Image budget: native vs 8 MP vs 6.29 MP vs 4 MP
Measured on 206,020 training pages (200K random + all 6,020 newspapers), weighted by the batch shares (hw-vs 25, hw-dps 8, hw-loc 10, hw-iitm 7; ncert, indicdlp, archives, newspapers, arxiv 10 each). Answer = compact block markers + digit boxes + text, figures without text.
Answer tokens stock → v3: mean 1,406 → 1,052; p99 13,835 → 8,874; max 34,777 → 21,387.
| image cap | tokenizer | mean total | ≤16K | ≤24K | ≤32K | max total |
|---|---|---|---|---|---|---|
| 2MP | stock | 3,287 | 99.172% | 99.977% | 99.995% | 36,804 |
| 2MP | cbpe_v3 | 2,933 | 99.957% | 100.0% | 100.0% | 23,414 |
| 4MP | stock | 4,227 | 98.492% | 99.962% | 99.993% | 38,871 |
| 4MP | cbpe_v3 | 3,873 | 99.908% | 99.997% | 100.0% | 25,481 |
| 6.29MP | stock | 4,826 | 97.801% | 99.9% | 99.988% | 40,888 |
| 6.29MP | cbpe_v3 | 4,472 | 99.694% | 99.993% | 100.0% | 27,498 |
| 8MP | stock | 5,330 | 97.51% | 99.694% | 99.985% | 42,868 |
| 8MP | cbpe_v3 | 4,977 | 99.037% | 99.985% | 100.0% | 29,478 |
| native16MP | stock | 6,053 | 91.493% | 98.832% | 99.972% | 45,755 |
| native16MP | cbpe_v3 | 5,699 | 92.368% | 99.9% | 100.0% | 32,365 |
Total = 64 prompt + image + answer tokens. One image token per 32×32 px after resize.
What the other OCR models use
| model | base | max pixels | where |
|---|---|---|---|
| Chandra OCR 2 | Qwen3.5 4B | 6.29 MP (3072×2048 area); PDFs at 192 DPI; output ≤ 12,384 tok | github/chandra_model_util.py scale_to_fit, chandra_settings.py |
| Surya OCR 2 | Qwen3.5 0.8B arch | 6.29 MP (3072×2048 area) | github/surya_inference_util.py, vLLM max_pixels 6291456 |
| LightOnOCR-3 4B / 0.8B | Qwen3.5 | trained at 2,048 px longest edge (~3 MP for A4); benchmarks allow 6.29 MP | github/lightonocr_models.py, benchmarks_ocr_bench_config.py |
| Logics-Parsing-V3 | Qwen3.5 0.8B | 4.19 MP (2048²) | github/inference_v3.py MAX_PIXELS |
| Infinity-Parser2 | Qwen3.5 | 16.8 MP (4096²) | README min/max_pixels |
| HunyuanOCR | Hunyuan-VL 1B | 4.19 MP processor (our adapter 9.4 MP) | preprocessor_config |
| GLM-OCR | GLM 0.9B | 11.8 MP (SDK) / 9.6 MP (config) | sdk/image_utils.py |
| dots.ocr / dots.mocr | dots 3B | 11.3 MP | preprocessor_config |
| Dolphin v2 | Qwen2.5-VL 3B | long side ≤ 1,600 px (~2.6 MP) | demo_page.py (our adapter) |
| MinerU2.5 / PaddleOCR-VL | two-stage | 1.6 MP / 1.0 MP per crop | preprocessor_config, image_processing_paddleocr_vl.py |
The two strongest unified Qwen3.5 models (Chandra 2, Surya 2, same base as ours) feed at most 6.29 MP. Two-stage systems use small inputs because they OCR crops, not full pages.
Text size per source
Each crop shows the same region of a real training page at the resolution the model would receive. All three versions are drawn at the same display width with image-rendering: pixelated, so a lower cap looks blockier: that blockiness is the information the model loses. One image token covers 32×32 pixels at the model's resolution; one vision patch is 16×16.
Line height = block height / number of text lines (includes line spacing; blocks under 8 px per line are ignored as line-count errors). Most hurt by 4 MP = largest line-height drop that ends below 20 px. Densest = page at the 5th percentile of line height at 4 MP; the crop is a text-rich block near that page's median line height. Table: median and 10th-percentile line height per source, and % of pages whose median line is under 16 px (one patch).
| source | pages | native | 8MP | 6.29MP | 4MP | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| median px | p10 px | <16 px | median px | p10 px | <16 px | median px | p10 px | <16 px | median px | p10 px | <16 px | ||
| arxiv-data | 414 | 51.8 | 37.8 | 5.6% | 51.1 | 37.3 | 5.6% | 44.5 | 32.3 | 5.8% | 36.3 | 26.4 | 6.0% |
| hw-dps | 480 | 81.1 | 54.9 | 0.0% | 81.1 | 54.9 | 0.0% | 80.4 | 54.6 | 0.0% | 77.5 | 48.7 | 0.0% |
| hw-iitm | 479 | 78.4 | 52.6 | 0.8% | 77.2 | 52.6 | 0.8% | 77.1 | 49.7 | 0.8% | 75.9 | 48.5 | 1.0% |
| hw-loc | 502 | 114.5 | 67.9 | 0.6% | 104.1 | 57.7 | 1.2% | 94.2 | 50.3 | 1.6% | 80.7 | 41.4 | 1.8% |
| hw-vs | 462 | 58.0 | 27.0 | 5.4% | 58.0 | 27.0 | 5.4% | 58.0 | 27.0 | 5.4% | 58.0 | 27.0 | 5.4% |
| indicdlp-v2-final | 605 | 35.7 | 15.7 | 10.4% | 35.7 | 15.7 | 10.4% | 35.5 | 15.7 | 10.4% | 35.0 | 15.7 | 10.4% |
| indicdlp-v2-final-np | 593 | 26.5 | 12.0 | 23.6% | 25.8 | 12.0 | 23.6% | 22.6 | 12.0 | 23.8% | 18.9 | 12.0 | 28.2% |
| ncert-scert | 470 | 44.7 | 21.9 | 7.7% | 44.3 | 20.6 | 7.9% | 44.3 | 20.3 | 7.9% | 44.3 | 20.3 | 8.1% |
| printed-nat-archives | 398 | 37.0 | 19.5 | 1.5% | 35.9 | 19.4 | 1.5% | 35.2 | 19.4 | 1.8% | 32.7 | 19.3 | 2.3% |
arxiv-data · biggest page · 4955×3505 (17.4 MP)
2001.06548__2001.06548_page_30




