
How do you create images with Super Intelligence? You translate a visual objective into a specification: subject, composition, viewpoint, environment, lighting, colour, style, text, aspect ratio and the details that must remain consistent. Then you generate, inspect, edit and verify the image instead of treating the first result as final.
Image creation with SI is both a creative and a control problem. Generative systems can produce rich visual options quickly, but they can also misread spatial relationships, labels, counts, text, diagrams or exact brand details. Strong visual workflows separate artistic freedom from factual requirements.
This eduKateSG guide explains how to create, prompt, edit and evaluate AI images for education, websites, presentations, social content, diagrams and creative projects. It follows How to Brainstorm With Super Intelligence in Stage 5 of the How to Learn Super Intelligence Quickly curriculum.
Terminology: SI is our editorial term for practical contemporary AI learning. Image-generation features change over time and vary by tool. Current OpenAI documentation for Images in ChatGPT describes creation and editing, text inside images, transparent backgrounds, aspect-ratio choices and other image workflows; always check the current documentation for the tool you use.
The First Principle: Describe the Visual Job, Not Only the Subject
“Create a school image” names a subject but not a job. Is the image a website header, worksheet diagram, social post, poster, presentation background or conceptual illustration?
The job determines composition, text space, detail level, aspect ratio and what must be accurate.
Start with the receiver and placement before aesthetic style.
Step 1 — Define the Image Purpose
State what the image should accomplish: explain, attract attention, establish mood, illustrate a concept, show a process or provide a reusable asset.
An explanatory diagram and a decorative header should not be prompted the same way.
Step 2 — Define the Receiver
A Primary student, Secondary student, parent, technical audience and public social-media user have different visual needs.
Age, domain knowledge and viewing context influence complexity, labels and visual metaphor.
Step 3 — Define the Canvas
Choose orientation and aspect ratio according to use. A website hero may need a wide composition; a social tile may be square; a worksheet illustration may need vertical or transparent space.
Leave intentional negative space when text or interface elements will be added later.
Step 4 — Define the Subject
Name the main subject precisely: three Secondary students reviewing mathematics notes, a simplified electric circuit, a futuristic library interior, a labelled cell diagram.
Specify attributes that matter and omit unnecessary details that only constrain creativity.
Step 5 — Define Composition
Composition tells the system where elements belong and how the viewer should read the image.
Useful instructions include foreground, background, centre, left/right placement, visual hierarchy, symmetry, depth and empty space.
Step 6 — Define Viewpoint and Lens Feel
Point of view changes the story. Eye-level feels direct; overhead can clarify layout; close-up emphasises detail; wide shots establish environment.
You can describe cinematic or photographic qualities without needing exact camera specifications when those details do not matter.
Step 7 — Define Lighting
Lighting affects mood, readability and realism. Soft daylight, classroom fluorescent light, dramatic rim light or flat diagram lighting create different results.
For instructional visuals, clarity often matters more than dramatic mood.
Step 8 — Define Colour
Use colour to support purpose. Brand palettes, subject conventions or accessibility requirements may constrain colour choices.
Avoid relying on colour alone to distinguish important categories in diagrams.
Step 9 — Define Style
Style can be photographic, editorial illustration, flat vector, isometric, watercolour, 3D render, technical diagram or another visual language.
Describe observable characteristics rather than only naming a style label: clean geometric shapes, muted palette, fine ink lines, realistic texture.
Step 10 — Define Text Requirements
If the image must contain text, provide the exact wording and where it should appear. Keep text short where possible.
Inspect spelling, line breaks and label placement. Even strong image generators can produce visual text that requires correction.
Step 11 — Define Invariants
List details that must not change during iteration: number of people, logo placement, diagram labels, product shape, character identity or required colour.
These invariants form the preservation set for image editing.
Step 12 — Define Creative Freedom
Also state what may vary: background detail, clothing colour, lighting nuance, props or camera angle.
A good prompt protects essentials while leaving enough room for visual invention.
The Visual Prompt Canvas
- Purpose.
- Receiver.
- Placement.
- Aspect ratio.
- Main subject.
- Composition.
- Viewpoint.
- Environment.
- Lighting.
- Colour.
- Style.
- Exact text.
- Required details.
- Forbidden details.
- Negative space.
- Editing invariants.
- Verification checks.
Prompting for Photorealistic Images
Describe subject, environment, viewpoint, lighting, materials and mood. Use specific visual observations rather than abstract praise such as “beautiful”.
If authenticity matters, avoid impossible combinations of lighting, perspective or physical objects unless the image is intentionally surreal.
Prompting for Illustrations
Define line quality, shape language, palette, texture and level of simplification.
For educational illustration, make the learning target more important than decorative complexity.
Prompting for Diagrams
Start from the relationships that must be correct. List nodes, labels, arrows and direction before choosing style.
For fact-sensitive diagrams, verify labels and connections independently. If precision is critical, a deterministic diagramming tool may be better than free-form generation.
Prompting for Infographics
Infographics combine data, hierarchy and visual explanation. Use a verified data source and decide what the viewer should understand in seconds.
Generated icons and layouts can help, but exact charts and numerical labels should be created or verified through data-driven tools.
Prompting for Website Headers
A header should leave space for title or interface elements and avoid important faces or details under expected overlays.
Define the crop and focal point. Test desktop and mobile layouts if the page will be responsive.
Prompting for Social Images
Social content needs fast visual hierarchy. Define one dominant subject and one clear message.
Avoid tiny text and overpacked compositions. Verify platform dimensions when they matter.
Prompting for Presentation Images
Presentation visuals should support the slide argument rather than compete with it.
Use clean composition and space for slide text. Keep visual metaphor connected to the actual concept.
Prompting for Worksheet Images
Worksheets benefit from high contrast, simple shapes and limited visual clutter.
If the image forms part of a question, verify that generated details do not accidentally reveal or change the answer.
Prompting for Concept Art
Concept art can explore atmosphere, form, environment and design direction before precision matters.
Generate several visual territories rather than polishing one result too early.
Prompting for Characters
Define character role, age range, clothing, posture, expression and recurring identifying features.
For consistency across several images, keep a character sheet or reference image and preserve identity constraints during editing.
Prompting for Environments
Describe spatial layout, materials, time of day, activity and atmosphere.
If the environment is a real place, distinguish creative interpretation from documentary accuracy.
Prompting With Reference Images
Reference images can communicate composition, subject or visual identity more directly than text. Clarify what should be preserved and what should change.
