the best value in each of the first two columns.

The naive baseline is described in one sentence of the paper: it "simply overlays the Arial English text onto the original image containing Japanese onomatopoeia". No inpainting, no style matching, no generation. It scores 0.8918 NED. OnomatoBridge scores 0.7826 and Gemini 3 Pro Image 0.7182. Every generative system in the table renders the intended English text less accurately than pasting it in a system font.

That result is in the paper and the paper says so, in a subordinate clause: "our method outperforms competing models except for the naive baseline in both English onomatopoeia readability and removal of the original Japanese onomatopoeia". The abstract does not mention it. The control appears in Table 1 with three empty cells and is not discussed again.

What the trade actually is

Our arithmetic on Table 1 of arXiv:2609.21199. The NED column is the control's 0.8918 minus each system's score, in points. The residual column is the relative reduction in JPDet rate against the control's 0.5108.

Read as a trade, OnomatoBridge gives up 10.9 NED points to remove 56.0% of the residual Japanese. That is a defensible exchange and the pipeline's real contribution. Two other numbers in that column are less comfortable.

Qwen-Image Edit removes 11.0% of the residual Japanese. Its JPDet rate is 0.4547 against the control's 0.5108. A model whose entire job in this comparison is to erase the Japanese and draw English gets within a tenth of the score you get by pasting text over the top and walking away, and it loses 32.5 NED points doing it.

The control's residual score is 0.5108, not 1.0. It does not remove anything, so in principle a detector should find Japanese in every one of its outputs. It finds it in 51%. Either the pasted Arial block physically covers the glyph about half the time, or the vision-language model doing the detection misses it about half the time. The paper does not report detector accuracy, so the JPDet column has an unmeasured error floor sitting under every row of it.

The abstract says the method "improves English text correctness by roughly 10 to 25 points". Against the three image-editing baselines the actual gains are 24.4 points over FLUX.1 Kontext, 21.5 over Qwen-Image Edit, and 6.4 over Gemini 3 Pro Image. Two of three sit in the stated range.

The evaluation set was chosen by the method being evaluated

The experimental setup is three sentences long and one of them decides the result. 384 Japanese onomatopoeia from Manga109 with bounding boxes of at least 300 x 300 pixels go in. 92 images are evaluated, "after filtering by OnomatoBridge". That is 24.0% retention, and the three baselines are scored on the same 92.

This is the part of the design a reader should push on. OnomatoBridge's filters are pixel-change scores that reject an inpaint when too little changed inside the mask or too much changed outside it, and reject an extraction when the glyph is not confined to the mask polygon or its SSIM to the original is too low. Those are reasonable rejection rules. They are also the proposed method's own quality judgement, applied before the comparison. The 292 dropped cases are the ones OnomatoBridge found hard, and nobody reports what the baselines would have scored on them.

The compute side is in the same sentences. Four inpaint candidates, four extract candidates, two FLUX.1 Kontext Pro refiner passes: ten image generations per sound effect, minimum, against one for each baseline. A single manga volume carries hundreds of sound effects.

The human study has five raters, three questions, no reported inter-rater agreement and no intervals. OnomatoBridge wins all three. The system with the best text score was not put in front of them, so we cannot say whether a reader would prefer a correctly spelled Arial block over a stylised near-miss.

What this means for anyone localizing comics

Score the do-nothing option, every time. The whole finding here exists because the authors did include a pasted-text control and printed its number. Most image-editing papers in localization do not. If your vendor demo does not show you what a system font in the right place scores, ask for it.

Text correctness and style fidelity are separate purchases. A stylised sound effect that reads SPLOOSH when it should read SPLASH has failed at the one job translation has. On these numbers the generative route costs you 6 to 33 NED points of text accuracy in exchange for a style match that five people rated between 1.99 and 3.10 out of 5. Neither end of that trade is finished technology.

Retention is a cost line. A pipeline that produces acceptable output on 24% of inputs needs a human for the other 76%, and the ten-generation-per-effect bill applies to the failures too. Price the per-panel cost against a letterer, not against zero.

Judgement, marked as judgement: the honest reading of this table is that automated visual onomatopoeia translation is not yet better than a legible overlay on the axis that matters most to a reader, and that the paper is more useful as a measurement of that gap than as a product. It is a two-page poster and does not claim otherwise.

