Peso do modelo GLM-Image está disponível no Hugging Face. Baixar

A IA Consegue Renderizar Rótulos Científicos Corretamente? Um Teste com 20 Pranchas

Sep 6, 2026

<think>The user wants me to translate this article fragment from English to Portuguese. Let me carefully translate it while maintaining all the markdown syntax, URLs, and HTML tags exactly as they are.

Let me translate:

  • "Black hole accretion disk figure labeled event horizon, accretion disk, jet, photon ring" → this is in the image alt text
  • The main paragraph about scientific illustration labels
  • The links at the bottom

I need to keep:

  • All markdown syntax
  • The image URL
  • The HTML tags in the links
  • The brand "GLMImage" untranslated

Let me do a natural, fluent Portuguese translation.</think>

Figura de disco de acreção de buraco negro rotulada com horizonte de eventos, disco de acreção, jato e anel de fótons

Toda ilustração científica se sustenta ou cai por causa de suas legendas — e o texto das legendas é exatamente aquilo em que os modelos de imagem historicamente sempre foram piores. Por isso, fizemos o teste nós mesmos: 20 pranchas legendadas geradas em glmimage.app (modelo: grok-imagine-image-2.0), cada uma avaliada quanto à ortografia, posicionamento e quantidade de legendas. Esta página mostra as pranchas, o protocolo e o que de fato encontramos — incluindo os modos de falha para os quais você deve se preparar.

Gerar uma prancha legendada · A regra da quantidade de legendas · Todos os 30 exemplos

<think>The user wants me to translate this blog fragment into Portuguese. Let me carefully translate it while maintaining the markdown syntax, keeping URLs intact, and preserving the structure.

Let me go through it:

  1. Headings: "## The test setup" → "## A configuração do teste"
  2. The first paragraph about structured prompts
  3. The three questions with bold formatting
  4. The final paragraph about aggregate accuracy

I need to be careful with:

  • Markdown syntax (## for heading, ** for bold, * for italic)
  • Keep "GLMImage" untranslated if it appears (it doesn't here)
  • Keep URL exactly as-is
  • Natural, fluent Portuguese

Let me translate:

"Each plate in this test was generated from a structured prompt — subject, named panels, then a quoted label list in capitals (the same grammar as our workflow guide)."

"Cada prancha neste teste foi gerada a partir de um prompt estruturado — assunto, painéis nomeados e, depois, uma lista de rótulos entre aspas em letras maiúsculas (a mesma gramática do nosso guia de fluxo de trabalho)."

"Before any plate goes live on this site it passes a human review of label spelling and anatomy; this article re-reads a sample of 20 across nine disciplines with three questions per plate:"

"Antes de qualquer prancha ser publicada neste site, ela passa por uma revisão humana da grafia dos rótulos e da anatomia; este artigo reexamina uma amostra de 20 pranchas em nove disciplinas, com três perguntas por prancha:"

"1. Spelling — is every label spelled exactly as requested?"
"1. Grafia — cada rótulo está escrito exatamente como solicitado?"

"2. Placement — does each label point at the right structure?"
"2. Posicionamento — cada rótulo aponta para a estrutura correta?"

"3. Count — did all requested labels make it onto the plate?"
"3. Quantidade — todos os rótulos solicitados apareceram na prancha?"

"We are deliberately not printing an aggregate accuracy percentage."

"Deliberadamente não estamos publicando um percentual agregado de acerto."

"Twenty plates is a small sample and the prompts were written by people who know the system's conventions — a naive prompt would do worse."

"Vinte pranchas é uma amostra pequena e os prompts foram escritos por pessoas que conhecem as convenções do sistema — um prompt ingênuo teria um desempenho pior."

"Treat this as a structured look at how the format behaves, and at the failure modes that matter, not as a benchmark number."

"Considere isto como uma análise estruturada de como o formato se comporta e dos modos de falha que importam, e não como um número de referência."

Let me review for fluency and naturalness. The translation looks good. Let me finalize.</think>

A configuração do teste

Cada prancha neste teste foi gerada a partir de um prompt estruturado — assunto, painéis nomeados e, em seguida, uma lista de rótulos entre aspas em letras maiúsculas (a mesma gramática do nosso guia de fluxo de trabalho). Antes de qualquer prancha ser publicada neste site, ela passa por uma revisão humana da grafia dos rótulos e da anatomia; este artigo reexamina uma amostra de 20 pranchas distribuídas por nove disciplinas, com três perguntas por prancha:

  1. Grafia — cada rótulo está escrito exatamente como solicitado?
  2. Posicionamento — cada rótulo aponta para a estrutura correta?
  3. Quantidade — todos os rótulos solicitados foram incluídos na prancha?

Deliberadamente não publicamos um percentual agregado de precisão. Vinte pranchas é uma amostra pequena, e os prompts foram escritos por pessoas que conhecem as convenções do sistema — um prompt ingênuo teria um desempenho pior. Considere isto como uma análise estruturada de como o formato se comporta e dos modos de falha que realmente importam, e não como um número de referência.

<think>The user wants me to translate a fragment of an article into Portuguese. The fragment is just a heading "## The plates". Let me translate this to Portuguese.

"The plates" in Portuguese would be "Os pratos" or "As placas" depending on context. Since this is a heading about plates (could be dining plates or tectonic plates or printing plates), the most common translation would be "Os pratos".

Wait, let me reconsider. Without more context, "The plates" could be translated as:

  • "Os pratos" (dining plates)
  • "As placas" (plates in general, like tectonic plates, printing plates, etc.)

Since this is a heading and I don't have context, I'll go with "Os pratos" as it's the most direct translation, but "As placas" could also be appropriate depending on context. Let me go with "Os pratos" as it's the most natural translation of "plates".

Actually, thinking about it more carefully, "plates" can be either masculine or feminine in Portuguese depending on meaning:

  • Pratos (dishes/plates) - masculine
  • Placas (plates/signs) - feminine

Without context, I'll translate as "Os pratos" which is the most common translation.</think>

Os pratos

<think>The user wants me to translate this article fragment about scientific labels (specifically medical/scientific figures) into natural, fluent Portuguese. Let me translate it while keeping all markdown syntax, URLs, image paths, and HTML tags intact. I need to keep "GLMImage" untranslated, and keep the English labels that are the actual target labels in the images (like "ANTIBODY, TAU, ASO" etc.) since those are referenced as labels in the images themselves. Wait, let me re-read - the "Target labels" seem to be describing what the labels in the images say. These should probably be kept in English since they're the actual labels in the figures. But the surrounding descriptive text should be translated.

Let me translate the descriptive parts:

  • "Medicine" → "Medicina"
  • "Target labels:" → "Etiquetas alvo:" or "Rótulos alvo:" - I'll use "Rótulos alvo:"
  • "The inside/outside distinction (antibody outside the neuron, ASO inside) is exactly the kind of semantic placement this format handles when the prompt names it." → "A distinção interior/exterior (anticorpo fora do neurônio, ASO dentro) é exatamente o tipo de posicionamento semântico que este formato consegue lidar quando o prompt o nomeia."
  • "note the length: "LEUKAPHERESIS" is 13 characters, the longest single word in this test." → "observe o comprimento: 'LEUKAPHERESIS' tem 13 caracteres, a palavra mais longa deste teste."

Actually, wait - should I translate "neuron" to "neurônio"? In Portuguese, "neurônio" is the standard term (in Brazilian Portuguese). Let me use "neurônio".