stored text of this block
The image is a correlation plot, specifically a correlogram, displaying the correlation coefficients between various variables. The variables are listed on both the x-axis (top) and y-axis (left). The plot uses a grid of circles to represent the correlations. The size and color of each circle indica
arxiv-data · densest page · 2550×3300 (8.4 MP)
2409.04107__2409.04107_page_7




stored text of this block
<table>
<tr>
<td>
$\boldsymbol{\varepsilon}$
</td>
<td>
$\tilde{\mathbf{B}}$
</td>
<td>
Selected Nodes
</td>
<td>
Original and Reconstructed Signals
</td>
<td>
Normalized
Error (in dB)
</td>
</tr>
<tr>
<td rowspan="2">
$0.03$
</td>
<td>
$\tilde{\mathbf{B}}_{\text{svd}}$
</td>
<td>
<img src="LABELID_arxiv-data · most-hurt-by-4MP page · 2550×3300 (8.4 MP)
hep-ex_0403040__hep-ex_0403040_page_17




stored text of this block
$$B_{\Lambda_c}=(4.7\pm1.1\text{ (stat)}\pm1.3\;(Br(\Lambda_c\rightarrow pK^-\pi^+))\pm1.3\text{ (other syst)})\%.$$
(19)hw-dps · biggest page · 2339×3312 (7.7 MP)
bf9e543afedf0206a10ed47946c726af