Do not assume every tool treats reference images the same way; check current capabilities and permissions.
Editing Existing Images With SI
Image editing should use the same preservation logic as text editing. Define the region to change and the features to keep.
Current OpenAI help documentation describes both selection-based and conversational edits in ChatGPT Images. Regardless of interface, review surrounding areas because edits can extend beyond the intended region.
Edit Type 1 — Remove an Object
Name the object and describe what should plausibly appear behind it.
Check shadows, reflections and geometry after removal.
Edit Type 2 — Add an Object
Specify placement, scale, perspective and lighting relation to the scene.
The new object should belong physically to the image rather than look pasted in.
Edit Type 3 — Change Colour
Identify the exact object and colour target. Preserve lighting variation so the result remains natural.
For brand colours, compare against the required colour reference.
Edit Type 4 — Change Background
Protect subject edges, hair, transparency and contact shadows.
A new background may change the implied lighting; decide whether the subject lighting should also adapt.
Edit Type 5 — Change Style
Style transfer can alter details as well as surface appearance. Recheck identity, text, geometry and required objects after transformation.
Edit Type 6 — Expand the Canvas
When extending an image, define what should continue into the new area and where the original focal point should remain.
Canvas expansion is useful for adapting one asset to multiple layouts.
Edit Type 7 — Correct Text
Provide exact replacement text and location. Check every character afterward.
For large amounts of text, consider adding typography in a design tool rather than relying on generated text.
Edit Type 8 — Create Transparency
Transparent backgrounds are useful for cut-out assets, diagrams and compositing.
Inspect edges and semi-transparent details such as hair, glass or shadows.
The Image Iteration Loop
- Generate a broad first version.
- Inspect composition and purpose.
- Check required details.
- Identify the single highest-value change.
- Edit that change.
- Recheck invariants.
- Repeat until the image meets the receiver’s job.
- Export or place in the real layout and review again.
One targeted change per iteration often produces more controllable results than replacing the whole prompt after every issue.
Visual Diffing
Compare before and after images for intended and unintended changes. Check subject identity, count, pose, labels, background, colour and crop.
A visual edit can fix one object while changing another unnoticed detail.
Factual Image Verification
If an image depicts a scientific, historical, geographic or technical fact, treat it like factual prose. Verify the details.
A convincing render of an incorrect circuit or anatomical label is still misinformation.
Counting Problems
When exact counts matter, inspect them manually. Generative images can include extra or missing repeated objects.
Use deterministic graphics when count accuracy is non-negotiable.
Spatial Relationship Problems
Check left/right, inside/outside, direction of arrows, relative scale and physical support.
Visual plausibility can hide spatial errors that matter in diagrams.
Text Problems
Check spelling, punctuation and whether text is attached to the correct object.
For important text, separate image generation from final typography when practical.
Identity Consistency
Characters or products can drift across iterations. Use references and preserve defining features.
After every edit, compare identity-critical details rather than assuming the model remembered them.
Brand Consistency
Maintain logo, palette, typography and visual tone through a brand reference pack.
Do not let SI invent or alter logos when exact brand identity is required.
Accessibility in Images
Use sufficient contrast, readable labels and alternative text. Avoid conveying a required distinction only through colour.
Decorative images can have concise alt text; instructional visuals need descriptions that convey the learning content.
Privacy and Likeness
Be careful with identifiable people, private images and sensitive contexts. Use appropriate consent and permissions.
For practice or generic educational imagery, fictional subjects can reduce unnecessary privacy exposure.
Copyright and Source Discipline
Use tools and source material according to applicable rights, licences and product terms. Do not assume that an available image is free to reuse or transform.
For public publishing, keep track of the origin of reference assets when rights matter.
A Worked Example: Secondary Education Header
Purpose: website article header. Receiver: parents and students. Composition: three Secondary students around open books, natural classroom environment, clear faces, no embedded text, enough negative space for page title.
Generate several versions. Select one with the right tone and crop, then check student count, hands, book geometry and background distractions.
A Worked Example: Science Diagram
Purpose: explain series circuit. Required elements: cell, switch, two lamps in one loop, arrows showing conventional current, clear labels.
Generate or draft visually, then verify every connection. If exact circuit topology matters, recreate the final diagram in a deterministic vector tool.
A Worked Example: Presentation Concept Image
Topic: bottlenecks in systems. Visual concept: narrow bridge causing traffic queue, clean editorial illustration, space on left for title.
The metaphor supports the concept without pretending to be empirical data.
A Worked Example: Social Campaign Image
Goal: promote study-planning article. Use one dominant visual idea, short optional text and high contrast.
Check that the image still reads when viewed small on a phone.
A Worked Example: Character Series
Create a character sheet first with age, hairstyle, clothing palette and defining features. Use it as reference for later scenes.
Test consistency across three different poses before building a large series.
A Worked Example: Image Edit
Original classroom photo-like illustration contains a distracting poster. Edit only the poster area while preserving students, lighting and composition.
Compare the result with the baseline to confirm no identity or hand details changed.
Image Failure 1 — Prompt Overload
Too many aesthetic and content constraints conflict.
Prioritise purpose and invariants; simplify secondary style details.
Failure 2 — Vague Composition
The right objects appear but visual hierarchy is wrong.
Specify placement, focal point and negative space.
Failure 3 — Text Errors
The image looks good but labels are misspelled.
Correct text separately or add typography outside the generator.
Failure 4 — Wrong Count
Repeated objects appear inconsistently.
Check manually or use a deterministic method when exact count matters.
Failure 5 — Edit Drift
One local edit changes unrelated areas.
Compare before and after and restore from baseline if necessary.
Failure 6 — Style Over Accuracy
A beautiful scientific or historical image contains incorrect details.
Verify factual content independently.
Failure 7 — Wrong Aspect Ratio
The image cannot fit the real placement without destructive cropping.
Design for the target canvas from the beginning or expand carefully.
Failure 8 — No Receiver Test
The image is aesthetically pleasing but fails at actual size or context.
Place it in the website, slide or worksheet before final acceptance.
The Image Quality Audit
- Purpose fit.
- Receiver fit.
- Subject accuracy.
- Composition.
- Aspect ratio.
- Required details.
- Text accuracy.
- Spatial accuracy.
- Identity consistency.
- Colour and contrast.
- Accessibility.
- Factual verification.
- Rights and permissions.
- Real-layout test.
A Practice Lab: One Subject, Five Compositions
Choose one subject and keep style constant. Generate eye-level, overhead, close-up, wide environmental and asymmetric negative-space compositions.
