How to Make AI Pictures

Understanding AI Image Generation

Learning how to make AI pictures involves interacting with generative artificial intelligence models that convert natural language text descriptions (prompts) or reference visual data into high-resolution synthetic images. These systems do not simply search an existing image database or stitch collage fragments together. Instead, they generate novel visual compositions pixel by pixel based on mathematical patterns learned from massive training datasets consisting of image-text pairs.

The process of generating AI images generally follows a structured workflow:

1. Model Selection: Choosing an appropriate generative system (cloud-hosted service or locally run open-source model) based on required control, fidelity, and privacy needs.
2. Prompt Construction: Drafting a descriptive prompt specifying subject matter, artistic style, medium, composition, and lighting.
3. Parameter Configuration: Adjusting generation variables such as resolution, aspect ratio, guidance scale, sampling steps, and random seeds.
4. Inference & Sampling: Executing the model to iteratively denoise random data into a coherent visual output.
5. Post-Processing & Refinement: Modifying outputs through selective regeneration (inpainting), expansion (outpainting), upscaling, or fine-tuning with structural control adapters.

The Underlying Technology: How Generative Models Work

To make effective use of AI image tools, it helps to understand the underlying computational architectures that power them.

+-----------------------------------------------------------------------+
|                         TEXT PROMPT INPUT                             |
|               "A cinematic portrait of a red fox in snow"              |
+-----------------------------------------------------------------------+
                                    │
                                    ▼
+-----------------------------------------------------------------------+
|                   TEXT ENCODER (e.g., CLIP / T5)                      |
|        Translates human words into mathematical vector embeddings      |
+-----------------------------------------------------------------------+
                                    │
                                    ▼
+-----------------------------------------------------------------------+
|                    LATENT DIFFUSION / DENOISING                       |
|  Random Noise  ──►  Step-by-step Guided Filtering  ──►  Latent Image  |
+-----------------------------------------------------------------------+
                                    │
                                    ▼
+-----------------------------------------------------------------------+
|                       DECODER (VAE / Autoencoder)                     |
|        Transforms mathematical latent space into final visual pixels   |
+-----------------------------------------------------------------------+

Diffusion Models

Modern state-of-the-art tools—such as Stable Diffusion, FLUX, Midjourney, and DALL-E 3—rely primarily on latent diffusion models (LDMs) or diffusion transformers (DiTs). Diffusion models work through a two-phase process:

• Forward Diffusion (Training): The system takes a clean image and systematically adds Gaussian noise across hundreds of steps until it becomes pure, unrecognizable static.
• Reverse Diffusion (Inference): The model is trained to reverse this process. Starting with a field of pure noise, a neural network predicts and removes the noise step-by-step, conditioned on text prompts, until a crisp image emerges.

Latent diffusion executes this denoising process inside a compressed mathematical space called latent space rather than at full pixel resolution. A variational autoencoder (VAE) then decodes this compressed data back into standard visual pixels, drastically reducing computational overhead while retaining fine detail.

Multimodal Text Encoders

Generators bridge language and imagery using multimodal encoders such as CLIP (Contrastive Language-Image Pre-training) or T5. These encoders convert textual words into multidimensional vectors (embeddings). The diffusion backbone references these embeddings at every denoising step to ensure that the emerging visual features correlate with the descriptive terms in the prompt.

Primary Platforms and Tool Selection

Creating AI images can be done through fully managed cloud applications or through open-weights architectures run on personal hardware or dedicated cloud instances.

Step-by-Step Guide: Generating Your First AI Image

1. Selecting an Environment

For immediate access without technical setup, cloud-based tools like ChatGPT Plus (DALL-E 3), Midjourney, or Adobe Firefly offer turn-key web interfaces. For granular parameter control, privacy, and zero recurring subscription fees, open-source web interfaces like ComfyUI or Automatic1111 (Stable Diffusion WebUI) can be installed locally on systems with capable NVIDIA or Apple Silicon graphics hardware.