stored text of this block
2 There are many political parts in our contry. This Multiparty system is very important part of our democracy. Ex: BJP, Congress
hw-dps · densest page · 2339×3312 (7.7 MP)
34e25d4018d80d2b3980c15dd83f5079




stored text of this block
most (a) tasty (b) healthier (c) best (d) exotic X
hw-iitm · biggest page · 9012×6251 (56.3 MP)
0a560f1d54f43a3715d643e003596679




stored text of this block
from F.B.D. of brake lever AB
$\Sigma M_A \geqslant 0 \quad P \times 2 - T_1 \times 1 = 0$
$T_1 \geqslant 200 N$
F.B.D. of pulley
$T_2 = T_1 e^{\mu \beta}$
$\beta = 270^{\circ} \geqslant \frac{3}{2} \pi \text{ rad}$
$\mu = 0.3$
$T_2 \geqslant 200 \times e^{(0.3 \times 3\pi/2)}$
$T_2 = 822.2 \text{ Khw-iitm · densest page · 1654×2338 (3.9 MP)
6591eb218b9b8dd215db169780937513




stored text of this block
- Cranial - CC & Brain - Skull shape - CSF - Foramen Magnum - BF - Brow ridge - PMC - Forehead - Prognathism - Mandible & chin - Dentition -> Parabolic -> Size -> Diastema (canine) - Post cranial - Linea Aspera lumbar curve - Pelvis broad
hw-iitm · most-hurt-by-4MP page · 5023×5812 (29.2 MP)
fd9a4ea04acd15311a9e9448069d79e3




stored text of this block
<table> <tr> <td>Job</td> <td colspan="2">Machine A</td> <td colspan="2">Machine B</td> <td colspan="2">Machine C</td> </tr> <tr> <td></td> <td>IN</td> <td>OUT</td> <td>IN</td> <td>OUT</td> <td>IN</td> <td>OUT</td> </tr> <tr> <td>4</td> <td>0</td> <td>2</td> <td>2</td> <td>4</td> <td>4</td> <td>10</
hw-loc · biggest page · 6826×9055 (61.8 MP)
ca8fab67e9aee549ac162e53083d3929




stored text of this block
not criminatively of anybody, nor accusatory of those who have in so unusual and ungenerous a manner sought to undermine me in your good opinion. You have not given me the names of those who have so unfairly, un- justly, and causelessly arraigned me; nor furnished the specifica- tions of the guilt &
hw-loc · densest page · 3632×4636 (16.8 MP)
13d05563fe4544028c40c45940c97d79




stored text of this block
I am sorry to see you leave the government service where you have done much valuable work.
hw-loc · most-hurt-by-4MP page · 6150×4245 (26.1 MP)
72c1e00b4f9eccc19be772ee12537875




stored text of this block
New York Catch'd it from them & so it passed to this Government: and if administration had put their Invention upon the utmost Stretch to Contrive a plan of union for the Colonies, I cannot well conceive of any one measure that would tend more Effectually to unite the Colonies than the present act i
hw-vs · biggest page · 8000×10667 (85.3 MP)
404cdd56-839a-412e-aa6c-3697821a5597_page_2




stored text of this block
"তৎকালে পৃথিবী ঈশ্বরের সাক্ষাতে ভ্রষ্ট... আমাদের মুখ্য উদাহরন যিশু খ্রিষ্টকে বিবেচনা করুন। অন্ততপক্ষে, সে তা জানত না, কিন্তু, সা. প্রেরিতরা দৃঢ়ভাবে বিশ্বাস করত যে, মৃতেরা পুনরুত্থিত হবে। ১৯ জন ক্ষুদ্র অপরাধী রেকর্ডের সাথে সম্ভবত চিহ্নিত করা হয়েছিল।
hw-vs · densest page · 1182×1625 (1.9 MP)
236c9ed1-ac0c-467c-88e2-fb3113628098_page_17