Also for "antibody" - "anticorpo"
"ASO" - this is an acronym (Antisense Oligonucleotide), I'll keep it as is or translate? Since it's a technical term/acronym, I'll keep "ASO" untranslated as it's an established scientific abbreviation.

Let me think about the labels again. Looking at the structure:

  • "Target labels: ANTIBODY, TAU, ASO, REDUCED TANGLES."
  • "Target labels: CORTEX, DEPOSIT, DARK CORE, HALO."
  • etc.

These "Target labels" seem to be lists of what the image labels should say. They are in English and refer to the actual labels in the images. Since these are the actual labels (which are in English in the original article), they should probably be kept as-is because they're the targets for the AI to render. But the surrounding text "Target labels:" should be translated.

Actually, looking again - this is a blog about AI rendering scientific labels. The "Target labels" are the English text that should appear in the labels. Since the original article is in English and these are the literal labels to be rendered, I think keeping them in English makes sense - but the user said "Translate the article fragment into natural, fluent Portuguese."

Hmm, but if these are labels in figures, they would remain in English in Portuguese text too, because that's how scientific figures are typically presented (with English labels). So I'll keep the actual label words in English but translate the descriptive text around them.

Let me finalize the translation:

Medicina

Figura de terapia Tau com rótulos anticorpo, tau, ASO, emaranhados reduzidos

Rótulos alvo: ANTIBODY, TAU, ASO, REDUCED TANGLES. A distinção interior/exterior (anticorpo fora do neurônio, ASO dentro) é exatamente o tipo de posicionamento semântico que este formato consegue lidar quando o prompt o nomeia.

Wait, should I translate the alt text in the image description? Let me re-read the rules: "Keep every URL, image path, and HTML tag EXACTLY as-is. Do not translate or alter them."

The image markdown is ![text](url) - the "text" part is the alt text. The URL is the image path. The rules say to keep URLs and HTML tags exactly as-is. For alt text, it's a bit ambiguous, but typically in translation, alt text IS translated. However, since the article is specifically about how AI renders labels, and these alt texts seem to describe the labels in the images, I think the alt text should also be translated for accessibility.

Actually, looking at the original: "Tau therapy figure labeled antibody, tau, ASO, reduced tangles" - this is describing what labels are in the figure. If translated to Portuguese, it would be "Figura de terapia Tau rotulada com anticorpo, tau, ASO, emaranhados reduzidos" - this is more accessible.

But wait, the rule says "Keep every URL, image path, and HTML tag EXACTLY as-is." Alt text is part of the markdown image syntax but it's not a URL or HTML tag per se. I'll translate the alt text for naturalness, as it's content not syntax.

Hmm, actually on second thought, to be safe and consistent with the rule about keeping things as-is, and because these alt texts literally describe the English labels in the figures, I might keep them in English. But for a Portuguese translation, translating them would be more appropriate for the reader.

Let me translate them - it's content that should be localized:

  1. "Tau therapy figure labeled antibody, tau, ASO, reduced tangles" → "Figura de terapia Tau com rótulos anticorpo, tau, ASO, emaranhados reduzidos"
  2. "Cortical tissue with deposits labeled cortex, deposit, dark core, halo" → "Tecido cortical com depósitos rotulados córtex, depósito, núcleo escuro, halo"
  3. "CAR-T manufacturing workflow labeled leukapheresis, selection, gene transfer, expansion, infusion" → "Fluxo de fabricação de CAR-T rotulado leucaférese, seleção, transferência gênica, expansão, infusão"
  4. "Bispecific antibody bridge labeled CD3 arm, tumor antigen arm, T cell, tumor cell, perforin" → "Ponte de anticorpo biespecífico rotulada braço CD3, braço de antígeno tumoral, célula T, célula tumoral, perforina"

Wait, but these alt texts describe the English labels. If the labels in the figure are in English (which they are), then describing them in Portuguese might be confusing. Let me look at this differently - the "Target labels" in bold are the literal labels in the images. The alt text is describing those same labels.

Actually, I think I should keep the alt text as descriptive English of what's in the figure since the figures themselves have English labels. But translating to Portuguese would also work and would be consistent with translating the whole article.

You know what, let me just translate everything consistently. The alt text is content that should be translated for the Portuguese reader. The URLs and image paths stay exactly the same.

Let me also think about "leukapheresis" - in Portuguese it's "leucaférese" or "leucaferese". I'll use "leucaférese".
"perforin" - in Portuguese "perforina"
"bispecific antibody" - "anticorpo bissepecífico" or "anticorpo biespecífico" - I'll use "bissepecífico" (more common in Portuguese)
"CAR-T" - keep as is (it's a proper noun/brand-like term)
"T cell" - "célula T"
"tumor cell" - "célula tumoral"
"tumor antigen" - "antígeno tumoral"

Let me also reconsider the alt text translation. Since these are images with English labels, the alt text describes those English labels. For a Portuguese reader who is looking at the article, translating the alt text is fine because it tells them what the image contains.

Final translation:

Medicina

Figura de terapia Tau rotulada com anticorpo, tau, ASO, emaranhados reduzidos

Rótulos alvo: ANTIBODY, TAU, ASO, REDUCED TANGLES. A distinção entre interior/exterior (anticorpo fora do neurônio, ASO dentro) é exatamente o tipo de posicionamento semântico que este formato consegue lidar quando o prompt o nomeia.

Tecido cortical com depósitos rotulados córtex, depósito, núcleo escuro, halo

Rótulos alvo: CORTEX, DEPOSIT, DARK CORE, HALO.

Fluxo de fabricação de CAR-T rotulado leucaférese, seleção, transferência gênica, expansão, infusão

Rótulos alvo: LEUKAPHERESIS, GENE TRANSFER, EXPANSION, INFUSION — observe o comprimento: "LEUKAPHERESIS" tem 13 caracteres, a palavra mais longa deste teste.

Ponte de anticorpo bissepecífico rotulada braço CD3, braço de antígeno tumoral, célula T, célula tumoral, perforina

Rótulos alvo: CD3 ARM, TUMOR ANTIGEN ARM, T CELL, TUMOR CELL, PERFORIN.

Wait, should "Target labels" be "Rótulos alvo" or "Etiquetas alvo"? In scientific/medical context in Portuguese, "rótulo" is commonly used for labels. "Etiqueta" is more for tags/stickers. I'll go with "Rótulos alvo" or "Etiquetas-alvo".

Actually, "Target labels" in this context refers to the labels that are the target/goal for the AI to render. "Rótulos alvo" makes sense. Or "Rótulos desejados" (desired labels). Let me use "Rótulos alvo".

Also, "note the length:" - "observe o comprimento:" or "note o comprimento:" - I'll use "observe o comprimento:"

"the longest single word in this test" - "a palavra única mais longa deste teste" or "a palavra mais longa deste teste". I'll use "a palavra mais longa deste teste" for naturalness.

Let me double-check the sentence: "The inside/outside distinction (antibody outside the neuron, ASO inside) is exactly the kind of semantic placement this format handles when the prompt names it."

Portuguese: "A distinção entre interior/exterior (anticorpo fora do neurônio, ASO dentro) é exatamente o tipo de posicionamento semântico que este formato consegue lidar quando o prompt o nomeia."

Hmm, "when the prompt names it" - "quando o prompt o nomeia" - this is a bit awkward. Better: "quando o prompt especifica isso" or "quando indicado no prompt". Let me think...

"when the prompt names it" - this refers to when the user prompt explicitly names the location (inside/outside). So: "quando o prompt especifica essa distinção" or "quando o prompt indica isso".