2. Formulating the Prompt

Prompting is the practice of translating visual intent into descriptive language. Prompts generally perform best when organized into distinct components:

• Core Subject: Define the primary focal point with concrete nouns (a cybersecurity analyst, a vintage mechanical watch, a Siberian tiger).
• Action / Context: Specify what the subject is doing and the broader environment (inspecting glowing holographic server racks inside a brutalist concrete data center).
• Artistic Medium & Style: Declare the visual medium (35mm documentary photography, oil on canvas with heavy impasto, isometric 3D vector illustration, macro photography).
• Lighting & Atmosphere: Detail the light sources and mood (volumetric golden-hour sunlight, high-contrast chiaroscuro, diffused overcast ambient light, neon rim lighting).
• Composition & Framing: Direct the camera (close-up portrait, wide-angle lens, Dutch angle, shallow depth of field with creamy bokeh, f/1.8 aperture).

[Subject + Context] + [Medium / Art Style] + [Lighting & Color Palette] + [Camera / Framing]

Example Prompt:
"An elderly horologist repairing an intricate brass clockwork mechanism, macro photography, warm side-lighting from an incandescent desk lamp, rich earth tones, shallow depth of field, 85mm lens, f/2.0, highly detailed mechanical gears."

3. Setting Technical Parameters

Before executing the generation, adjust the technical hyperparameters that govern output dimensions and algorithmic behavior:

• Aspect Ratio / Resolution: Set the horizontal-to-vertical proportion. Standard defaults include 1:1 (square, e.g., 1024×1024), 16:9 (widescreen landscape, e.g., 1344×768), and 9:16 (vertical portrait, e.g., 768×1344). Generating at non-native aspect ratios directly on older models (like SD 1.5) can cause duplicate subjects or warped anatomy; newer models handle variable aspect ratios natively.
• CFG Scale (Classifier-Free Guidance): Governs how strictly the AI adheres to the prompt versus exploring latent variations. A low CFG (e.g., 2–4) yields softer, more creative interpretations; a balanced CFG (e.g., 6–8) balances fidelity and coherence; an excessively high CFG (>12) often causes color saturation, burning, and visual artifacts.
• Sampling Steps: The number of iterative denoising calculations performed. For modern samplers (such as DPM++ 2M Karras or Euler Ancestral), 20 to 35 steps are typically sufficient. Extremely low step counts (<10) leave images blurry or unfinished, while excessive steps (>60) offer diminishing returns and waste computational power.
• Seed Number: A numerical identifier that determines the initial random noise pattern. Keeping the seed constant while tweaking text phrases allows you to isolate and modify specific elements without changing the overall composition.

Advanced Composition and Control Techniques

While simple text prompts produce good results, achieving precise creative control over composition, pose, and consistency requires advanced workflows.

+-----------------------------------------------------------------------------------+
|                         ADVANCED GENERATION WORKFLOW                              |
+-----------------------------------------------------------------------------------+
                                          │
     ┌────────────────────────────────────┼────────────────────────────────────┐
     ▼                                    ▼                                    ▼
+-----------------------+     +-----------------------+     +-----------------------+
|       CONTROLNET      |     |       INPAINTING      |     |         LoRA          |
| Enforces strict poses |     | Modifies isolated     |     | Injects specific      |
| and edge outlines via |     | areas using masked    |     | subjects, characters, |
| reference sketches.   |     | regeneration.         |     | or artistic styles.   |
+-----------------------+     +-----------------------+     +-----------------------+

Inpainting and Outpainting

• Inpainting (Regional Editing): Allows you to paint a digital mask over an unsatisfactory part of an existing image (such as an irregularly rendered hand or an unwanted background object) and regenerate only that masked region using a targeted prompt.
• Outpainting (Canvas Extension): Extends the canvas beyond the original borders of an image. The AI analyzes the edges of the original visual data and generates continuous, contextually consistent scenery to expand the frame.

Structural Guidance via ControlNet

ControlNet is a neural network structure that provides explicit spatial conditioning to diffusion models. Instead of relying solely on descriptive words to position objects, ControlNet allows creators to guide the output using structural reference maps:

• Canny / Lineart: Extracts edges from a reference image and forces the AI to construct its generation strictly within those boundaries.
• OpenPose: Detects human skeletal joint positions from an input image, enabling you to transfer complex body postures to an entirely new character.
• Depth Maps: Uses z-axis spatial depth information from 3D models or 2D photos to preserve foreground-to-background spatial relationships.

Low-Rank Adaptation (LoRA)

LoRAs are small, modular model patches (typically 10MB to 200MB) trained on top of a base diffusion model. They introduce specific concepts without requiring the retraining of the entire multi-gigabyte foundation model. LoRAs are widely used to maintain consistent character appearances across multiple scenes, replicate niche artistic styles, or generate proprietary products with exact brand fidelity.