Compare which composition best serves the final placement.
A Practice Lab: One Composition, Five Styles
Keep subject and layout constant while varying style: photographic, flat vector, editorial, watercolour and technical.
Observe how style changes interpretation and suitability for the receiver.
A Practice Lab: Controlled Editing
Generate one image, then make three separate edits: background, one object and colour.
After each edit, check the preservation set. This trains visual change control.
Frequently Asked Questions
How detailed should an image prompt be?
Detailed enough to define purpose, subject, composition and invariants. Add style detail only where it changes the result.
Should I generate many versions?
Use enough variants to explore composition or visual direction, then converge. Endless generation can delay real evaluation.
Can I use generated images for diagrams?
For conceptual illustrations, often yes. For exact technical diagrams, verify every relationship and consider deterministic tools for the final version.
Can SI edit an existing image?
Many current image systems can. For example, OpenAI’s current Images documentation describes uploading or selecting an image and requesting edits. Check the current tool’s capabilities and permissions.
How do I keep a character consistent?
Use a reference image or character sheet, define identity invariants and recheck them after each edit.
What comes next?
Continue Stage 5 with presentations, documents and spreadsheets through the complete SI learning hub.
The Visual Production System
A serious image workflow needs more than a prompt. It needs a visual brief, reference assets, iteration history, preservation rules, output specifications and a final placement test.
For one-off creative exploration, this system can remain lightweight. For a website, brand, character series or educational library, it becomes the difference between random attractive images and a coherent visual system.
The production state should record what is approved and what remains exploratory, just as a writing workflow distinguishes drafts from final copy.
The Visual Brief
- Purpose and receiver.
- Placement and aspect ratio.
- Subject.
- Composition.
- Visual hierarchy.
- Reference assets.
- Style characteristics.
- Colour and lighting.
- Required text.
- Required factual details.
- Preservation invariants.
- Forbidden or unwanted elements.
- Output format.
- Accessibility needs.
- Final verification.
A good brief gives the generator enough direction without specifying decorative details that do not affect the receiver’s job.
Reference Packs
A reference pack can contain approved colours, logos, character sheets, product photos, visual examples and layout guidance.
Label the role of each reference. One image may define composition, another style, another subject identity. Do not assume the system knows which property you want transferred.
Keep rights and permissions for reference assets clear, especially in public or commercial work.
Character Sheets
A character sheet defines identity across several images: approximate age, face shape, hair, clothing, accessories, palette, posture and recurring features.
Use front, side or varied-expression references when consistency matters. Keep the sheet stable while scene, pose and environment change.
After every generation, compare identity-critical features before accepting the image into the series.
Brand Visual Systems
Brand consistency needs more than a colour prompt. Define subject treatment, composition, texture, lighting, typography handling, spacing and emotional tone.
Create several approved exemplars and a short description of why they belong together.
SI can generate within the system, but final selection should protect brand identity rather than model preference.
Visual Grammar
Visual grammar is the recurring way a series communicates. It may use centred subjects, generous negative space, natural classroom lighting and realistic materials; another series may use flat geometry and strong iconography.
Once the grammar is explicit, new images can vary while still feeling related.
A grammar should be flexible enough to handle different subjects without forcing identical compositions.
Composition Systems
For a visual series, define a small set of useful compositions: hero-wide, three-quarter portrait, overhead workspace, centred diagram, left-subject/right-space and close-up detail.
This makes image generation more predictable and reduces random layout exploration for every new asset.
Choose composition based on placement rather than aesthetic habit.
Negative Space as a Design Variable
Negative space is not empty failure. It creates room for text, interface controls, cropping and visual rest.
Specify where the empty space should appear. “Three students on the right half, clean classroom wall on the left for headline overlay” is operationally useful.
Test the real overlay before final acceptance.
Foreground, Midground and Background
Depth can be controlled by assigning elements to planes. Foreground may contain books, midground students, background classroom.
This improves scene readability and lets you control where detail should concentrate.
For small web images, avoid background complexity that disappears or becomes visual noise.
Lighting Consistency Across a Series
Lighting strongly affects whether a set feels coherent. Decide on daylight, warm indoor, high-key studio or another recurring treatment.
When editing one image, check whether the new background or object still belongs under the same lighting conditions.
Mismatched shadow direction is a common sign of weak visual integration.
Colour Consistency Across a Series
Use a palette with primary, secondary and accent roles rather than one vague colour description.
For educational categories, colour can help organise content, but keep meaning available through labels or shape as well.
Check output on different screens when colour accuracy matters operationally.
Typography Strategy
AI image systems can produce text, but typography-heavy designs still benefit from a separate design pass because exact spelling, hierarchy and alignment are easier to control deterministically.
Use generation for background, illustration or rough layout, then add final titles and body copy in a design or publishing tool when precision matters.
If text is generated directly, verify every character before use.
Icons and Symbol Sets
When generating an icon family, define stroke weight, corner radius, fill style, viewpoint, palette and canvas size.
Generate a small core set first and test whether they feel consistent side by side.
A single attractive icon proves little about the coherence of a full library.
Multi-Image Sequences
A sequence needs continuity: same character, environment, palette and progression. Create a storyboard before generating individual frames.
Label each frame’s purpose and what changes from the previous frame. Keep identity and setting invariants explicit.
Review the sequence together, not only each image individually.
Storyboards
Storyboards are low-cost tools for testing visual narrative before high-detail generation.
Use simple frames to validate shot order, composition, character movement and information flow.
Only polish after the sequence works. This prevents expensive iteration on the wrong narrative.
Mood Boards
Mood boards explore visual territory: colour, material, atmosphere, composition and references.
Do not use them as factual source material. Their job is direction and vocabulary.
Convert the final mood board into observable style rules before production.
Sketch-to-Image Workflows
A sketch can communicate spatial intent faster than prose. Current OpenAI image documentation describes a mobile sketch workflow that can be combined with written instructions.
Whether using that tool or another system, keep the sketch’s role explicit: layout, subject position or shape. State which details the generator may reinterpret.
Sketch-to-image is especially useful when composition is easier to draw than describe.
Template-Based Image Workflows
Templates can accelerate repeatable formats such as posters, tiles or logos. Current OpenAI documentation describes image templates as one way to begin generation in supported contexts.
Treat templates as starting structures. Replace sample content deliberately and verify that inherited elements do not conflict with the new purpose.
A template should reduce setup, not force every asset into the same visual solution.