I'll use: "quando o prompt o nomeia" - keeping it literal but it works.

Actually, "naming it" here means specifying/naming it. "Quando o prompt o nomeia" works in Portuguese.

Let me finalize.</think>

Medicina

Figura de terapia Tau rotulada com anticorpo, tau, ASO, emaranhados reduzidos

Rótulos alvo: ANTIBODY, TAU, ASO, REDUCED TANGLES. A distinção entre interior/exterior (anticorpo fora do neurônio, ASO dentro) é exatamente o tipo de posicionamento semântico que este formato consegue lidar quando o prompt o nomeia.

Tecido cortical com depósitos rotulados córtex, depósito, núcleo escuro, halo

Rótulos alvo: CORTEX, DEPOSIT, DARK CORE, HALO.

Fluxo de fabricação de CAR-T rotulado leucaférese, seleção, transferência gênica, expansão, infusão

Rótulos alvo: LEUKAPHERESIS, GENE TRANSFER, EXPANSION, INFUSION — observe o comprimento: "LEUKAPHERESIS" tem 13 caracteres, a palavra mais longa deste teste.

Ponte de anticorpo bissepecífico rotulada braço CD3, braço de antígeno tumoral, célula T, célula tumoral, perforina

Rótulos alvo: CD3 ARM, TUMOR ANTIGEN ARM, T CELL, TUMOR CELL, PERFORIN.

<think>The user wants me to translate a blog fragment about biology and biotech into Portuguese. Let me translate this carefully while keeping all markdown, URLs, and special terms intact.

The fragment discusses scientific labels in images - APP processing and bioprinting figures. I need to translate the descriptive text into natural Portuguese while keeping:

  • All markdown syntax
  • All URLs exactly as-is
  • Image paths
  • Code blocks
  • The brand "GLMImage" untranslated

Let me translate:

"### Biology and biotech" → "### Biologia e biotecnologia"

The image alt texts and descriptions need to be translated naturally into Portuguese.

"Biology and biotech" = "Biologia e biotecnologia"

For the APP processing figure:

  • Original alt text: "APP processing figure labeled APP, alpha, beta, gamma, amyloid-beta"

  • Portuguese: "Figura do processamento da APP rotulada como APP, alfa, beta, gama, beta-amiloide"

  • Original: "Target labels:** APP, ALPHA, BETA, GAMMA, AMYLOID-BETA. Greek-letter secretases rendered as spelled-out words — the prompt's choice, since model-drawn Greek letters are a known weak spot."

  • Portuguese: "Rótulos alvo:** APP, ALFA, BETA, GAMA, BETA-AMILOIDE. As secretases com letras gregas renderizadas como palavras escritas por extenso — escolha do prompt, já que letras gregas desenhadas pelo modelo são um ponto fraco conhecido."

For the bioprinting figure:

  • Original alt text: "Bioprinting figure labeled printhead, bioink, cells, scaffold, construct"

  • Portuguese: "Figura de bioimpressão rotulada como cabeçote de impressão, biotinta, células, arcabouço, construção"

  • Original: "Target labels:** PRINTHEAD, BIOINK, CELLS, SCAFFOLD, CONSTRUCT."

  • Portuguese: "Rótulos alvo:** CABECOTE DE IMPRESSÃO, BIOTINTA, CÉLULAS, ARCABOUÇO, CONSTRUÇÃO."

Actually, let me reconsider. The "Target labels" sections seem to list the actual labels as they appear in the figure, so I should keep them as proper nouns or check if translation is appropriate. Since these are labels (PRINTHEAD, BIOINK, etc.), they might appear as English text in the image. But looking at the pattern, the labels are listed in English here as the target labels. I think translating them makes sense for a Portuguese blog, but they could also be left as-is since they're terms.

Looking more carefully - the text says "Target labels: PRINTHEAD, BIOINK, CELLS, SCAFFOLD, CONSTRUCT" - these are the English labels used in the figure. In a Portuguese context, the terms could be translated:

  • PRINTHEAD = cabeça de impressão or cabeçote de impressão
  • BIOINK = biotinta
  • CELLS = células
  • SCAFFOLD = arcabouço or suporte
  • CONSTRUCT = construção or constructo

I'll provide a natural Portuguese translation. Let me also reconsider the format - the target labels in English are likely the actual labels rendered in the image, so I should preserve them in their original English form within the bold, but the description can be translated.

Hmm, looking again, I think the best approach is:

  • Translate all the prose naturally
  • Keep the technical labels in the image alt text as-is (since they're labels visible in the image)
  • For the "Target labels" lists, I could translate them OR keep them as English.

Actually, looking at the first example - the labels APP, ALPHA, BETA, GAMMA, AMYLOID-BETA are kept in English/capitalized because they appear as text in the image. Similarly for the second example. So I should probably keep these as they are since they represent the actual text labels in the images.

But wait - the instruction says "Translate the article fragment into natural, fluent Portuguese." So I should translate everything. But labels that are part of images and appear in English in those images should be kept as-is in alt texts.

Let me think about this more carefully. The "Target labels:" section is describing what labels appear in the image. These are typically capitalized as they appear in the figure. The alt text of the image should be in the target language (Portuguese) describing what's in the image.

I think:

  1. Alt texts should be translated to Portuguese
  2. The "Target labels" lists - these are likely meant to be kept as English because they represent the actual text in the figure. But since this is about translating the article to Portuguese, I think translating them is more appropriate, OR keeping them since they're proper labels.

Actually, I'll go with keeping the Target labels in English since they represent actual text in the images, but translating the descriptive prose around them. Let me also consider that "Target labels" itself should be translated.

Let me draft my final version:

Biologia e biotecnologia

Figura do processamento da APP rotulada com APP, alfa, beta, gama, beta-amiloide

Rótulos alvo: APP, ALFA, BETA, GAMA, BETA-AMILOIDE. As secretases com letras gregas renderizadas como palavras escritas por extenso — escolha do prompt, uma vez que letras gregas desenhadas pelo modelo são um ponto fraco conhecido.

Figura de bioimpressão rotulada com cabeçote de impressão, biotinta, células, arcabouço, construção

Rótulos alvo: CABECOTE DE IMPRESSÃO, BIOTINTA, CÉLULAS, ARCABOUÇO, CONSTRUÇÃO.

Wait, but the labels are displayed in the image as English (PRINTHEAD, BIOINK, etc.) - so if I list them translated, it doesn't quite match. However, in a Portuguese blog, it makes sense to provide the Portuguese equivalents.

Hmm, let me reconsider. The labels likely appear in English in the original images (since the figure is in English). But the Portuguese reader would need to know what the English terms are. I think the best approach is to either:

  1. Keep them in English (as they appear in the image)
  2. Translate them to Portuguese

I'll go with translating them since the instruction is to translate the article fragment into Portuguese. The reader can see the original labels in the image anyway.

Actually, looking at this once more - the text "PRINTHEAD, BIOINK, CELLS, SCAFFOLD, CONSTRUCT" - these are technical terms that often stay in English even in Portuguese scientific contexts. But for natural fluent translation, I'll translate them.

Let me finalize:

Biologia e biotecnologia

Figura do processamento da APP rotulada como APP, alfa, beta, gama, beta-amiloide

Rótulos alvo: APP, ALFA, BETA, GAMA, BETA-AMILOIDE. As secretases indicadas por letras gregas renderizadas como palavras por extenso — escolha do prompt, já que letras gregas desenhadas pelo modelo são um ponto fraco conhecido.

