Evidence

MOSS-Image benchmark: text and layout tests

We generated 22 fixed-seed images with MOSS-Image SFT, scored every required text field, and published the prompts, raw outputs and failures. This is a reproducibility snapshot, not a cross-model leaderboard.

Updated 2026-10-23Open weights · Apache-2.0Hosted: MOSS-Image SFT

Cases passed

20 / 22

15/15 core · 5/7 boundary

Required text fields

283 / 294

96.3% readable and correct

Price pairs

170 / 170

Item or service next to the right price

Core cases

15 / 15

Bilingual, menu and layout tests

01 / Method

What we ran and how we scored it

The hosted model is MOSS-Image SFT through fal.ai. Every case uses the same server-side stack as the product: 50 inference steps, CFG 3, the default negative prompt, the safety checker on, and prompt expansion on for text prompts. Each case has a fixed seed. Most cases render at 1K (about 1 MP); the Draft boundary tests render at 512 px. The exact prompt, seed, tier and size are on every card.

Scoring is manual and deliberately conservative. A text field counts only when it is readable and correct after normalizing case, whitespace and punctuation variants such as typographic apostrophes and separator dots. A price pair counts only when the price sits with the correct dish or service. Layout cases are checked against the requested zones with a 5% canvas tolerance.

VerdictRule
PassEvery required text field is readable and correct; every price is paired correctly; layout is within tolerance.
PartialAt least 80% of text fields are correct, no price is paired with the wrong item, and layout is mostly correct.
FailFewer than 80% of text fields are correct, a price is paired with the wrong item, or core copy is unreadable.

Single draw per case

Each case was rendered once, not rerun until it looked good. The page keeps the failures. The hosted UI uses a random seed, so a replay from the UI may differ from the fixed-seed result.

02 / Results

Results by category

Core cases test the product promises. Boundary tests explore where the model starts to break.

CategoryPassedText fieldsPrice pairs
Bilingual long text5 / 534 / 34—
Menus and price lists5 / 596 / 9676 / 76
Layout and packaging5 / 539 / 3916 / 16
Boundary stress tests5 / 7114 / 12578 / 78

03 / Bilingual long text

Bilingual long text

5 fixed-seed cases. Every prompt and raw output is published.

Bilingual event poster — MOSS-Image benchmark outputpassA1

Bilingual event poster

3:4 · 864×1152 · seed 1101

Text 7/7

All seven bilingual fields rendered; line breaks preserved.

Full promptopen
A flat, front-facing event poster on deep charcoal paper, printed in cream and vermilion ink. At the top, the title "NIGHT SIGNALS" in large condensed sans-serif, with "夜间信号" directly below in smaller Chinese type. Under it, the subtitle "A Festival of Light and Sound" and "光与声音的节日". At the bottom, three lines: "SAT 14 NOV 2026 · 20:00", "WAREHOUSE 9 · RIVERSIDE", "Doors open at 19:30 / 19:30 入场". Keep every line separate. No extra words. High contrast, straight-on, no perspective, no glow, no texture.
Bilingual product ad — MOSS-Image benchmark outputpassA2

Bilingual product ad

3:4 · 864×1152 · seed 1102

Text 7/7

Headline, Chinese subtitle, price badge, feature lines and CTA all correct.

Full promptopen
A studio product advertisement for a matte stoneware mug on a warm oatmeal background. At the top left, the headline "SLOW MORNINGS" with "慢晨" below it. Centered, the mug. At the bottom right, a circular price badge reading "$24". Under the mug, three short feature lines: "Hand-thrown stoneware / 手工陶器", "Matte glaze / 哑光釉", "Dishwasher safe / 可机洗". At the very bottom, "SHOP NOW / 立即购买". Keep every line separate. No extra words, no logos, no watermark.
Chinese long-form poster — MOSS-Image benchmark outputpassA3

Chinese long-form poster

3:4 · 864×1152 · seed 1103

Text 7/7

Chinese title, subtitle, dates and the three-line paragraph are all legible.