Check it yourself

Every number above is in Table 1 of the paper. The derived columns are two lines of arithmetic.

python3 - <<'EOF'
# Table 1, arXiv:2609.21199
rows = {
 'Naive baseline':     (0.8918, 0.5108),
 'FLUX.1 Kontext':     (0.5391, 0.2772),
 'Qwen-Image Edit':    (0.5673, 0.4547),
 'Gemini 3 Pro Image': (0.7182, 0.3949),
 'OnomatoBridge':      (0.7826, 0.2246),
}
n_ned, n_jp = rows['Naive baseline']
for k, (ned, jp) in rows.items():
    print(f'{k:<19} NED {ned:.4f}  '
          f'behind control {(n_ned-ned)*100:5.1f} pts  '
          f'residual removed {100*(n_jp-jp)/n_jp:5.1f}%')
EOF
# Naive baseline      NED 0.8918  behind control   0.0 pts  residual removed   0.0%
# Qwen-Image Edit     NED 0.5673  behind control  32.5 pts  residual removed  11.0%
# OnomatoBridge       NED 0.7826  behind control  10.9 pts  residual removed  56.0%

The dataset is Manga109, which is free for academic use and requires a separate agreement for commercial use, plus the Manga109-Onomatopoeia annotation layer. The pipeline components named in the paper are FLUX.1 Kontext and FLUX.1 Kontext Pro (inpainting and the two refiner passes), Gemini 2.5 Flash Image (extraction), and GPT-4o (generating the English onomatopoeia text from the Japanese). The masks come with the dataset, and the paper assumes the English text and the mask are given, so it is not solving translation or detection.

What would prove this wrong

The 292 dropped cases. If the authors score all 384 onomatopoeia for every system, including the ones OnomatoBridge's filters rejected, and OnomatoBridge still leads the generative baselines by roughly 20 NED points, the selection concern is empty and the comparison stands as published. Our prediction: the gap narrows, because the filter rejects exactly the cases the pipeline handles badly.

The control in the user study. If five or more readers are shown the pasted-Arial output alongside OnomatoBridge and rate it below 2.5 on readability, then NED is measuring the wrong thing and a correctly spelled system font really is unacceptable in a panel. That single experiment would settle the article's central claim in either direction, and it costs an afternoon.

The detector. If the vision-language model behind the JPDet column is reported with its own accuracy and turns out to be near-perfect, then the control's 0.5108 means the Arial overlay genuinely hides half the glyphs and our error-floor point is wrong.

Sources

1. Takara Taniguchi, Wataru Shimoda, Kota Yamaguchi, Hideki Nakayama, OnomatoBridge: Onomatopoeia Translation and Rendering Pipeline in Manga, arXiv:2609.21199, 18 September 2026. SIGGRAPH Posters '26, Los Angeles, 19-23 July 2026. University of Tokyo and CyberAgent.
2. Manga109 and the Manga109-Onomatopoeia annotations, manga109.org, the source of the 384 candidate sound effects.
3. Tokyo Side of Life, Are Mangas Properly Translated? Bilingual Looks at Onomatopoeias in Mangas, YouTube, 25 February 2025.

Related BLOMEGA guides: Japanese subtitle scores and a segmentation artefact · Prime Video lip-sync visual dubbing · A copy baseline beats two models on Teochew

FAQ

What is visual onomatopoeia translation in manga?

Replacing a Japanese sound effect that is drawn into the artwork with an English one in the same visual style: removing the original glyph, inpainting the art underneath, and rendering the English text so it matches the drawn style. Publishers otherwise either redraw sound effects by hand, which costs lettering and touch-up time, or leave the Japanese and add a small gloss.

Which model is best at translating manga sound effects?

On the 92 evaluated Manga109 images in arXiv:2609.21199, OnomatoBridge scores 0.7826 NED for text correctness and 0.2246 JPDet for residual Japanese, ahead of Gemini 3 Pro Image (0.7182 / 0.3949), Qwen-Image Edit (0.5673 / 0.4547) and FLUX.1 Kontext (0.5391 / 0.2772). The naive control that pastes Arial over the original scores 0.8918 on text correctness, the best figure in the table, and 0.5108 on residual.

Why does a pasted system font beat image generation here?

Because rendering exact text is the one thing an image editing model is not reliable at. Pasting the intended English guarantees the letters are correct, and the metric that measures correctness, normalised edit distance, rewards that directly. The pasted version fails on the other axis: it does not remove the Japanese underneath, so a detector still finds Japanese in 51% of its outputs.

How many images does OnomatoBridge process successfully?

92 of 384 candidate onomatopoeia, or 24.0%, survive its inpainting and extraction filters and reach the evaluation. The three baselines are scored on that same 92-image set, which the proposed method selected. What the baselines score on the other 292 is not reported.

How expensive is the pipeline per sound effect?

At least ten image generations: four inpainting candidates, four extraction candidates, and two FLUX.1 Kontext Pro refiner passes. Each baseline uses one generation. A manga volume carries hundreds of sound effects, and the ten-generation cost applies to the 76% that get filtered out as well.