Image-to-Image Variation
Use an accepted image as a reference and change one dimension at a time: environment, clothing, lighting, crop or style.
One-variable iteration helps preserve identity and reveals which change improves the result.
Large simultaneous transformations make it difficult to diagnose drift.
Controlled Restyling
When restyling, separate content invariants from visual treatment. A diagram’s relationships, a product’s geometry or a character’s identity should survive the style change.
After restyling, rerun the factual or identity check.
Style is a surface transformation only when the underlying content remains intact.
Background Replacement
A background edit should preserve edge quality, contact shadows, reflections and scene lighting.
The new environment should match subject scale and perspective.
For public figures, products or documentary-looking scenes, avoid creating a misleading impression of a real event unless clearly framed as synthetic or illustrative.
Object Replacement
When replacing an object, specify the new object’s size, orientation and physical relation to the scene.
Check hands, grips, shadows and occlusion. These local relationships often reveal edit problems.
If the object carries factual meaning, verify its correct form.
Pose Changes
Pose editing can change anatomy, clothing and identity. Preserve character references and review hands, limbs and facial consistency.
For instructional movement diagrams, free-form image generation may be insufficient where exact biomechanics matter.
Use specialist or deterministic illustration methods when precision is required.
Facial Expression Changes
Expression edits can subtly change perceived identity or age. Compare the result with the approved reference.
Define the intended emotion through observable cues such as gentle smile, attentive gaze or neutral concentration rather than exaggerated labels.
Keep emotional portrayal appropriate to the context.
Product Visualisation
For concept products, image generation can explore form and environment. For an existing product, exact geometry, ports, controls and brand details may require real product photography or 3D assets.
Label concept imagery clearly when it does not depict the actual product.
Do not let attractive visualisation create false expectations about features.
Architecture and Interior Visualisation
Generative imagery can explore mood, materials and layout concepts, but spatial dimensions may be inconsistent.
Use real plans and deterministic tools for construction or safety decisions.
Concept renders are visual hypotheses, not engineering documents.
Educational Diagrams
Start with a fact specification before visual style. Define every required label, relationship, arrow and exception.
Generate or illustrate, then compare against the specification and a trusted source.
When correctness matters more than artistry, rebuild the final asset in vector or diagram software after using SI for ideation.
Process Diagrams
A process diagram needs order, branching and state clarity. Write the process as structured steps first.
Then create the visual. Verify every arrow and label after generation.
Do not infer process correctness from visual neatness.
Charts and Data Visualisation
Charts should be generated from verified data using tools that respect exact values. SI can help choose chart types, write labels and explain patterns.
If an image generator is used for a stylised chart concept, do not treat it as a data-accurate final visual unless every value has been reconstructed and checked.
Data integrity outranks decorative style.
Maps
Map-like illustrations can be useful for conceptual geography or itineraries, but exact routes and boundaries require authoritative geographic data.
Do not rely on a free-form generated map for navigation or precise location.
Use a mapping tool for exact spatial information and SI for supporting explanation or visual concept.
Historical Illustrations
Historical scenes require source discipline when presented as educational material. Clothing, architecture, objects and context should be researched.
A generated reconstruction should be labelled appropriately when it is interpretive rather than documentary.
Visual realism does not turn reconstruction into evidence.
Scientific Illustration
Scientific images should begin with a verified content specification. Anatomy, molecular structure, circuit topology and scale relationships can be wrong despite professional-looking rendering.
Use subject experts or authoritative references for validation.
For high-stakes scientific communication, specialist illustration may be more appropriate.
Visual Metaphors
Metaphors can explain abstract ideas without claiming literal reality. A bottleneck represented as traffic at a narrow bridge can communicate constrained throughput.
The metaphor should clarify the mechanism rather than distort it. Ask where the analogy breaks.
Metaphor images work well in presentations and educational articles when clearly conceptual.
Hero Images Versus Evidence Images
A hero image establishes context or emotion. An evidence image communicates facts: chart, screenshot, diagram, document or photograph of an actual object.
Do not substitute a generated hero for evidence. Keep the role clear to the reader.
This distinction is especially important in journalism, research and technical content.
Decorative Versus Functional Images
Decorative images can be judged mainly by composition, brand fit and accessibility. Functional images must also pass task-specific correctness checks.
Classify the image before review so the team knows which quality dimensions matter.
Not every asset needs the same verification burden.
Visual Asset Naming
Name files consistently by series, topic, version and purpose when managing many assets.
Avoid names such as final-final2.png. Stable naming reduces accidental use of the wrong image.
For WordPress libraries, titles and alternative text should help future maintainers identify the asset.
Visual Asset Versioning
Keep baseline and accepted edits when the asset is important. A later edit may look better while losing a required detail.
Versioning is particularly useful for brand images, diagrams and recurring characters.
Do not preserve every exploratory generation indefinitely; archive selectively.
Visual Asset Registry
A registry can contain asset ID, page, purpose, source or generation state, rights note, aspect ratio, featured/header use and last review.
This is useful when a site contains thousands of images or when the same asset appears in several pages.
The registry turns media management into an operational system.
Image Provenance
Record whether an image is original photography, generated, edited, licensed, public-domain or another source category where provenance matters.
Provenance helps with rights, maintenance and appropriate disclosure.
Do not rely on filename alone to remember origin months later.
Visual Disclosure
Whether disclosure is needed depends on context, policy and the risk of misleading the viewer. Generated concept art and decorative illustration differ from synthetic imagery that could be mistaken for documentary evidence.
Use clear labelling where a reasonable viewer could be misled about whether an event, person or product depiction is real.
Follow applicable platform, organisational and legal requirements.
Visual Consent
When using identifiable people or private images, obtain appropriate permission and consider whether the transformation could create a misleading or harmful portrayal.
For generic educational assets, fictional or model-generated characters can avoid unnecessary use of real student identities.
Consent is part of visual workflow design, not an afterthought.
Visual Accessibility
Write alternative text according to function. Decorative images need little or no descriptive burden in contexts that support decorative marking; instructional images need the information necessary to understand the concept.
For complex diagrams, consider a longer textual description nearby.
Test colour contrast and do not encode categories through colour alone.
Image File Formats
Choose format based on use. Raster formats are common for photographic images; transparent assets may need formats supporting alpha; vector formats are useful for scalable deterministic graphics.
Do not force one format for every asset. Consider quality, transparency, file size and publishing compatibility.
The final publishing system may create additional responsive sizes automatically.
Image Optimisation
Large files slow pages. Resize and compress according to placement while preserving enough quality for the display.