Full promptopen
A vertical Chinese exhibition poster on warm rice paper with a single ink-brush mountain shape. At the top, the title "山中来信" in large Song-style type. Below it, the subtitle "一场关于茶与雾的展览". At the bottom, a block of three lines: "2026年11月8日—12月20日", "杭州·天目里美术馆 3号展厅", "山雾从竹林里升起。
每一片茶叶都记得雨水的方向。
在缓慢的时间里,重新学习观看。" Keep every line separate and legible. Minimal, straight-on, no perspective, no extra words.
English editorial cover — MOSS-Image benchmark outputpassA4

English editorial cover

3:4 · 864×1152 · seed 1104

Text 7/7

Masthead, issue line, three headlines and both credits correct; headlines wrap across lines.

Full promptopen
A front-facing editorial magazine cover on off-white paper. At the top, the masthead "THE MOSS REVIEW" in a high-contrast serif. Below it, "ISSUE 04 · AUTUMN 2026". In the center, a square photograph of a foggy forest. Around it, three headlines: "The Quiet Return of Print", "A Studio in the Fog", "Notes on Slow Type". At the bottom, two caption lines: "Words by Ada Lin", "Photographs by Milo Chen". Keep every line separate. No extra words, no barcode, no price.
Bilingual exhibition card — MOSS-Image benchmark outputpassA5

Bilingual exhibition card

3:4 · 864×1152 · seed 1105

Text 6/6

All six bilingual lines correct; no dropped characters.

Full promptopen
A minimal bilingual exhibition information card on cream paper with a thin vermilion rule. At the top, "MOSS ARCHIVE / 苔藓档案" in large type. In the middle, four separate lines: "10.16—11.30", "TUE–SUN 10:00–18:00", "GETMOSS GALLERY · 2F", "FREE ENTRY / 免费入场". At the bottom, "Closed Mondays / 周一闭馆". Straight-on, flat, high contrast, no extra words.

05 / Layout and packaging

Layout and packaging

5 fixed-seed cases. Every prompt and raw output is published.

Four-zone poster layout — MOSS-Image benchmark outputpassC1

Four-zone poster layout

3:4 · 864×1152 · seed 3101

Text 4/4 · Layout 4/4

All four text blocks landed in the requested corners around the image.

Requested layout: Title top-center; dates top-right; venue bottom-left; ticket bottom-right; image centered.

Full promptopen
A vertical poster with a strict four-zone layout on cream paper with black and vermilion ink. Top center: the title "THE MOSS ARCHIVE" in large serif. Top right: "10.16—11.30" in small type. Bottom left: "GETMOSS GALLERY · 2F". Bottom right: "FREE ENTRY". Center: a single dark green moss photograph in a thin square frame. Keep the four text blocks in their named corners; do not move them. Flat, straight-on, high contrast, no extra words.
Product ad with price badge — MOSS-Image benchmark outputpassC2

Product ad with price badge

1:1 · 1024×1024 · seed 3102

Text 4/4 · Layout 4/4

Brand, bag label, price badge and CTA landed in the requested zones.

Requested layout: Brand top-left; label on the centered bag; price badge bottom-right; CTA bottom-center.

Full promptopen
A square product advertisement on a pale sand background. Top left: the headline "MOSS & EMBER". Center: a matte black coffee bag with a small cream label reading "SINGLE ORIGIN". Bottom right: a circular price badge reading "$18". Bottom center: "SHOP NOW". Keep each text block in its named position; do not add extra words. Flat, studio lighting, straight-on.
Coffee packaging label — MOSS-Image benchmark outputpassC3

Coffee packaging label

3:4 · 864×1152 · seed 3103

Text 6/6 · Layout 6/6

All six label fields correct; the blank barcode box sits at bottom center.

Requested layout: Brand top; product name center; origin and roast below; net weight bottom-left; ingredients bottom-right; barcode bottom-center.