stored text of this block
2. The five colours of the five rings are red, green, yellow, blue and black a) pink b) orange c) maroon d) black
indicdlp-v2-final · biggest page · 8420×12200 (102.7 MP)
indicdlp_nv_ml_000700_0




stored text of this block
ഒരു മണിക്കൂർ കൂടി കഴിഞ്ഞശേഷം ഉള്ള ഇപ്ര കാരം പറഞ്ഞു:—ഞാൻ അവിടെ ഇരുന്നു് എഴുതിക്കൊ ണ്ടിരിക്കുമ്പോൾ വാതലരികിൽ നിന്നുകൊണ്ടു് ഒരാൾ “കയ്യുയർത്തുക” എന്ന് ആജ്ഞാപിച്ചു. ഞാൻ ചുററും നോക്കി. ഒരാൾ മുഖം മൂടിക്കൊണ്ടു നിൽക്കുന്നതു കണ്ടു. മുഖം മൂടിയിരുന്നെങ്കിലും അതു മോഹനലാൽ ആണെന്ന് എനിക്ക് മനസ്സിലായി. അയാൾ കൈത്തോക്
indicdlp-v2-final · densest page · 204×248 (0.1 MP)
indicdlp_v2_nt_ur_000023_0




stored text of this block
گریجویٹ کی کارکردگی اور انفرادی شعراء میں مساواتوں سے بے دہی کی کارروائی *گریجوائن* کے مجموعے سے بڑھتا ہے۔ کار، بے دہی، تو ایک معنی بناتا ہے: ہیئت میں کسی بھی قسم کے لیسین کی ذمہ داری پر رجحان یا ثبوتیت لمحہ پر مکانی اور دہی کی حتمی کمی کو پورا کرنا۔ مثال کے لیے: پروازوں یا دیگر پاؤڈروں کے مالکان، ب
indicdlp-v2-final-np · biggest page · 4400×6917 (30.4 MP)
indicdlp_v2_np_ur_000395_1




stored text of this block
کی (ایسی) ریاست مہاراشٹرا کے وزیر اعلی کے ذریعے مراٹھا سماج کو ریزرویشن دینے کے اعلان کا اساتذہ کی متحرک تنظیم اکھل بھارتیہ اردو شکشک سنگھ نے استقبال کیا ہے۔ نیز مسلمانوں کو ریزرویشن کیوں نہیں، حکومت سچر کمیٹی، رنگا ناتھ مشرا کمیشن، محمود الرحمن کمیٹی کی سفارشات پر کب عمل کریگی؟ اسطرح کا سوال حکومت
indicdlp-v2-final-np · densest page · 954×1503 (1.4 MP)
indicdlp_np_bn_000432_1




stored text of this block
নিজস্ব প্রতিবেদন, মেন্দীপুর: পশ্চিম মেনীপুরের পুলিশ সুপার হিসেবে ফেসানো হল বুদ্ধিমান সরকারকে। এই জেলার দুটি লোকসভা কেন্দ্র ঘাটল এবং মেনীপুরের ২৫ মে নির্বাচন হল। যেখানে তিনি এবং তিনি আমাদের জেলা থেকে বুদ্ধিমান সরকারকে সরিয়ে দেয়। কমিশনের পক্ষ থেকে বলা হয় নির্বাচন সংক্রান্ত কোনও পদে থাকতে পারবেন না এই
indicdlp-v2-final-np · most-hurt-by-4MP page · 4175×6563 (27.4 MP)
indicdlp_v2_np_ne_000216_0