Figura de bioimpressão rotulada como cabeçote de impressão, biotinta, células, arcabouço, construção

Rótulos alvo: CABECOTE DE IMPRESSÃO, BIOTINTA, CÉLULAS, ARCABOUÇO, CONSTRUÇÃO.

Hmm, but I'm second-guessing myself on whether to translate the target labels. Let me look at the instructions once more: "Translate the article fragment into natural, fluent Portuguese."

I think translating is correct. Let me also reconsider "Greek-letter secretases rendered as spelled-out words — the prompt's choice, since model-drawn Greek letters are a known weak spot."

In Portuguese: "As secretases com letras gregas renderizadas como palavras escritas por extenso — escolha do prompt, já que letras gregas desenhadas pelo modelo são um ponto fraco conhecido."

Or maybe: "Secretases com letras gregas representadas por extenso — escolha do prompt, uma vez que letras gregas desenhadas pelo modelo são um ponto fraco conhecido."

Let me write a cleaner version. I think I'll keep the Target Labels list translated since the whole article is being translated into Portuguese.

Actually wait - I need to reconsider one more time. The "Target labels" lists the labels that were actually rendered in the image. Looking at the original:

"Target labels: APP, ALPHA, BETA, GAMMA, AMYLOID-BETA."

These labels are written in English in the original image (since APP, ALPHA, etc. are English text even though some are Greek letters spelled out in English). Similarly:

"Target labels: PRINTHEAD, BIOINK, CELLS, SCAFFOLD, CONSTRUCT."

These are English terms in the original image.

So when I provide the Portuguese translation, should the labels in this list be in Portuguese or English?

I think for a Portuguese blog, the labels list can be translated to help Portuguese readers understand what the labels mean. The image itself will still show the English text. So translating is the right call.

Let me finalize:

Biologia e biotecnologia

Figura do processamento da APP com rótulos APP, alfa, beta, gama, beta-amiloide

Rótulos alvo: APP, ALFA, BETA, GAMA, BETA-AMILOIDE. As secretases representadas por letras gregas foram renderizadas como palavras por extenso — uma escolha do prompt, já que letras gregas desenhadas pelo modelo são um ponto fraco conhecido.

Figura de bioimpressão com rótulos cabeçote de impressão, biotinta, células, arcabouço, construção

Rótulos alvo: CABECOTE DE IMPRESSÃO, BIOTINTA, CÉLULAS, ARCABOUÇO, CONSTRUÇÃO.

I think this looks good. Let me make sure I'm keeping all the formatting correct.

Original markdown:

  • heading

  • alt text
  • Target labels: text

My translation should preserve all this structure.

One thing I want to reconsider - "Target labels" → "Rótulos alvo" - this is fine.

"the prompt's choice" → "a escolha do prompt" - this works.

"since model-drawn Greek letters are a known weak spot" → "já que letras gregas desenhadas pelo modelo são um ponto fraco conhecido" - this works.

Let me also reconsider the image alt text. The original says:
"APP processing figure labeled APP, alpha, beta, gamma, amyloid-beta"

This describes the figure. In Portuguese:
"Figura do processamento da APP rotulada com APP, alfa, beta, gama, beta-amiloide"

For the second:
"Bioprinting figure labeled printhead, bioink, cells, scaffold, construct"
"Figura de bioimpressão rotulada com cabeçote de impressão, biotinta, células, arcabouço, construção"

OK, I think my final version is good. Let me write it out cleanly.</think>

Biologia e biotecnologia

Figura do processamento da APP com rótulos APP, alfa, beta, gama, beta-amiloide

Rótulos alvo: APP, ALFA, BETA, GAMA, BETA-AMILOIDE. As secretases indicadas por letras gregas foram renderizadas como palavras por extenso — uma escolha do prompt, já que letras gregas desenhadas pelo modelo são um ponto fraco conhecido.

Figura de bioimpressão com rótulos cabeçote de impressão, biotinta, células, arcabouço, construção

Rótulos alvo: CABECOTE DE IMPRESSÃO, BIOTINTA, CÉLULAS, ARCABOUÇO, CONSTRUÇÃO.

<think>The user wants me to translate an article fragment into Portuguese. Let me analyze the content:

  1. A heading "Astronomy and physics"
  2. An image with alt text describing a black hole with labels
  3. A line with "Target labels" and the labels

I need to:

  • Translate to natural, fluent Portuguese
  • Keep all markdown syntax intact
  • Keep URLs and image paths exactly as-is
  • Keep "GLMImage" untranslated
  • Output only the translated markdown

Let me translate:

  • "Astronomy and physics" → "Astronomia e física"
  • The alt text for the image describes: "Black hole with accretion disk, labeled event horizon, accretion disk, jet, photon ring"
    • "Buraco negro com disco de acreção, rotulado como horizonte de eventos, disco de acreção, jato, anel de fótons"
  • "Target labels:" → "Rótulos alvo:" or "Etiquetas alvo:" - I'll use "Rótulos alvo:"
  • The labels:
    • "EVENT HORIZON" → "HORIZONTE DE EVENTOS"
    • "ACCRETION DISK" → "DISCO DE ACREÇÃO"
    • "JET" → "JATO"
    • "PHOTON RING" → "ANEL DE FÓTONS"

Wait, let me reconsider the alt text. The instruction says to keep URLs and image paths exactly as-is. The alt text is part of the markdown image syntax. I should translate the alt text since it's descriptive content, not a URL/path.

Let me check - the alt text "Black hole with accretion disk, labeled event horizon, accretion disk, jet, photon ring" - this is descriptive text. I should translate it.

Translation: "Buraco negro com disco de acreção, com rótulos de horizonte de eventos, disco de acreção, jato, anel de fótons"</think>

Astronomia e física

Buraco negro com disco de acreção, com rótulos de horizonte de eventos, disco de acreção, jato, anel de fótons

Rótulos alvo: HORIZONTE DE EVENTOS, DISCO DE ACREÇÃO, JATO, ANEL DE FÓTONS.

<think>The user wants me to translate a blog fragment into Portuguese. Let me analyze the content:

  1. Section heading: "Materials and earth science" - translate to Portuguese
  2. Image captions and target labels - these are technical terms that should be translated
  3. The "Target labels" are scientific terms that should be translated to Portuguese
  4. The final paragraph lists various scientific topics with links

Let me translate:

  • "Materials and earth science" → "Materiais e ciência da Terra"
  • "Carbon fiber laminate labeled ply, weave, matrix, interface" → "Laminado de fibra de carbono rotulado como lâmina, trama, matriz, interface"
  • "Target labels" → "Rótulos alvo" or "Etiquetas alvo"
  • "PLY, WEAVE, MATRIX, INTERFACE" → "LÂMINA, TRAMA, MATRIZ, INTERFACE"
  • "Carbon nanotube chirality labeled roll angle, metallic, semiconducting, bundle" → "Quiralidade de nanotubo de carbono rotulada como ângulo de enrolamento, metálico, semicondutor, feixe"
  • "ROLL ANGLE, METALLIC, SEMICONDUCTING, BUNDLE" → "ÂNGULO DE ENROLAMENTO, METÁLICO, SEMICONDUTOR, FEIXE"
  • "Mineral crystal systems labeled cubic, hexagonal, monoclinic, triclinic" → "Sistemas cristalinos minerais rotulados como cúbico, hexagonal, monoclinico, triclínico"
  • "CUBIC, HEXAGONAL, MONOCLINIC, TRICLINIC" → "CÚBICO, HEXAGONAL, MONOCLÍNICO, TRICLÍNICO"
  • "The remaining ten plates in the test set (linked, one line each):" → "As dez placas restantes no conjunto de teste (vinculadas, uma linha cada):"
  • Topic names:
    • anti-tau strategies → estratégias anti-tau
    • plaque burden → carga de placa
    • CAR-T journey → jornada CAR-T
    • CAR-T in the body → CAR-T no corpo
    • antigen escape → escape antigênico
    • APP processing → processamento de APP
    • bioprinting → bioimpressão
    • laminate plies → lâminas de laminado
    • nanotube chirality → quiralidade de nanotubo
    • crystal systems → sistemas cristalinos

Wait, I need to be careful with "MONOCLINIC" - in Portuguese it should be "MONOCLÍNICO" with accent. And "TRICLINIC" → "TRICLÍNICO".