Full promptopen
A straight-on packaging label for a coffee bag, cream paper with black ink and a thin vermilion rule. Top: "MOSS & EMBER" in a small serif. Center: "SINGLE ORIGIN COFFEE" in large condensed type. Below: "ETHIOPIA · GUJI" and "MEDIUM ROAST". Bottom left: "NET WT. 250 g". Bottom right: "100% ARABICA COFFEE". A blank barcode box sits at the very bottom center. Flat, straight-on, no perspective, no extra words.
Two-column menu zones — MOSS-Image benchmark outputpassC4

Two-column menu zones

4:3 · 1152×864 · seed 3104

Text 20/20 · Prices 16/16 · Layout 20/20

16/16 items and prices correct; header and footer landed in the requested zones.

Requested layout: Header top-center; two column headings; eight items per column; footer bottom-center.

Full promptopen
A straight-on two-column menu board on cream paper. Header at the top center: "MOSS KITCHEN". Left column heading "ALL DAY": "Toast 6", "Granola 8", "Eggs 10", "Soup 7", "Salad 9", "Sandwich 11", "Pasta 14", "Bowl 13". Right column heading "DRINKS": "Coffee 3", "Tea 3", "Juice 5", "Lemonade 4", "Beer 6", "Wine 8", "Sparkling Water 3", "Kombucha 5". Footer at the bottom center: "12 MOSS STREET · OPEN 7–4". Keep the header, column headings, items, prices and footer in their named zones. No extra words.
Shelf price label — MOSS-Image benchmark outputpassC5

Shelf price label

1:1 · 1024×1024 · seed 3105

Text 5/5 · Layout 5/5

All five shelf-label fields correct in their zones.

Requested layout: Name top-center; size center-left; price center-right; origin bottom-left; descriptor bottom-right.

Full promptopen
A square shelf label for a bottle of olive oil, cream paper with dark green ink. Top center: "MOSS OLIVE OIL". Center left: "500 ml". Center right: "$18.00". Bottom left: "Kalamata, Greece". Bottom right: "Extra virgin · cold pressed". Keep each text block in its named position. Flat, straight-on, no extra words.

06 / Failures

Where it fails

These are the cases that did not reach 80% text accuracy. We publish them unchanged.

Long body copy is the weak point

Both failures are dense paragraphs at small type. At 512 px the middle sentences collapse into overlapping glyphs. At 1024 px, with the same prompt and seed, almost none of the paragraph is readable. Higher resolution alone is not a fix: split long copy into short blocks or separate text items.
Long paragraph at Draft 512 — MOSS-Image benchmark outputfailS6

Long paragraph at Draft 512

1:1 · 512×512 · seed 4106

Text 6/9

Long paragraph at Draft 512: the first sentence, the closing section and FREE ADMISSION survive, but the middle three sentences collapse into overlapping glyphs. 6 of 9 required fields are readable.

Full promptopen
A vertical poster at small size with a long artist statement in small type. Title "THE SLOW FIELD". Body text exactly: "The Slow Field is a study of moss, stone and weather across four seasons. The project began in a narrow valley where the fog stays until noon. Each photograph records a surface that is usually overlooked: a crack in a wall, a stain on a step, a line of water at the edge of a path. The work asks what it means to look slowly, without a destination, and to let a place change the pace of attention. The exhibition includes new prints, a reading table and a small archive of field notes. It is not a survey of landscape photography. It is an attempt to stay with one subject long enough for the ordinary to become strange. The artist would like to thank the studio, the printers, and everyone who walked the same path." At the bottom, "FREE ADMISSION". Keep the paragraph as one block of small text. No extra words.
Long paragraph at Final 1024 — MOSS-Image benchmark outputfailS7

Long paragraph at Final 1024

1:1 · 1024×1024 · seed 4106

Text 1/9

The same prompt and seed at Final 1024: the paragraph is almost entirely garbled and only FREE ADMISSION is correct. One fixed-seed sample, not a resolution benchmark; higher resolution alone did not fix long body copy.

Full promptopen
A vertical poster at small size with a long artist statement in small type. Title "THE SLOW FIELD". Body text exactly: "The Slow Field is a study of moss, stone and weather across four seasons. The project began in a narrow valley where the fog stays until noon. Each photograph records a surface that is usually overlooked: a crack in a wall, a stain on a step, a line of water at the edge of a path. The work asks what it means to look slowly, without a destination, and to let a place change the pace of attention. The exhibition includes new prints, a reading table and a small archive of field notes. It is not a survey of landscape photography. It is an attempt to stay with one subject long enough for the ordinary to become strange. The artist would like to thank the studio, the printers, and everyone who walked the same path." At the bottom, "FREE ADMISSION". Keep the paragraph as one block of small text. No extra words.