Do not use a massive source image where the page shows only a small thumbnail.
Check mobile performance and image sharpness after publication.
Responsive Cropping
An image can work on desktop and fail on mobile because the subject is cropped. Keep important content away from vulnerable edges and test responsive layouts.
Alternative crops may be needed for some placements.
Negative-space planning during generation reduces this problem.
Visual SEO
Use descriptive filenames and alternative text where appropriate, but do not keyword-stuff. The image should support the page topic naturally.
The page’s substantive content remains more important than decorative media for answering the reader’s query.
Image metadata should help accessibility and maintenance first.
Visual Consistency Audits
Periodically review a series side by side. Are aspect ratios, subject treatment, quality and brand tone coherent? Are old images using outdated media families or visual standards?
Replace only where improvement is meaningful. Avoid endless visual churn that creates work without improving receiver value.
For large sites, start from latest legacy content and move backward under a controlled rule.
Visual Regression Tests
Keep representative failures: wrong count, misspelled text, identity drift, inaccurate diagram, broken crop or mismatched header image.
After changing prompts or tools, rerun these cases.
Visual workflows improve when failure lessons remain available.
The Image Production Handoff
Before publication, hand off the accepted asset with its purpose, target page, alt text, crop guidance, rights/provenance note and whether it should be featured or in-content header.
This prevents a correct image from being used in the wrong placement.
For series work, include the media family rule.
The Final Image Governance Gate
Before an important visual goes live, confirm purpose, factual accuracy, text, identity, rights, consent, accessibility, file size, crop and real-layout appearance.
For generated edits, compare with the baseline to detect unintended changes.
For technical diagrams, charts or maps, return to deterministic or authoritative data when exactness matters.
The mature SI image workflow treats generation as one production stage inside a visual system whose final authority remains with the human publisher or designer.
Contact Sheets for Fast Visual Selection
When exploring several directions, generate a controlled set and review them together as a contact sheet or visual grid. Side-by-side comparison makes composition, identity drift and style inconsistency easier to notice than reviewing images one by one.
Label versions by prompt or design direction. Choose the strongest composition first, then invest editing effort in that candidate rather than polishing every draft.
This is the visual equivalent of clustering ideas before convergence.
Prompt Cards for Reusable Visual Work
A prompt card stores the stable parts of a visual workflow: purpose, subject treatment, composition grammar, palette, lighting, aspect ratio, negative space and invariants.
When creating a new asset, change only the topic-specific fields. This keeps a series coherent while avoiding copy-paste prompts whose hidden assumptions no longer fit.
Version prompt cards when the visual standard changes, and keep one approved example beside the card.
Visual Regression Sets
For repeated generation, keep a small set of difficult cases: hands interacting with objects, several people with fixed counts, text labels, transparent edges, precise diagrams and responsive crops.
After a tool or prompt change, rerun representative cases. Improvement in one image should not be assumed to generalise to the whole workflow.
A regression set makes visual quality cumulative rather than dependent on memory.
Image Model Changes and Revalidation
Image-generation systems change over time. A prompt that produced stable results in one version may behave differently after a model update.
Current OpenAI release notes and Images in ChatGPT documentation describe ongoing changes to image creation and editing. Treat model changes as a reason to revalidate important prompt cards and reference workflows rather than assuming identical behaviour.
The durable asset is the visual specification and evaluation method, not one model’s exact response pattern.
Visual Asset Approval States
Use states such as Exploratory, Candidate, Approved and Published. This prevents attractive drafts from entering production before rights, accessibility or factual checks are complete.
For series work, only Approved assets should be assigned to live pages unless a deliberate exception is recorded.
Approval state is especially important when many images are generated faster than humans can review them.
Image Selection Criteria
- Purpose fit.
- Receiver fit.
- Composition and hierarchy.
- Required subject details.
- Factual correctness.
- Identity consistency.
- Text correctness.
- Brand or series coherence.
- Accessibility.
- Rights and consent.
- Crop resilience.
- File performance.
- Editability and future reuse.
Do not collapse these criteria into one aesthetic score. A gorgeous image with wrong labels or unusable crop can still fail its job.
Visual Asset Reuse
A strong image may serve several pages or formats if the subject and rights permit. Reuse can strengthen brand consistency and reduce production overhead.
However, avoid forcing one crop into every placement. Create approved derivatives or extensions for different aspect ratios where necessary.
Keep the canonical source asset identifiable so later edits do not begin from a degraded copy.
Derivative Management
When one source image produces several crops, transparent versions or text overlays, record the relationship. This helps future editors know which file contains the original visual information.
A derivative should not quietly become the new master if it has lost resolution or important content.
For WordPress, keep media titles and alt text useful enough that maintainers can distinguish variants.
Images in a Large Publishing Series
A long article series benefits from a media rule. Decide whether images rotate, follow subject families, use one consistent style or use a defined collection such as edukatesecondary###.
The rule should be stored outside individual memory so later batches continue the same standard. Before publication, verify both the first in-content header and the featured media record.
This protects visual continuity without touching global theme or homepage design.
Replacing Legacy Images Safely
When upgrading older articles, separate image replacement from content rewriting. Confirm the page category, current media date and replacement family before making the change.
Change featured and header image together when the publishing standard requires them to match. Do not alter page layout, menus or unrelated blocks.
After the replacement, inspect the page state to confirm the intended asset actually appears in both places.
Image Alt Text as Receiver Information
Alternative text should communicate the image’s function. A decorative classroom header may need a concise description. A diagram requires the relationships needed to understand the concept.
Do not stuff keywords into alt text. Write for a user who cannot see the image.
If nearby text already explains the image fully, the alt description can avoid unnecessary repetition.
Images and Page Performance
Visual quality includes loading behaviour. An oversized file can damage mobile experience even when the image itself is excellent.
Use appropriate dimensions and compression for the placement. Prefer responsive delivery where the publishing platform supports it.
Test the actual page rather than judging only the source file.
Images and Search Intent
A relevant header can reinforce topic context, but images do not compensate for thin article content. The page’s written explanation remains the main answer to a text-heavy search query.
Use visuals where they add recognition, explanation or brand coherence. Avoid inserting many generic images simply to increase media count.
Image strategy should serve the reader’s job first.
A Full Production Case: Four-Article Batch
Suppose a learning series publishes four new articles. The media standard requires edukatesecondary### images. Select four approved assets, record their IDs and map one to each article.