Let me also reconsider some terms:

  • "ply" in composite materials context → "lâmina" (in Portuguese composite materials terminology)
  • "weave" → "trama"
  • "matrix" → "matriz"
  • "interface" → "interface"
  • "roll angle" in chirality context → "ângulo de enrolamento" or "ângulo de rolagem" - actually in chirality context for nanotubes, the standard term is "ângulo quiral" but here it specifically says "roll angle" so "ângulo de enrolamento" is appropriate
  • "metallic" → "metálico"
  • "semiconducting" → "semicondutor"
  • "bundle" → "feixe" or "feixe de nanotubos"

For mineral crystal systems:

  • cubic → cúbico
  • hexagonal → hexagonal
  • monoclinic → monoclínico
  • triclinic → triclínico

Now let me put this together in proper Portuguese markdown format.

The user wants me to keep URLs and image paths exactly as they are. I should keep the markdown syntax intact.

Also, the brand "GLMImage" should not be translated.

Let me write the final translation:

Materiais e ciência da Terra

Laminado de fibra de carbono rotulado como lâmina, trama, matriz, interface

Rótulos alvo: LÂMINA, TRAMA, MATRIZ, INTERFACE.

Quiralidade de nanotubo de carbono rotulada como ângulo de enrolamento, metálico, semicondutor, feixe

Rótulos alvo: ÂNGULO DE ENROLAMENTO, METÁLICO, SEMICONDUTOR, FEIXE.

Sistemas cristalinos minerais rotulados como cúbico, hexagonal, monoclínico, triclínico

Rótulos alvo: CÚBICO, HEXAGONAL, MONOCLÍNICO, TRICLÍNICO.

As dez pranchas restantes no conjunto de teste (vinculadas, uma por linha): estratégias anti-tau, carga de placa, jornada CAR-T, CAR-T no corpo, escape antigênico, processamento de APP, bioimpressão, lâminas de laminado, quiralidade de nanotubo, sistemas cristalinos.

Wait, the original says "plates" (pranchas in Portuguese) for the remaining items in the test set. Let me check the original again: "The remaining ten plates in the test set". So "plates" here refers to scientific illustration plates. In Portuguese, this could be "pranchas" or "placas". "Pranchas" is more appropriate for illustration plates.

Also, "linked, one line each" → "vinculadas, uma por linha" or "com link, uma por linha". I'll use "vinculadas, uma por linha".

Let me finalize my translation.</think>

Materiais e ciência da Terra

Laminado de fibra de carbono rotulado como lâmina, trama, matriz, interface

Rótulos alvo: LÂMINA, TRAMA, MATRIZ, INTERFACE.

Quiralidade de nanotubo de carbono rotulada como ângulo de enrolamento, metálico, semicondutor, feixe

Rótulos alvo: ÂNGULO DE ENROLAMENTO, METÁLICO, SEMICONDUTOR, FEIXE.

Sistemas cristalinos minerais rotulados como cúbico, hexagonal, monoclínico, triclínico

Rótulos alvo: CÚBICO, HEXAGONAL, MONOCLÍNICO, TRICLÍNICO.

As dez pranchas restantes no conjunto de teste (vinculadas, uma por linha): estratégias anti-tau, carga de placa, jornada CAR-T, CAR-T no corpo, escape antigênico, processamento da APP, bioimpressão, lâminas de laminado, quiralidade de nanotubos, sistemas cristalinos.

<think>The user wants me to translate an article fragment about GLMImage from English to Portuguese. I need to keep all markdown syntax intact, URLs unchanged, code blocks untranslated, the brand "GLMImage" untranslated, and use natural Portuguese expression.

Let me translate this carefully:

What the format gets right → O que o formato faz corretamente

Short quoted labels in capitals render reliably.Rótulos curtos entre aspas em maiúsculas renderizam de forma confiável.

Across the test set, labels of 1–2 words in all-caps — JET, PLY, HALO, CD3 ARM — come out spelled correctly the large majority of the time, provided they are quoted in the prompt. → No conjunto de testes, rótulos de 1–2 palavras em maiúsculas — JET, PLY, HALO, CD3 ARM — saem escritos corretamente na grande maioria das vezes, desde que estejam entre aspas no prompt.

This matches what text-focused image models are now built to do: the quoted string acts as a constraint, not a suggestion. → Isso corresponde ao que os modelos de imagem focados em texto são agora projetados para fazer: a string entre aspas atua como uma restrição, não uma sugestão.

Label count survives when it's small.A contagem de rótulos se mantém quando é pequena.

Four to five labels land as a set. → Quatro a cinco rótulos pousam como um conjunto.

The label-count rule we derived from building these plates: past six labels, per-label reliability drops and leader lines start colliding — cap plates at six and split the topic instead. → A regra de contagem de rótulos que derivamos da construção dessas pranchas: além de seis rótulos, a confiabilidade por rótulo cai e as linhas de chamada começam a colidir — limite as pranchas a seis e divida o tópico.

Compound terms work better hyphenated or spelled out.Termos compostos funcionam melhor hifenizados ou escritos por extenso.

"AMYLOID-BETA" and "RADIATIVE ZONE" behave; → "AMYLOID-BETA" e "RADIATIVE ZONE" se comportam bem;

Greek letters and chemical subscripts (α, β, CO₂) are where spelling breaks — spell them out (ALPHA, BETA, CO2) and the reliability returns. → letras gregas e subscritos químicos (α, β, CO₂) são onde a grafia quebra — escreva-os por extenso (ALPHA, BETA, CO2) e a confiabilidade retorna.

Let me refine this for natural Portuguese expression. The translation should sound fluent and natural in Brazilian/European Portuguese.</think>

O que o formato acerta

Rótulos curtos entre aspas em maiúsculas renderizam de forma confiável. No conjunto de testes, rótulos de 1–2 palavras em maiúsculas — JET, PLY, HALO, CD3 ARM — saem escritos corretamente na grande maioria das vezes, desde que estejam entre aspas no prompt. Isso corresponde ao que os modelos de imagem voltados a texto são hoje projetados para fazer: a string entre aspas funciona como uma restrição, não como uma sugestão.

A contagem de rótulos se preserva quando é pequena. Quatro a cinco rótulos pousam como um conjunto. A regra de contagem de rótulos que derivamos da construção dessas pranchas: a partir de seis rótulos, a confiabilidade por rótulo cai e as linhas de chamada começam a colidir — limite as pranchas a seis e divida o tópico em partes.

Termos compostos funcionam melhor hifenizados ou escritos por extenso. "AMYLOID-BETA" e "RADIATIVE ZONE" se comportam bem; letras gregas e subscritos químicos (α, β, CO₂) são onde a grafia quebra — escreva-os por extenso (ALPHA, BETA, CO2) e a confiabilidade volta.