Upscaling and Hi-Res Fix

Generating raw images at extremely high resolutions (such as 4K or 8K) can overwhelm system memory and introduce repeating anatomical artifacts. Professional workflows generate images at standard baseline dimensions (e.g., 1024×1024) and upscale them using specialized neural upscalers (such as RealESRGAN, Ultimate SD Upscale, or Topaz Gigapixel). These algorithms intelligently reconstruct fine textures, skin pores, and sharp edges without introducing blurriness.

Troubleshooting Common Generation Issues

Creating AI imagery often involves iterative troubleshooting. Below are common generation artifacts and the technical methods used to resolve them:

1. Distorted Hands and Limbs

• Cause: Diffusion models process spatial regions locally without an internal 3D structural model of human biomechanics. Training datasets also feature hands in diverse, foreshortened angles, making statistical consistency difficult.
• Solution: Use newer generation architectures (FLUX, SDXL) with improved anatomical priors; use ControlNet OpenPose to supply explicit finger/hand skeletal maps; or apply Inpainting at high resolution specifically over the hand area with targeted prompts.

2. Illegible or Garbled Typography

• Cause: Traditional text encoders break words down into sub-word tokens rather than individual visual letterforms, making it difficult for the model to map text strings to spatial strokes.
• Solution: Wrap intended words in quotation marks when using models built on large language model backbones (such as DALL-E 3 or FLUX), or composite typography manually using graphic design software (Photoshop, Illustrator) after generating the visual background.

3. Oversaturated, "Burned," or Plastic Appearance

• Cause: Setting the CFG scale too high, using over-generalized buzzwords (e.g., "hyperrealistic, masterpiece, 8k trending on Artstation"), or running aggressive denoising values during image-to-image conversions.
• Solution: Lower the CFG scale to between 4.0 and 7.0; replace hyperbolic buzzwords with concrete photographic parameters (lens type, film stock, lighting style); and use specialized negative prompts or embeddings to reduce unwanted high-gloss plastic finishes.

Legal, Ethical, and Professional Considerations

Deploying AI images in professional, commercial, and editorial contexts requires awareness of the evolving regulatory and technical landscape:

• Copyright and Authorship: In many major jurisdictions (including the United States, via guidance from the U.S. Copyright Office), raw outputs generated purely by artificial intelligence without significant human creative modification are not eligible for copyright protection. Establishing copyright generally requires demonstrating substantial human creative expression, such as extensive manual editing, sequencing, or hybrid digital compositing.
• Commercial Licensing: Commercial usage rights depend entirely on the terms of service of the platform used and the training pedigree of the model. Enterprise platforms like Adobe Firefly are trained predominantly on licensed Adobe Stock and public domain content to provide indemnification against intellectual property claims, whereas other platforms place the legal responsibility for output usage on the end user.
• Provenance and Provenance Standards: To combat deceptive media and deepfakes, industry consortiums have established standards such as C2PA (Coalition for Content Provenance and Authenticity). Many modern generative platforms automatically embed cryptographic metadata into generated image files, recording the tool origin, generation date, and modification history.

The basic idea

How to make AI pictures depends on the image generator you use, but the underlying process is broadly the same: describe the desired image in natural language, choose relevant settings, generate one or more results, and refine the prompt or edit the output until it serves its purpose. You can use a web application, a design platform, a chatbot with image-generation capabilities, or locally installed software. Some tools create an image entirely from text, while others modify an existing photograph, extend a canvas, remove objects, or combine several visual references.

The quality of the result is influenced by more than the number of descriptive words in a prompt. A useful request identifies the subject, action, setting, composition, visual style, lighting, viewpoint, aspect ratio, and any constraints that matter. It is usually more effective to begin with a clear, relatively simple concept and then make targeted changes than to include every possible detail in one long instruction.

A typical workflow looks like this:

1. Choose a suitable image tool. Consider whether you need photorealism, illustration, editing, text in an image, transparent backgrounds, commercial-use rights, or control over a reference image.
2. Describe the image. State what should appear, where it should appear, and how it should look.
3. Set the format. Select a square, portrait, landscape, banner, or other aspect ratio when the tool provides that option.
4. Generate several candidates. The first result is often a draft rather than a finished image.
5. Evaluate the output. Check anatomy, hands, faces, text, perspective, object counts, lighting, and whether the image actually communicates the intended idea.
6. Refine selectively. Change one or two important details at a time, or use an editing feature such as inpainting, masking, or outpainting.
7. Review rights, privacy, and accuracy. These issues matter especially when the image will be published, used commercially, or presented as a real event.