07 / Boundary

Boundary tests that passed

Low contrast, a 30-item menu at 512 px, a hand-written chalkboard and a 10-line poster all stayed readable.

Low-contrast 30-item menu — MOSS-Image benchmark outputpassS1

Low-contrast 30-item menu

4:3 · 1152×864 · seed 4101

Text 35/35 · Prices 30/30

30/30 items and prices correct despite low-contrast beige ink.

Full promptopen
A straight-on 30-item cafe menu on cream paper with low-contrast beige ink. Header "MOSS CAFE". Three columns, ten items each, small type, prices right-aligned. Column 1 heading "COFFEE": Espresso 3.50, Americano 4.00, Flat White 4.50, Cappuccino 4.50, Latte 4.75, Mocha 5.25, Cold Brew 5.00, Pour Over 5.50, Cortado 4.25, Macchiato 4.00. Column 2 heading "TEA": English Breakfast 3.50, Earl Grey 3.50, Green Tea 3.50, Jasmine 3.50, Oolong 4.00, Mint 3.50, Chamomile 3.50, Rooibos 3.50, Chai 4.50, Matcha 5.00. Column 3 heading "PASTRIES": Croissant 3.75, Almond Croissant 4.50, Pain au Chocolat 4.50, Banana Bread 4.00, Scone 3.50, Cinnamon Roll 4.75, Seasonal Tart 5.50, Cookie 3.25, Brownie 4.00, Muffin 4.25. Footer: "12 MOSS STREET · OPEN 7–4". Keep each item and price on the same line. No extra words.
Low-contrast bilingual paragraph — MOSS-Image benchmark outputpassS2

Low-contrast bilingual paragraph

3:4 · 864×1152 · seed 4102

Text 6/6

Six long bilingual lines correct at low contrast; no dropped characters.

Full promptopen
A vertical bilingual poster on light gray paper with gray ink and low contrast. At the top, "THE SLOW FIELD / 慢野" in large type. Below, four lines of small text: "A study of moss, stone and weather in four seasons. / 苔藓、石头与四季天气的研究。", "Exhibition dates: 2026.11.08—2026.12.20 / 展览时间:2026年11月8日—12月20日", "Opening reception: Saturday 7 November, 18:00 / 开幕酒会:11月7日周六 18:00", "Gallery 3, 2F, 88 Moss Road, Hangzhou / 杭州苔藓路88号 2层3号展厅". At the bottom, "FREE ADMISSION / 免费参观". Keep every line separate and legible. Minimal, no images, no extra words.
Hand-written chalkboard menu — MOSS-Image benchmark outputpassS3

Hand-written chalkboard menu

4:3 · 1152×864 · seed 4103

Text 20/20 · Prices 18/18

18/18 items and prices correct on a hand-written, slightly angled chalkboard.

Full promptopen
A hand-written chalkboard cafe menu, white chalk on a dark green board, slightly angled with soft glare. Header "MOSS CAFE". Two columns, nine items each, with prices: "Espresso 3.50", "Americano 4.00", "Flat White 4.50", "Cappuccino 4.50", "Latte 4.75", "Mocha 5.25", "Cold Brew 5.00", "Pour Over 5.50", "Cortado 4.25", "Croissant 3.75", "Almond Croissant 4.50", "Pain au Chocolat 4.50", "Banana Bread 4.00", "Scone 3.50", "Cinnamon Roll 4.75", "Seasonal Tart 5.50", "Cookie 3.25", "Brownie 4.00". Footer: "OPEN 7–4". Keep each item and price on the same line. No extra words.
30-item menu at Draft 512 — MOSS-Image benchmark outputpassS4

30-item menu at Draft 512

1:1 · 512×512 · seed 4104

Text 35/35 · Prices 30/30

30/30 items and prices correct at 512 px; small but legible.