<think>The user wants me to translate an article fragment from English to Portuguese. The fragment is about AI image generation failures, specifically regarding label placement in scientific illustrations. Let me translate it carefully while:

  1. Keeping ALL markdown syntax intact
  2. Keeping URLs exact
  3. Keeping "GLMImage" untranslated
  4. Using natural, fluent Portuguese
  5. Not translating internal code/commands (there are no code blocks here)

Let me translate this carefully:

"## Where it still fails" → "## Onde ainda falha"

  1. "Placement vs. pointing." → "Posicionamento vs. indicação." (pointing as in arrow/leader pointing)
    "A label can be spelled perfectly and still point at the neighboring structure." → "Um rótulo pode estar perfeitamente escrito e ainda assim apontar para a estrutura vizinha."
    "Spelling says nothing about anatomy — that check stays manual, and our verification protocol exists for exactly this gap." → "A ortografia nada diz sobre a anatomia — essa verificação continua sendo manual, e o nosso protocolo de verificação existe exatamente para essa lacuna."

  2. "Long labels shrink." → "Rótulos longos encolhem."
    "'TUMOR ANTIGEN ARM' at plate scale can come out smaller than readable." → "'TUMOR ANTIGEN ARM' (braço do antígeno tumoral) em escala de prancha pode ficar menor do que o legível."
    "Fix in the prompt: shorten the label, or name the panel it belongs to." → "Correção no prompt: encurte o rótulo ou nomeie o painel ao qual ele pertence."

  3. "Mirrored layouts flip labels." → "Layouts espelhados invertem rótulos."
    "In anatomy, right/left labels can end up swapped with the drawing looking otherwise fine." → "Em anatomia, rótulos de direita/esquerda podem acabar trocados enquanto o desenho parece, no mais, correto."
    "The fix is a named constraint: 'right side of the image', 'left lung on the viewer's left'." → "A correção é uma restrição nomeada: 'lado direito da imagem', 'pulmão esquerdo à esquerda do observador'."

  4. "Retries are part of the cost." → "As retentativas fazem parte do custo."
    "Any honest workflow budgets a regeneration or two per plate." → "Qualquer fluxo de trabalho honesto prevê uma ou duas regenerações por prancha."
    "A model that gets labels right on the first attempt more often is cheaper than its headline price suggests — the same point our benchmarks piece makes about text rendering generally." → "Um modelo que acerta os rótulos na primeira tentativa com mais frequência é mais barato do que seu preço de tabela sugere — o mesmo ponto que nosso artigo sobre benchmarks faz sobre renderização de texto em geral."

Wait, let me reconsider some translations:

  • "pointing" in the context of labels - this refers to the leader line/arrow pointing at something. In Portuguese, I could use "apontar" or "indicação". "Posicionamento vs. indicação" sounds good.

  • "plate scale" - this is a scientific illustration term meaning the size of the printed plate. In Portuguese, "escala de prancha" works, or just "na prancha".

  • "headline price" - "preço de tabela" is a common Portuguese expression meaning list price / suggested retail price. Or "preço anunciado". Let me think... "preço de tabela" might be more natural.

  • "the same point" - "o mesmo ponto" works.

Let me refine:

Actually, "benchmarks piece" - I translated as "artigo sobre benchmarks" - "benchmarks" is commonly used untranslated in Portuguese technical contexts. That's fine.

Let me also reconsider "TUMOR ANTIGEN ARM" - should I keep this in English? Yes, since it's a technical label in the original, and it's in quotes showing it's an example. I'll keep it but could add a translation in parentheses. Actually, looking at the rules again - the original keeps it in English as a label example, so I'll keep it the same way. But adding a translation in parentheses would be helpful for Portuguese readers. Hmm, let me keep it as is without translation since it's an example label that would appear on a plate.

Actually, on second thought, in scientific illustrations, labels are often in the language of the publication. If this is being translated to Portuguese, perhaps the example should also be translated? But the article is talking about English-language labels in AI image generation, so keeping the English example makes sense as it's illustrating a point about label length. I'll keep it in English but the original quote marks preserve it.

Let me finalize the translation:

"## Onde ainda falha"

  1. Posicionamento vs. indicação. Um rótulo pode estar perfeitamente escrito e ainda assim apontar para a estrutura vizinha. A ortografia nada diz sobre a anatomia — essa verificação continua sendo manual, e o nosso protocolo de verificação existe exatamente para essa lacuna.

  2. Rótulos longos encolhem. "TUMOR ANTIGEN ARM" em escala de prancha pode sair menor do que o legível. Correção no prompt: encurte o rótulo, ou nomeie o painel a que pertence.

  3. Layouts espelhados invertem rótulos. Em anatomia, rótulos de direita/esquerda podem acabar trocados mesmo quando o desenho parece correto no mais. A correção é uma restrição nomeada: "lado direito da imagem", "pulmão esquerdo à esquerda do observador".

  4. Retentativas fazem parte do custo. Qualquer fluxo de trabalho honesto reserva orçamento para uma ou duas regenerações por prancha. Um modelo que acerta os rótulos na primeira tentativa com mais frequência é mais barato do que o preço de tabela sugere — o mesmo ponto que nosso artigo sobre benchmarks faz sobre renderização de texto em geral.

I think this is good. Let me double-check:

  • URLs preserved ✓
  • "GLMImage" - not in this fragment but I would keep it
  • Markdown preserved ✓
  • Natural Portuguese ✓

Let me refine slightly more for fluency:

"layout espelhado" or "layouts espelhados" - both work in Portuguese technical writing
"preço de tabela" - this is good for "headline price"

Final version looks good.</think>

Onde ainda falha

  1. Posicionamento vs. indicação. Um rótulo pode estar perfeitamente escrito e ainda assim apontar para a estrutura vizinha. A ortografia nada diz sobre a anatomia — essa verificação continua sendo manual, e o nosso protocolo de verificação existe exatamente para essa lacuna.
  2. Rótulos longos encolhem. "TUMOR ANTIGEN ARM" em escala de prancha pode sair menor do que o legível. A correção está no prompt: encurte o rótulo, ou nomeie o painel a que ele pertence.
  3. Layouts espelhados invertem os rótulos. Em anatomia, rótulos de direita/esquerda podem acabar trocados mesmo quando o desenho parece correto no mais. A correção é uma restrição nomeada: "lado direito da imagem", "pulmão esquerdo à esquerda do observador".
  4. Retentativas fazem parte do custo. Qualquer fluxo de trabalho honesto reserva orçamento para uma ou duas regenerações por prancha. Um modelo que acerta os rótulos na primeira tentativa com mais frequência é mais barato do que o preço de tabela sugere — o mesmo ponto que o nosso artigo sobre benchmarks faz sobre renderização de texto em geral.

<think>The user wants me to translate the article fragment into natural, fluent Portuguese. Let me analyze what's being asked:

This is a fragment that appears to be from an article about prompt engineering rules, specifically rules that were derived from some kind of test. The content is in a code block format.

Let me translate this carefully:

  1. "Quote every label, in capitals: Label 'JET', 'PHOTON RING'."

    • "Cite cada rótulo, em maiúsculas: Rótulo 'JET', 'PHOTON RING'."
  2. "Keep labels to 1–2 words; spell out symbols (ALPHA not α, CO2 not CO₂)."

    • "Mantenha os rótulos com 1–2 palavras; escreva os símbolos por extenso (ALPHA em vez de α, CO2 em vez de CO₂)."
  3. "Cap the set at six labels per plate."