Choosing a method for creating AI images

There is no single best way to generate AI images. The right method depends on the job, the level of control required, and the consequences of mistakes.

Text-to-image generation

Text-to-image systems create a new image from a written description. They are useful for concept art, mood boards, editorial illustrations, fictional scenes, background images, product concepts, and visual experimentation. The system interprets the prompt rather than retrieving a single existing picture, although generated images may still resemble patterns found in its training data.

Text-to-image is often the easiest starting point because it requires no source image. It is less predictable when the prompt demands exact geometry, a particular person, precise logos, long readable text, or a complicated arrangement of many objects.

Image-to-image generation and editing

Image-to-image tools use an existing image as a starting point. Depending on the application, you may be able to change its style, replace the background, alter clothing, recolor an object, improve a sketch, or transform a rough composition into a polished illustration. An image-to-image workflow generally preserves more of the original composition than a purely text-based request, although the extent of preservation varies with the tool and its settings.

Common editing operations include:

• Inpainting: changing a selected region while leaving the rest of the image mostly intact.
• Outpainting: extending an image beyond its original borders.
• Background replacement: isolating a subject and generating a new environment.
• Upscaling: producing a larger version, sometimes with added detail.
• Style transfer: applying characteristics of a visual medium or broad aesthetic to an existing image.
• Object removal or replacement: reconstructing the area behind an unwanted object.

Reference-image workflows

Some generators allow one or more reference images to guide subject identity, pose, color palette, layout, or visual style. This can be useful when an image must follow a rough sketch, match a brand palette, depict a recurring fictional character, or preserve a product’s general shape. A reference is guidance, not necessarily an exact template. Small details may change, and different tools provide different levels of control.

Local and hosted tools

A hosted service runs the model and interface on a provider’s infrastructure. This is usually convenient and may offer integrated editing, storage, collaboration, and automatic updates. A locally run system can offer greater control over models, workflows, and privacy, but it may require suitable hardware, installation, model management, and technical knowledge. Local use does not automatically eliminate legal or ethical concerns: the model, source material, output, and intended use still require review.

How to write an effective image prompt

An image prompt is an instruction that tells the generator what to create or change. There is no universal grammar, and different systems respond differently to natural sentences, keyword lists, weighting controls, or negative prompts. Nevertheless, a structured description gives the model a clearer target.

A useful prompt can contain the following components:

For example, instead of writing a bicycle, you might write:

A vintage red touring bicycle leaning against a weathered brick wall in a quiet city courtyard, soft early-morning light, muted blue and ochre palette, wide editorial illustration, bicycle placed on the right third of the frame, no people.

This prompt identifies the principal object, its environment, lighting, palette, composition, medium, and an exclusion. It does not guarantee a particular result, but it reduces ambiguity.

Put the most important information first

Many systems weigh some parts of a prompt more strongly than others, and long prompts can cause secondary details to be ignored. Start with the central subject and intended action. Add the details that are essential to the message, then include style and finishing preferences. If the image repeatedly misses a critical requirement, move that requirement earlier and state it more plainly.

Describe appearance rather than naming a vague mood

Words such as “beautiful,” “professional,” or “cinematic” can be useful but are broad. They become more informative when paired with observable features. For example, “professional headshot” can be expanded into “shoulders-up portrait, neutral background, even soft lighting, natural skin texture, calm expression, centered composition.”

Similarly, “dramatic” might mean deep shadows, a low camera angle, strong directional light, saturated colors, or an unusual perspective. Specify which visual properties you want rather than relying entirely on an abstract label.

Use concrete spatial language

If arrangement matters, describe it explicitly: “three jars in a row,” “the person standing behind the table,” “a close-up from eye level,” or “the title area left blank at the top.” Spatial instructions are not always followed perfectly, but they are more useful than simply listing objects. For a complex composition, generating a rough layout first and then editing separate regions may work better than requesting everything at once.

Be cautious with named artists and living creators

A prompt that names a visual creator may produce an imitation of that person’s recognizable style and can raise ethical, contractual, or policy questions. A safer and often more controllable approach is to describe the relevant characteristics: flat areas of color, visible ink outlines, compressed perspective, limited palette, loose brushwork, geometric forms, or a particular historical medium. The same principle applies to living photographers, designers, illustrators, and other creators.