For every page, set the same asset as featured media and first image block. Add descriptive alt text. Publish the text only after the image, slug, SEO fields and internal links are correct.
Run a final batch audit showing article word count, status, featured media, first image, noindex state and schema. This is the visual equivalent of a release checklist.
A Full Production Case: Educational Diagram
Suppose the topic is a water cycle. First write the factual specification: evaporation, condensation, precipitation, collection and the direction of movement.
Generate a visual draft with appropriate arrows and labels. Check every label against the specification. If the image cannot maintain exact topology or text, rebuild the final diagram in vector software using the generated draft only as a composition reference.
The teaching objective determines whether generative speed or deterministic precision should dominate.
A Full Production Case: Social Image Series
Create a series rule: one subject, one short headline area, shared palette and consistent crop. Generate several base compositions and approve one grammar.
Each new post changes subject and supporting detail while preserving hierarchy. Review the series together every few posts to detect drift.
When the grammar becomes stale, update deliberately rather than letting gradual random changes break coherence.
The Image Maintenance Cycle
- Review live placement.
- Check source and rights state.
- Check visual accuracy.
- Check crop and mobile rendering.
- Check alt text.
- Check file size.
- Check whether the style still matches the series.
- Replace or edit only when a real improvement is available.
- Preserve canonical source and approved derivatives.
The Final Visual Receiver Test
Place the image in the actual page, slide, worksheet or post and view it at realistic size. Ask whether the receiver immediately understands the visual role and whether any detail creates confusion.
A visual can pass generation review and fail layout review. Cropping, overlays, responsive behaviour and nearby text all change how the image works.
The image is finished when it works in context, not when it looks attractive in the generation window.
The Final Visual Portability Test
Take the visual brief and use another image tool or a human designer. The purpose, composition, invariants and checks should remain understandable.
If the workflow depends on one prompt phrase that nobody can explain, the system is fragile.
Portable visual design lives in the brief, reference pack and quality criteria.
The Final Visual Production Gate
Before publishing, confirm the exact asset, approved state, purpose, receiver, required factual details, alt text, rights or consent, crop, file performance, featured/header assignment and destination page.
For an edited image, compare it with the approved baseline. For a factual image, return to the authoritative source. For a generated visual that could be mistaken for documentary evidence, use appropriate framing or disclosure.
This final gate turns SI image generation from attractive experimentation into a controlled visual-production system.
How Image Generation Produces Pictures: A Mechanism and Repair Lab
A prompt names a picture that does not yet exist. An image generator must turn that request into numerical image information, then into visible colour and structure. This lab explains two documented ways to do that and gives you a complete, inspectable exercise. You will calculate a small reconstruction, compare supplied layouts and choose a repair that preserves an approved design.
What the examples are: the numbers, diagrams and candidate outcomes below are teaching fixtures authored for this exercise. They are not outputs collected from a commercial image model, evidence of a product’s performance or photographs of an event. No image-generation call is needed to answer the questions. The coordinate descriptions are the supplied output, so every assessed change is visible in the packet.
1. Start With a Representation, Not a Tiny Picture in a Word
A colour raster image can be represented as a height-by-width grid with red, green and blue values at each position. For example, a 1,200-by-800 RGB canvas contains 1,200 × 800 × 3 = 2,880,000 channel values. That counts numerical entries, not file bytes: storage also depends on numeric precision, compression and file format.
A learned representation need not retain one number per visible channel. In the latent-diffusion approach described by Rombach and colleagues, an encoder compresses an image into a smaller spatial array of learned features; a decoder maps such features back to pixels. These latent values are not ordinarily human-readable labels such as “circle” or “red”. Compression and decoding can also lose detail. See the latent-diffusion paper, Section 3.1.
Keep three meanings separate throughout this lab: pixels are the final display values; learned latents or image tokens are internal representations; our named grid cells are a human-readable design specification. A cell called C4 is not an actual model token, and our colour names do not reveal how a neural network stores colour.
2. Two Documented Generation Routes
Latent diffusion: in the documented LDM family, training uses encoded images with added noise. A network learns a noise-prediction task across noise levels. At generation time, the sampler starts with noisy latent values and repeatedly uses the learned prediction to update them; the decoder converts the final representation to an image. Conditioning information, such as an encoded text prompt, can influence those predictions through cross-attention. This is a description of the LDM mechanism, not a claim about every image tool.
Training changes model parameters using examples. Inference uses the learned parameters to produce a candidate. For the diffusion formulation discussed here, forward noising corrupts a known training example; reverse sampling learns to produce plausible data from noise. Generation does not require an unseen, correct photograph already hidden under the starting noise. The DDPM paper specifies this forward/reverse distinction.
Autoregressive image tokens: the original DALL·E paper describes a different route. A discrete autoencoder represents a 256-by-256 RGB image as a 32-by-32 grid, giving 1,024 image tokens. A transformer models text and image tokens as one sequence. At generation time, image tokens are selected successively, conditioned on the text and preceding tokens, then decoded into an image. This is a historical, documented example, not an architectural description of every later product with a related name. See Zero-Shot Text-to-Image Generation, Section 2.
The key distinction is the update being learned: the first example repeatedly refines a noisy representation; the second builds a sequence through conditional token predictions. Both can use compressed image representations. Neither description means that a model must paint one recognisable object completely before starting another.
A request can admit many valid pictures. Sampling choices can therefore produce different candidates from the same brief. Our practical inference is that a prompt should be accompanied by acceptance tests: plausible appearance alone does not establish an exact count or the right label-to-object relationship. These examples are not an exhaustive classification. Do not infer a current provider’s private architecture from a progress animation, product name or the appearance of its pictures.
3. Worked Reconstruction: What a Noise Estimate Does
We now isolate one calculation from a noise-prediction formulation. It is a numerical teaching example, not a trained model, image decoder or complete reverse-sampling algorithm. Let x0 be a two-entry clean representation. Let ε be a specified noise vector, xt the noisy observation, and ε-hat a supplied estimate of that noise. The symbol t identifies a noise level, not elapsed seconds.
Use xt = a × x0 + b × ε, where a is the retained-signal coefficient and b is the noise coefficient. In the standard forward marginal, a = √ᾱt and b = √(1 − ᾱt), so a² + b² = 1. Here ᾱt means the cumulative signal-retention quantity at level t. This setup follows the DDPM forward marginal, Equation 4; our small values and supplied estimates are invented for transparent arithmetic. In that formulation ε is standard Gaussian noise; here its entries are fixed so you can check each operation.