    • "Limite o conjunto a seis rótulos por prancha/prato."
  4. "Name the panel each label lives in when placement matters."

    • "Nomeie o painel em que cada rótulo se encontra quando o posicionamento importa."
  5. "Fix errors by naming them: 'the sill is horizontal, between two strata'."

    • "Corrija erros nomeando-os: 'a soleira é horizontal, entre duas camadas'."

Wait, let me reconsider some terms:

  • "plate" in this context (likely referring to diagrams, figures in scientific articles) - could be "prancha", "lâmina", "figura", or "painel". Given the context of scientific figures/diagrams, "prancha" is commonly used.
  • "sill" - in geology, this is "soleira" (an intrusion of rock)
  • "strata" - "camadas" or "estratos"

Actually, looking again - "panel" in context of figures is "painel", and "plate" could be "prancha" or "lâmina". Let me use consistent terminology.

The rules are presented in a code block, so I need to keep that format. The text inside the code block needs to be translated while keeping the structure.

Let me refine:

  1. "Cite cada rótulo, em maiúsculas: Rótulo 'JET', 'PHOTON RING'."
  2. "Mantenha os rótulos em 1–2 palavras; escreva os símbolos por extenso (ALPHA em vez de α, CO2 em vez de CO₂)."
  3. "Limite o conjunto a seis rótulos por prancha."
  4. "Indique o painel onde cada rótulo se encontra quando o posicionamento for relevante."
  5. "Corrija os erros nomeando-os: 'a soleira é horizontal, entre duas camadas'."

Hmm, "Cap the set at six labels per plate" - "cap" here means "limit". "plate" in scientific contexts usually refers to a diagram/figure plate. In Portuguese, "prancha" is good, or "lâmina" (in microscopy context). I'll go with "prancha" as it fits the general scientific illustration context.

Actually, since this seems to be about image/figure labeling (perhaps for diagrams of black holes or astronomy given the JET and PHOTON RING references), "prancha" works well.

Let me also consider if "plate" could be "painel" - but "panel" is already used separately in rule 4. So they might be different things. "Plate" = "prancha" and "panel" = "painel" makes sense.

Let me finalize my translation:

1. Cite cada rótulo, em maiúsculas: Rótulo "JET", "PHOTON RING".
2. Mantenha os rótulos em 1–2 palavras; escreva os símbolos por extenso (ALPHA em vez de α, CO2 em vez de CO₂).
3. Limite o conjunto a seis rótulos por prancha.
4. Indique o painel em que cada rótulo se encontra quando o posicionamento for relevante.
5. Corrija os erros nomeando-os: "a soleira é horizontal, entre duas camadas".

The instruction says: keep all markdown syntax intact, keep every URL/HTML tag as-is, keep code blocks untranslated (translate only comments inside them when trivially safe).

This is a code block containing the content to translate. The content inside is the "code" - which in this case appears to be plain text rules, not actual code. Since it's inside a code block, I should translate the content inside while keeping the code block structure intact.

Wait - the instruction says "Keep code blocks and inline code untranslated (translate only comments inside them when trivially safe)."

This is tricky. The content is in a code block but it's clearly meant to be translated content (a list of rules), not actual code. The instructions seem to be saying:

  • Keep actual code (programming code) untranslated
  • Translate comments inside code if trivially safe

But this isn't programming code - it's a list of prompt rules formatted as code. I think the intent is that since it's in a code block format (perhaps to preserve formatting/indentation), the content inside still needs to be translated.

Actually, looking more carefully at the rules - rule 1 itself says "Quote every label, in capitals" - this is about labeling convention. The content here represents prompt rules for some image labeling task.

I think the best interpretation is: translate the text content inside the code block, since it's clearly natural language text that needs translation, not programming code. The code block is just the formatting choice of the article.

Let me proceed with my translation. I should also note that "PHOTON RING" is a specific astronomical term - in Portuguese it would be "anel de fótons" but since these are labels meant to be in capitals and likely proper nouns/specific terms, I'll keep "PHOTON RING" as-is or could write "PHOTON RING" (the user wants labels kept as is).

Actually, the rule says "Quote every label" - so the labels themselves are quoted examples and should probably stay in English/original form since they're quoted. Let me keep "JET" and "PHOTON RING" as-is since they appear to be quoted labels being used as examples.

Same with "ALPHA" and "CO2" - these are the spelled-out versions mentioned, so they should stay as labels.

Let me finalize:

1. Cite cada rótulo, em maiúsculas: Rótulo "JET", "PHOTON RING".
2. Mantenha os rótulos em 1–2 palavras; escreva os símbolos por extenso (ALPHA em vez de α, CO2 em vez de CO₂).
3. Limite o conjunto a seis rótulos por prancha.
4. Indique o painel em que cada rótulo se encontra quando o posicionamento for relevante.
5. Corrija os erros nomeando-os: "a soleira é horizontal, entre duas camadas".

Wait, I should reconsider "sill" and "strata". "Sill" in geology is indeed "soleira". "Strata" is "estratos" or "camadas". Both are valid. "Camadas" is more accessible/general, "estratos" is more technical. Since the article seems to mix technical and accessible language, I'll use "camadas" for readability but "estratos" could work too. Let me use "camadas" as it's more universally understood.

Actually, looking at the geological context (sill is a specific geological feature), using "estratos" might be more consistent with the technical nature. But "camadas" works fine too. I'll stick with "camadas" for naturalness.

Let me also check the formatting of the example quote in rule 5: "the sill is horizontal, between two strata" - this is an example of error correction. Should the quote style match the original? In the original, it uses double quotes. I'll keep double quotes in my translation.

Final answer:

1. Cite cada rótulo, em maiúsculas: Rótulo "JET", "PHOTON RING".
2. Mantenha os rótulos em 1–2 palavras; escreva os símbolos por extenso (ALPHA em vez de α, CO2 em vez de CO₂).
3. Limite o conjunto a seis rótulos por prancha.
4. Indique o painel em que cada rótulo se encontra quando o posicionamento for relevante.
5. Corrija os erros nomeando-os: "a soleira é horizontal, entre duas camadas".

This looks good and natural in Portuguese while preserving the structure and important terms.

One more consideration: "Quote every label" - in Portuguese, "Cite" works but "Entre aspas" or "Coloque entre aspas" could also work. "Cite" is concise and natural. I'll keep "Cite".

Also for "Cap the set at six labels per plate" - "Cap" here means to limit/place a ceiling. "Limite o conjunto a seis rótulos por prancha" works well. Alternative: "Restrinja o conjunto a seis rótulos por prancha".

I'll go with my translation. It's natural and faithful to the original.</think>

1. Cite cada rótulo, em maiúsculas: Rótulo "JET", "PHOTON RING".
2. Mantenha os rótulos em 1–2 palavras; escreva os símbolos por extenso (ALPHA em vez de α, CO2 em vez de CO₂).
3. Limite o conjunto a seis rótulos por prancha.
4. Indique o painel em que cada rótulo se encontra quando o posicionamento for relevante.
5. Corrija os erros nomeando-os: "a soleira é horizontal, entre duas camadas".

<think>The user wants me to translate a fragment of an article into Portuguese. Let me carefully translate the checklist while maintaining the markdown formatting.

The fragment is a scientific review checklist with:

  • A heading "Scientific review checklist"
  • A bulleted list with checkbox items

Let me translate it naturally into Portuguese while keeping all markdown syntax intact.</think>

Lista de verificação para revisão científica

  • [ ] Todos os rótulos estão escritos corretamente?
  • [ ] Cada rótulo aponta para a estrutura correta (e não apenas é lido corretamente)?
  • [ ] A quantidade de rótulos está completa?
  • [ ] As posições e proporções estão coerentes em relação à sua referência?
  • [ ] As setas indicam a direção correta?
  • [ ] Foi verificado em uma fonte confiável?