Use negative instructions carefully

Some tools support a dedicated negative prompt, while others handle exclusions in the main instruction. Negative prompts can request the absence of common problems such as extra fingers, distorted faces, clutter, watermarks, or unwanted objects. They are not guaranteed to work, and an excessively long list can distract from the positive description. State the desired result first and use only the exclusions that materially matter.

A practical generation and refinement workflow

1. Define the purpose before generating

Decide where the image will be used. A social post, book illustration, presentation slide, website banner, product concept, and printed poster have different requirements. A banner needs a suitable horizontal composition and perhaps empty space for text. A profile image needs a recognizable subject at a small size. A print asset may require a sufficiently large source file and careful inspection of fine detail.

Also decide whether the image is meant to be factual. A fictional illustration can be freely imaginative, but an image accompanying news, medical information, historical documentation, or a public statement must not imply that a fabricated scene is a real photograph.

2. Select the aspect ratio and composition

Aspect ratio is the relationship between an image’s width and height. Common choices include square, landscape, portrait, and wide banner formats. Choose it before generation when possible, because cropping a portrait subject into a landscape canvas can remove important elements or create awkward framing.

Composition should reflect the intended display size. A scene with many small objects may look impressive at full resolution but become unreadable as a thumbnail. A poster or advertisement may need negative space—an intentionally uncluttered area—for a headline or logo to be added later in a design application.

3. Generate variations rather than expecting one perfect result

Create several candidates with the same general prompt, then compare them. Variation helps reveal which elements are stable and which are accidental. If every result has the wrong camera angle, revise the composition instruction. If only one candidate has a good face but another has the right background, an editing workflow may combine the useful properties more effectively than starting over.

Some tools expose a seed or similar reproducibility control. When available, keeping the seed fixed while changing one phrase can help identify the effect of that phrase. This behavior is tool-dependent and may not produce identical results across model versions or settings.

4. Change one major variable at a time

Suppose an image has the right subject and color but the wrong viewpoint. Change the viewpoint without also changing the medium, background, lighting, and pose. Iterative changes make it easier to understand what improved or damaged the result. If the image is fundamentally wrong, however, rewriting the prompt from the beginning may be faster than making many small corrections.

5. Use targeted editing for local errors

Inpainting or masking is appropriate when most of an image works and one region does not. Select the area narrowly enough to protect surrounding details, then describe the replacement in relation to the existing scene. For example, “replace the cup with a blue ceramic mug, matching the existing perspective and warm window light” is more useful than simply “a mug.”

Check the boundary around the edited region. AI systems may create mismatched shadows, repeated textures, inconsistent reflections, or changes just outside the mask. Several small edits can sometimes be cleaner than one broad edit.

6. Add text separately when accuracy matters

AI image generators have historically been unreliable with exact spelling, long sentences, tables, labels, and logos. Even when a tool can generate legible text, inspect every character. For invitations, diagrams, advertisements, packaging, and social graphics, it is often more dependable to generate the visual background without text and add the wording afterward in a design or layout tool. Leave deliberate empty space where the text will go.

Common problems and why they occur

AI-generated images can look convincing while containing structural errors. A careful review is part of the creation process, not an optional final polish.

Hands, faces, and bodies

Human anatomy is difficult because the system must coordinate shape, pose, perspective, occlusion, and repeated structures. Inspect hands, fingers, ears, teeth, eyes, jewelry, and joints at full size. Crowded scenes and unusual poses increase the risk of errors. Cropping, changing the pose, generating fewer people, or repairing a selected region may help.

Objects that merge or multiply

The model may combine nearby objects, create extra wheels, duplicate furniture, or turn a partly hidden item into an implausible shape. Clarify the number and position of important objects, simplify the scene, and check whether the viewpoint makes the requested arrangement physically possible.

Perspective, shadows, and reflections

Generated scenes can contain inconsistent vanishing points, shadows that point in different directions, reflections showing objects that are not present, or lighting that does not match the stated time of day. These flaws may be subtle, especially in decorative images. If physical accuracy matters, use a reference photograph, a 3D or design workflow, or professional review rather than relying solely on generation.

Unreadable symbols and invented details

Signs, screens, maps, architectural details, and technical instruments may look plausible without being correct. Do not assume that a realistic appearance means the content is accurate. For educational, scientific, medical, financial, or instructional images, verify the underlying information independently and label generated or illustrative material where viewers could reasonably misunderstand it.