stored text of this block
एमालेले १ फागुनमा सत्ता गठबन्धनले महाभियोग प्रस्ताव दर्ता गरेपछि निलम्बनमा रहेका प्रधानन्यायाधीश चोलेन्द्रशमशेर राणाको पक्षमा समेत घोषणापत्रमा उल्लेख गरिएको छ। एमालेको घोषणापत्रमा सहकर्मी न्यायाधीशहरूको ४० दिनसम्म र नेपाल वारले १ सय दिनभन्दा लामो समय आन्दोलन गरेपछि महाभियोग लागेका राणामाथि सत्ता गठब
ncert-scert · biggest page · 5840×4479 (26.2 MP)
66987dd1-5362-4165-836b-f736abc26528__C06__BD_H-BSE__L_Punjabi__ch29__Page76




stored text of this block
<table> <tr> <td>ਲੜੀ ਨੰ.</td> <td>ਮਾਤਰਾ ਦਾ ਨਾਮ</td> <td>ਚਿੰਨ੍ਹ</td> <td>ਟਿੱਪੀ ਲੱਗਣ ਤੋਂ ਬਾਅਦ ਚਿੰਨ੍ਹ ਦਾ ਰੂਪ</td> <td>ਉਦਾਹਰਣ</td> </tr> <tr> <td>1.</td> <td>ਮੁਕਤਾ</td> <td>ਚਿੰਨ੍ਹ ਨਹੀਂ</td> <td>ਅੰ</td> <td>ਅੰਬ, ਰੰਗ, ਚੰਦ, ਸੰਤ, ਅੰਤ</td> </tr> <tr> <td>2.</td> <td>ਸਿਹਾਰੀ</td> <td>f</td> <td>fੰ</td> <td>ਪਿੰ
ncert-scert · densest page · 5840×4479 (26.2 MP)
66987dd1-5362-4165-836b-f736abc26528__C06__BD_H-BSE__L_Punjabi__ch29__Page76




stored text of this block
<table> <tr> <td>ਲੜੀ ਨੰ.</td> <td>ਮਾਤਰਾ ਦਾ ਨਾਮ</td> <td>ਚਿੰਨ੍ਹ</td> <td>ਟਿੱਪੀ ਲੱਗਣ ਤੋਂ ਬਾਅਦ ਚਿੰਨ੍ਹ ਦਾ ਰੂਪ</td> <td>ਉਦਾਹਰਣ</td> </tr> <tr> <td>1.</td> <td>ਮੁਕਤਾ</td> <td>ਚਿੰਨ੍ਹ ਨਹੀਂ</td> <td>ਅੰ</td> <td>ਅੰਬ, ਰੰਗ, ਚੰਦ, ਸੰਤ, ਅੰਤ</td> </tr> <tr> <td>2.</td> <td>ਸਿਹਾਰੀ</td> <td>f</td> <td>fੰ</td> <td>ਪਿੰ
ncert-scert · most-hurt-by-4MP page · 5840×4479 (26.2 MP)
66987dd1-5362-4165-836b-f736abc26528__C06__BD_H-BSE__L_Punjabi__ch17__Page51




stored text of this block
- ਪਤਾ - ਪੱਤਾ - ਜਗ - ਜੱਗ - ਸਜਾ - ਸੱਜਾ - ਕਤ - ਕੱਤ - ਬਚੇ - ਬੱਚੇ - ਸਤ - ਸੱਤ - ਬਦਲ - ਬੱਦਲ - ਜਿਦ - ਜਿੱਦ
printed-nat-archives · biggest page · 11182×7754 (86.7 MP)
879781c948308433fa568e06c6a2fe1d




stored text of this block
Achieving float altitudes in excess of 150, 000 feet has always been a balloon performance goal at Sea-Space Systems. To realize such performance, lightweight fabrics are mandatory. Our thin-film philosophy at Sea-Space, since the start of the Corporation three years ago, started from the premise th
printed-nat-archives · densest page · 1196×748 (0.9 MP)
811eae66ccaa8b13d62ecdabfd519679




stored text of this block
DECLASSIFIED Authority RF 74365
printed-nat-archives · most-hurt-by-4MP page · 8813×6555 (57.8 MP)
ca46712c967d932d14dbc1a78358bb4b




stored text of this block
Inflated Echo II shows panels and size in hangar. (Schjeldahl Photo)