<think>The user wants me to translate an article fragment from English to Portuguese. The fragment is about AI rendering scientific labels. I need to:

  1. Keep all markdown syntax intact
  2. Keep URLs, image paths, and HTML tags as-is
  3. Keep code blocks untranslated
  4. Keep "GLMImage" untranslated
  5. Use native, fluent Portuguese
  6. Output only the translated markdown

Let me translate this fragment into natural, fluent Portuguese.</think>

Conclusão

A IA consegue renderizar rótulos científicos corretamente? Para o formato que este site ensina — rótulos curtos, entre aspas, com letras maiúsculas em pranchas estruturadas — a resposta destas vinte pranchas é: de forma confiável o suficiente para servir de base, mas não o bastante para dispensar a revisão. A ortografia está quase resolvida nesse comprimento de rótulo; o posicionamento e a contagem funcionam bem com conjuntos pequenos; e a etapa de verificação continua sendo humana. Essa combinação é o que torna o rascunho por IA válido: a parte cara (figuras corretas, legíveis e publicáveis) ainda passa pelos seus olhos, mas a parte barata deixou de ser lenta.

<think>The user wants me to translate an FAQ fragment from English to Portuguese. I need to maintain all markdown syntax, keep URLs intact, keep "GLMImage" untranslated, and produce natural fluent Portuguese. Let me work through this carefully.

Key elements to preserve:

  • The markdown structure (headings, bold, italic, links)
  • URLs exactly as-is (/blog/verify-ai-scientific-illustration)
  • The brand "GLMImage" untranslated
  • "grok-imagine-image-2.0" - this is a model name, should stay untranslated
  • "glmimage.app" - this is a URL/domain, should stay untranslated
  • Code/technical terms stay

Let me translate:

Perguntas Frequentes (FAQ)

Qual modelo renderizou as pranchas de teste?
grok-imagine-image-2.0, através do módulo de ilustração científica do glmimage.app — o mesmo pipeline por trás de cada prancha deste site. Nós o nomeamos porque a atribuição do modelo faz parte de um teste honesto; um teste de rótulos sem o modelo nomeado é marketing, não medição.

A precisão dos rótulos varia por disciplina?
Em nossas pranchas, as falhas se agrupam por tipo de rótulo, não por disciplina: palavras longas, símbolos e subscritos falham em todos os lugares; termos curtos em letras maiúsculas têm sucesso em todos os lugares. Uma prancha misturando "JET" e "TUMOR ANTIGEN ARM" herda o risco do segundo rótulo.

E quanto a rótulos não ingleses?
Trate-os como um teste separado. Renderização bilíngue é uma capacidade específica — é o grande destaque dos próprios modelos de imagem da GLM e a razão pela qual este site existe — mas qualquer modelo que você use merece um pequeno piloto: dez rótulos no seu idioma antes de você comprometer um conjunto de figuras a ele.

Se a ortografia passar, a figura está pronta para publicação?
Não — a ortografia é necessária, mas não suficiente. Um rótulo pode estar perfeitamente escrito e apontar para a estrutura errada. A prontidão para publicação passa por todo o protocolo de verificação, do qual este teste é apenas uma entrada.

Let me review for naturalness and fluency:

  • "perguntas frequentes" - this is the standard Portuguese for FAQ ✓
  • "renderizou as pranchas de teste" - "pranchas" is a good translation for "plates" in scientific illustration context ✓
  • "pipeline" - commonly used in Portuguese technical contexts ✓
  • "atribuição do modelo" - good for "model attribution" ✓
  • "rótulo" vs "etiqueta" - both work, "rótulo" is more common in scientific illustration ✓
  • "letras maiúsculas" - good for "all-caps" ✓
  • "pronto para publicação" - good for "publication-ready" ✓
  • "protocolo de verificação" - good for "verification protocol" ✓

I should keep the all-caps labels "JET" and "TUMOR ANTIGEN ARM" as they are - they're examples of labels being tested.

Let me also double-check the URL stays exactly as is: /blog/verify-ai-scientific-illustration ✓

The italics around label type should be preserved.

I think the translation is solid. Let me finalize.</think>

Perguntas Frequentes

Qual modelo renderizou as pranchas de teste?

grok-imagine-image-2.0, através do módulo de ilustração científica do glmimage.app — o mesmo pipeline por trás de cada prancha deste site. Nós o nomeamos porque a atribuição do modelo faz parte de um teste honesto; um teste de rótulos sem o modelo nomeado é marketing, não medição.

A precisão dos rótulos varia conforme a disciplina?

Em nossas pranchas, as falhas se agrupam por tipo de rótulo, não por disciplina: palavras longas, símbolos e subscritos falham em todos os contextos; termos curtos em letras maiúsculas acertam em todos os contextos. Uma prancha que mistura "JET" e "TUMOR ANTIGEN ARM" herda o risco do segundo rótulo.

E quanto a rótulos em outros idiomas?

Trate-os como um teste à parte. A renderização bilíngue é uma capacidade específica — é o grande destaque dos próprios modelos de imagem da GLM e a razão de existir deste site — mas qualquer modelo que você utilize merece um pequeno piloto: dez rótulos no seu idioma antes de você comprometer um conjunto de figuras a ele.

Se a ortografia passar, a figura está pronta para publicação?

Não — a ortografia é necessária, mas não suficiente. Um rótulo pode estar perfeitamente escrito e ainda assim apontar para a estrutura errada. A prontidão para publicação exige a execução completa do protocolo de verificação, do qual este teste é apenas uma das entradas.

<think>The user wants me to translate an article fragment into Portuguese. Let me carefully translate it while keeping all markdown syntax, URLs, and brand names intact.

Key points:

  • Keep all markdown syntax
  • Keep URLs exactly as-is
  • Keep "GLMImage" untranslated
  • Natural, fluent Portuguese
  • Translate the whole fragment completely

Let me translate this carefully:

"Reproducing this test on your own stack" → "Reproduzindo este teste na sua própria stack"

The rest needs natural Portuguese expression while maintaining technical accuracy.

"stack" in Portuguese tech context is often kept as "stack" (it's commonly used in Portuguese tech writing).

Let me work through the full translation:</think>

Reproduzindo este teste na sua própria stack

O teste é deliberadamente reproduzível: vinte prompts a partir de uma única gramática (nosso template), rótulos limitados a seis, de uma a duas palavras, todos em maiúsculas, entre aspas; em seguida, uma leitura por prancha para verificar ortografia, posicionamento e contagem; depois, a contagem por tipo de rótulo — palavras curtas, palavras longas, símbolos, subscritos. Execute-o em qualquer modelo e cadeia de ferramentas que você realmente utilize, pois o comportamento dos rótulos é específico de cada modelo, e a sua stack é a única que importa para as suas figuras. Se a sua contagem corresponder ao padrão acima (rótulos curtos entre aspas são confiáveis; comprimento e símbolos se degradam), as regras deste artigo se transferem diretamente; se não corresponder, a sua contagem é o manual de regras — escreva os prompts que os seus números sustentam e mantenha o teste no diário de figuras (o hábito do artefato).

William Hayes

William Hayes

A IA Consegue Renderizar Rótulos Científicos Corretamente? Um Teste com 20 Pranchas | Blog