Identity drift

When generating several images of the same fictional or real person, facial features, clothing, age, and body proportions may change. Reference images, consistent descriptions, and controlled editing can reduce drift, but they may not eliminate it. A real person’s likeness should not be used deceptively or without appropriate permission, particularly in endorsements, political material, intimate imagery, or other sensitive contexts.

Improving quality after generation

A generated image may need both creative and technical treatment. First correct the concept: subject, composition, pose, and lighting. Then address local defects. Finally prepare the file for its actual destination.

Useful post-processing steps can include:

• cropping or extending the canvas to fit the intended layout;
• adjusting exposure, contrast, color balance, and saturation;
• removing small artifacts with conventional image-editing tools;
• adding typography, captions, diagrams, or branding separately;
• checking resolution and compression for the final platform;
• comparing the image at full size and at the size viewers will actually see.

Upscaling can make an image larger, but it cannot reliably restore information that was never present. It may also invent texture or sharpen artifacts. For print or technical use, inspect the enlarged image closely rather than assuming that more pixels mean more accuracy.

File format matters as well. A format that supports transparency is useful for isolated objects, whereas a compressed photographic format may be sufficient for a web image. Preserve an editable source when possible and export a separate delivery copy so that later revisions do not require rebuilding the image.

Rights, privacy, and responsible use

The legal status of AI-generated material varies by jurisdiction and depends on factors such as human creative contribution, the source material, the model provider’s terms, and how the image is used. An image that contains little human authorship may receive different protection from a substantially edited composition. These are legal questions, not merely technical ones, and a qualified lawyer or rights specialist should review important commercial or disputed uses.

Before publishing an image, consider:

• Provider terms: Check the service’s current rules concerning ownership, licensing, commercial use, attribution, prohibited content, and retention of uploaded files.
• Training and references: Do not upload confidential documents, private photographs, unreleased products, or personal data unless the service and your organization permit it and the privacy implications are understood.
• Likeness and consent: Avoid deceptive depictions of identifiable people. Extra care is required for minors, intimate contexts, impersonation, and political persuasion.
• Trademarks and logos: A generated brand mark may resemble an existing mark or fail to meet professional identity requirements. Obtain clearance before relying on it.
• Copyrighted source images: Having access to an image does not necessarily grant permission to transform or redistribute it.
• Disclosure: If viewers could mistake a fabricated scene for a real photograph or event, label it as generated, reconstructed, or illustrative as appropriate.
• Bias and stereotypes: Review depictions of people, occupations, cultures, disability, age, and gender for unwanted stereotypes or unequal treatment.

Do not use an AI-generated image as evidence merely because it appears photographic. The same visual realism that makes generated pictures useful for storytelling can make them misleading when context is omitted.

When AI generation is not the best tool

AI image generation is useful for exploration, but it is not always the most reliable or efficient method. A stock image, commissioned illustration, camera photograph, vector drawing, 3D model, diagram, or manually designed layout may be preferable when exact identity, technical precision, repeatability, accessibility, or legal clarity is essential.

For example, a technical assembly diagram should preserve correct part relationships; a medical illustration should reflect reviewed anatomy; a product advertisement should depict the actual product accurately; and a logo should remain legible and reproducible at many sizes. AI can assist with early concepts in these cases, but the final asset may need conventional design and expert verification.

The strongest results often come from a hybrid process: use AI to explore possibilities, select a promising composition, edit it with ordinary design tools, verify factual or brand-sensitive details, and retain a human decision-maker throughout. This approach treats the generator as a creative instrument rather than an authority.

A reusable prompt pattern

A flexible starting pattern is:

Create a [subject] [action or state] in [setting]. Show [composition and viewpoint], with [lighting and time of day]. Use [medium or broad visual treatment] and a [color palette]. The image is for [purpose and aspect ratio]. Preserve [important details] and avoid [specific unwanted elements].

For editing, adapt it to describe the change and the relationship to the existing image:

Modify the selected area to show [replacement]. Match the existing [perspective, lighting, color, scale, and texture], while preserving [areas that must not change]. Remove [unwanted detail].

These patterns are starting points, not formulas. The most dependable practice is to generate a clear first draft, inspect it critically, and refine the parts that affect the image’s meaning, usability, and trustworthiness.