Given: x0 = [0.50, -0.25] ε = [1.00, -0.50] a = 0.80; b = 0.60; ᾱt = 0.64 Forward calculation: xt[1] = 0.80(0.50) + 0.60(1.00) = 1.00 xt[2] = 0.80(-0.25) + 0.60(-0.50) = -0.50 xt = [1.00, -0.50]
We know the clean vector because we constructed this exercise. A deployed denoiser does not get that answer handed to it. Suppose its supplied estimate is ε-hat = [0.80, −0.40]. Rearrange the equation to estimate the clean representation: x0-hat = (xt − b × ε-hat) ÷ a, applying each operation to each entry.
Reconstruction with the imperfect estimate:
x0-hat[1] = (1.00 - 0.60(0.80)) / 0.80
= 0.52 / 0.80 = 0.65
x0-hat[2] = (-0.50 - 0.60(-0.40)) / 0.80
= -0.26 / 0.80 = -0.325
Error = x0-hat - x0 = [0.15, -0.075]
Mean squared error = (0.15² + (-0.075)²) / 2
= 0.0140625
Check the explanation: if the noise estimate had equalled [1.00, −0.50], reconstruction would have returned [0.50, −0.25] exactly. With our imperfect estimate, too little positive noise is removed from the first entry, so it remains too high. Too little negative noise is accounted for in the second, so it remains too negative. The error equals (b ÷ a) × (ε − ε-hat). Here b ÷ a is 0.75.
Where the analogy stops: these two numbers are not two image pixels, two objects or measured model activations. No trained predictor, text conditioning, noise schedule, iterative sampler or image decoder has been implemented. A real generation run does not reveal its “true intended picture” for this comparison. A low error on our vector would not establish correct spelling, geometry or factual truth in an image. The calculation teaches how prediction error affects a reconstruction estimate, and nothing more.
4. Source Packet: An Exact Three-Shape Teaching Card
The reader job is to distinguish shape, colour and label binding in a simple educational illustration. All objects and text are fictional. Use a landscape 1,200-by-800 canvas with white background, flat fills, black outlines, no textures, shadows, people, logos or extra objects. Coordinates below are pixels measured from the top-left: x increases rightward; y increases downward. The logical grid has 12 columns A–L and eight rows 1–8, each 100 pixels square.
- Title: THREE SHAPES, centred at (600, 60), black sans-serif text, 36 px. Footer: ILLUSTRATION, centred at (600, 750), black sans-serif text, 24 px. Neither is part of the object count.
- Circle: red fill #D64545, centre (250, 350), radius 30, black outline 3 px. Its centre is C4.
- Square: blue fill #2563EB, centre (550, 350), width and height 60, black outline 3 px. Its centre is F4.
- Triangle: green fill #228B22, vertices (850, 315), (815, 385), (885, 385), black outline 3 px. Its centre region is I4.
- Labels: exact words CIRCLE, SQUARE and TRIANGLE, centred respectively at (250, 550), (550, 550) and (850, 550), black sans-serif text, 24 px. Each label must identify the object directly above its own centre.
- One black, 3-px right-pointing arrow from (350, 350) to (475, 350), with a 12-px arrowhead at the endpoint. It points from the circle toward the square. No other arrows.
- Everything else is white. Preserve the canvas, title, footer, object positions, geometry, outlines, labels and arrow unless a later instruction explicitly permits a change.
These coordinates and exact text are the reference specification. The following compact maps show object centres, arrow direction and label anchors; they do not reproduce font rasterisation. A dash means no object or label anchor at that grid cell. Text occupies the space needed around its stated anchor. cR means red circle, sB blue square, tG green triangle; C, S and T mean the full labels CIRCLE, SQUARE and TRIANGLE.
Required layout, rows 4 and 6:
A B C D E F G H I J K L
r4 - - cR - -> sB - - tG - - -
r6 - - C - - S - - T - - -
Title and footer: as specified above
All other object/label anchors: absent
Task: before reading the answer, choose which prepared candidate meets the source packet. Count objects separately from arrows and labels. Then check labels by position, not by whether the right three words appear somewhere. All candidates below retain every source-packet field except their explicitly listed changes; that inheritance is part of the supplied candidate definition.
5. Prepared Candidates and a Fully Worked Decision
Candidate A: exactly the source packet, including all colours, geometry, text and the rightward arrow. Its map is the required layout above.
Candidate B: the source packet plus a second green triangle at K4, with vertices (1050, 315), (1015, 385), (1085, 385). There is no fourth label. Every original object and label remains in place.
Candidate B:
A B C D E F G H I J K L
r4 - - cR - -> sB - - tG - tG -
r6 - - C - - S - - T - - -
Candidate C: the source packet with SQUARE centred at (250, 550) and CIRCLE at (550, 550). TRIANGLE stays at (850, 550). No objects move, no text is misspelled and no other field changes.
Candidate C:
A B C D E F G H I J K L
r4 - - cR - -> sB - - tG - - -
r6 - - S - - C - - T - - -
Answer: A passes the supplied structural specification. B has four shapes rather than three; the extra triangle is still an object even though it lacks a label. C has three shapes and the correct vocabulary, but two bindings are wrong: SQUARE identifies the circle by placement and CIRCLE identifies the square. A word-presence check would miss this failure.
The least disruptive repair for B is to remove the added triangle and restore its area to the specified white background. Recreating the entire scene would risk correct content for no benefit. For C, exchange the two misplaced text labels while preserving the shapes. In an exact teaching card, editable text or a deterministic drawing tool makes this a direct correction. Recheck count, all three bindings and the arrow afterward.
A passes only what the packet lets us inspect. We have not measured a rendered font, colour contrast on a device, antialiasing or a commercial generator’s success rate. If these specifications became an actual worksheet, readability at final print size would remain an additional acceptance test.
6. Edit Packet: Permission to Change One Region Is Not Permission to Change Everything Inside It
Take approved Candidate A as the baseline. The new instruction is: “Change only the blue square’s fill to yellow #F4C542. Keep its centre, 60-by-60 size and black outline. Preserve everything else.” The allowed edit region M is F4: 500 ≤ x < 600 and 300 ≤ y < 400. It contains the complete square and a surrounding white margin. M describes permission in this exercise, not a file encoding for an actual provider’s mask.
The square’s fill is the only permitted semantic change. The outline and geometry remain protected even though they lie inside M. Outside M, the arrow, other shapes, all lettering and background must remain unchanged. This gives two separate tests: whether a change crosses the region boundary, and whether an allowed-region change exceeds the requested operation.