Full promptopen
A straight-on 30-item cafe menu on cream paper with dark ink. Header "MOSS CAFE". Three columns, ten items each, small type, prices right-aligned. Column 1 heading "COFFEE": Espresso 3.50, Americano 4.00, Flat White 4.50, Cappuccino 4.50, Latte 4.75, Mocha 5.25, Cold Brew 5.00, Pour Over 5.50, Cortado 4.25, Macchiato 4.00. Column 2 heading "TEA": English Breakfast 3.50, Earl Grey 3.50, Green Tea 3.50, Jasmine 3.50, Oolong 4.00, Mint 3.50, Chamomile 3.50, Rooibos 3.50, Chai 4.50, Matcha 5.00. Column 3 heading "PASTRIES": Croissant 3.75, Almond Croissant 4.50, Pain au Chocolat 4.50, Banana Bread 4.00, Scone 3.50, Cinnamon Roll 4.75, Seasonal Tart 5.50, Cookie 3.25, Brownie 4.00, Muffin 4.25. Footer: "12 MOSS STREET · OPEN 7–4". Keep each item and price on the same line. No extra words.
Dense event poster at Draft 512 — MOSS-Image benchmark outputpassS5

Dense event poster at Draft 512

1:1 · 512×512 · seed 4105

Text 11/11

11/11 lines correct at 512 px on a textured wall.

Full promptopen
A dense bilingual event poster at small size, photographed flat on a textured concrete wall at low contrast. Headline "NIGHT SIGNALS / 夜间信号". Ten lines of small text: "SAT 14 NOV 2026 · 20:00", "WAREHOUSE 9 · RIVERSIDE", "Doors 19:30 / 19:30 入场", "Tickets $18 / 票价 18 美元", "DJ set by Ada Lin", "Live visuals by Milo Chen", "Bar open until late", "No outside drinks", "Accessible entrance on 2F", "Follow @mossignals". Keep every line separate and legible. No extra words.

08 / Limits

What this proves — and what it does not

What this supports

  • On clean, flat, high-contrast layouts, MOSS-Image SFT rendered all required fields in 15 of 15 core cases.
  • All 170 dish and service prices stayed with the correct item.
  • Explicit placement instructions worked in all five layout and packaging cases.
  • A 30-item menu stayed correct at both 1152×864 and 512×512.

What this does not prove

  • It is not a comparison with other models. No other model was run for this snapshot.
  • It is one draw per case. It does not estimate variance or confidence intervals.
  • It does not cover photos of real menus, occlusion, extreme perspective, decorative or handwritten type, or scripts beyond English and Chinese.
  • It does not promise that every character will render correctly in your prompt. Long paragraphs still fail.

FAQ

Benchmark questions

Can I reproduce these results?

Yes. Every card includes the full prompt, seed, render tier and output size. The hosted UI uses a random seed, so a replay may differ; the API can pass the same seed. The other settings match the product: MOSS-Image SFT, 50 steps, CFG 3, the default negative prompt, safety checker on, prompt expansion on for text prompts.

Why do some cases fail?

The two failures are long paragraphs. At Draft 512 the middle sentences collapse into overlapping glyphs; the same prompt and seed at 1K 1024 produced almost no readable body text. Split long copy into short blocks or separate text items instead of assuming a higher resolution will fix it.

Do you compare other image models?

Not yet. This is a single-model reproducibility snapshot, not a cross-model leaderboard. We will add another model only when we can run the same prompts with fixed settings and publish the raw outputs.

How do you score text and prices?

Text is matched case-insensitively after normalizing whitespace and punctuation variants such as typographic apostrophes and separator dots. A field counts only if it is readable and correct. A price pair counts only when the price sits with the correct dish or service. Layout cases are checked against the requested zones with a 5% canvas tolerance.

Why 22 cases?

Version 1 has 15 core cases and 7 boundary tests. We are expanding to 30 and adding three-seed repeats for the core cases. Cases are added to this page; old results are not deleted to make the pass rate look better.

Are the published images edited?

No. The PNG is the provider output. The WebP version is a format conversion only: no cropping, retouching or text replacement. Every card links to the original PNG.

More

Use the tests