- Edit E1: change the square fill to #F4C542; all remaining fields exactly inherit Candidate A.
- Edit E2: make the E1 colour change and move the arrow tip from (475, 350) to (575, 350), lengthening the arrow into the square. All remaining fields inherit A.
- Edit E3: make the E1 colour change and enlarge the square to 140-by-140, keeping its centre (550, 350). Its new geometric bounds are x = 480–620 and y = 280–420. All remaining fields inherit A.
Answer: accept E1 for this symbolic specification. Reject E2: part of the changed arrow lies outside M, and the arrow endpoint was protected regardless of where it lies. Restore the original arrow from A rather than describing the whole scene again. Reject E3: the square crosses every side of the mask region and violates the size invariant. Restore the 60-by-60 square and apply only the fill change.
For the repaired result, compare the square bounds, outline and centre; inspect the neighbouring cells E4, G4, F3 and F5 for boundary spill; then check the distant circle, triangle, labels, title and footer. A local change can pass a close-up inspection while damaging distant content. Checking only yellow pixels does not test preservation.
In an actual tool, confirm its mask convention before supplying a mask: our M does not mean “black”, “white” or “transparent” universally. OpenAI’s current image-editing documentation says its mask guides editing and may not be followed with exact spatial precision. Its limitations section also discusses text placement, consistency and composition. A mask therefore does not by itself certify unchanged pixels. If pixel identity matters, retain the original outside-region pixels through controlled compositing and verify the resulting image.
7. Independent Transfer: A Different Layout and a Different Error
Now use a separate square 1,000-by-1,000 teaching card. The grid is A–J by rows 1–10, still 100 pixels per cell. The background is white. Title FOUR SYMBOLS is centred at (500, 60), black sans-serif, 36 px; footer ILLUSTRATION at (500, 960), black sans-serif, 24 px. All outlines and the single arrow are black, 3 px; labels are black sans-serif, 24 px. No shadows, other objects or other text are present.
- C3: navy triangle #183153, vertices (250, 215), (215, 285), (285, 285). Label ROOF centred at (250, 350).
- H3: orange circle #EA8A23, centre (750, 250), radius 30. Label SUN centred at (750, 350).
- C7: grey square #777777, centre (250, 650), width and height 60. Label BASE centred at (250, 750).
- H7: purple diamond #8057B8, vertices (750, 610), (790, 650), (750, 690), (710, 650). Label TILE centred at (750, 750).
- One downward arrow from (750, 430) to (750, 570), with a 12-px arrowhead at its endpoint, connecting the upper-right and lower-right regions. All other space is white.
Transfer baseline: object and label anchors
C H
r3 navy triangle orange circle
r4 ROOF SUN
r5 down arrow
r6 arrow tip
r7 grey square purple diamond
r8 BASE TILE
The edit request is to change only the purple diamond’s fill to teal #008B8B. Preserve its four vertices, outline and TILE label, and every other field. The allowed edit region is H7: 700 ≤ x < 800 and 600 ≤ y < 700. The following supplied outcomes inherit the entire transfer baseline except for the stated changes.
- Outcome U: the diamond becomes teal, and the arrow direction reverses: it starts at (750, 570) and ends with its arrowhead at (750, 430).
- Outcome V: the diamond becomes teal, its right vertex moves from (790, 650) to (770, 650), and its left vertex moves from (710, 650) to (730, 650). Its top and bottom vertices stay fixed.
- Outcome W: only the diamond fill becomes teal.
Questions: Which outcome passes? For each failure, identify the violated requirement and the least disruptive repair. Does staying inside H7 guarantee that the edit is correct? Which unchanged details must be checked after repair? Answer those before proceeding to the worked solution.
Transfer answer: W passes the supplied specification. U reverses a protected relationship outside H7; restore the downward arrow from the transfer baseline, retaining the accepted teal fill. V narrows the diamond from 80 to 40 pixels while staying inside H7. It fails the shape-geometry invariant even though it respects the region boundary. Restore the left and right vertices to (710, 650) and (790, 650), keeping teal.
Recheck four objects, all four shape-label bindings, the downward arrow, diamond vertices and outline, unchanged colours of the other shapes, both text banners, canvas and white background. The transfer changes object count, canvas shape, arrangement, labels, target shape and arrow direction. The answer depends on this new packet; copying the earlier “three shapes, rightward arrow” checks would be wrong.
Independent reconstruction question: use x0 = [0.50, −0.50], ε = [−0.25, 0.50], a = 0.60 and b = 0.80. First compute xt. Then reconstruct using ε-hat = [−0.25, 0.35]. Give both entries, the signed error and mean squared error. The equations are the same; none of the earlier numerical answers carry over.
Transfer calculation and answer:
xt = [0.60(0.50) + 0.80(-0.25),
0.60(-0.50) + 0.80(0.50)]
= [0.10, 0.10]
x0-hat = [(0.10 - 0.80(-0.25)) / 0.60,
(0.10 - 0.80(0.35)) / 0.60]
= [0.50, -0.30]
Error = [0.00, 0.20]
Mean squared error = (0.00² + 0.20²) / 2
= 0.02
The two noisy entries are equal, but their clean estimates differ because the noise estimates differ. The second noise estimate is too small, leaving that reconstruction too high by 0.20. This time b ÷ a is 4/3; do not reuse the earlier error multiplier of 0.75.
8. Why a Convincing Picture Does Not Establish an Event
Our approved card establishes that a specification was met within the supplied fixture. It says nothing about whether three physical shapes were ever placed on a table. The same boundary matters for photorealistic scenes: a generator can construct a visually plausible event from a request without anyone witnessing that event. Light, texture and detail are features of the rendering, not independent corroboration.
A generated scene may be useful as a labelled illustration, hypothetical reconstruction or design proposal. To establish that an event occurred, seek evidence tied to that event and verify its source and context. Neither a good prompt, a low reconstruction error nor an apparently perfect edit turns a synthetic image into that evidence.
Completion check: you should now be able to distinguish repeated denoising from successive image-token generation, calculate the supplied reconstruction, diagnose count and binding errors, separate mask boundaries from preservation rules, and explain why visual realism is not proof. Carry those distinctions into your next visual-production task.
Creating Images With SI Is Visual Specification Plus Review
The strongest results come from defining the visual job, giving the system enough freedom to create and then inspecting the output against purpose, accuracy and receiver needs.
Treat generated images as drafts when details matter. Iterate deliberately, preserve invariants and verify factual or textual content before publication.
