How to Create AI Images of Yourself

Core Mechanisms of AI-Driven Personal Image Generation

Creating an AI image of yourself involves generating synthetic photographs or digital illustrations that accurately preserve your facial features, skin texture, bone structure, and overall likeness. While early generative models struggled with facial coherence, modern diffusion architectures—such as Stable Diffusion (including SD 1.5, SDXL, and Flux.1), Midjourney, and specialized facial-conditioning neural networks—can generate photorealistic or stylized depictions across virtually any background, lighting condition, attire, or artistic style.

To understand how to make AI pictures of yourself, it is necessary to distinguish between the three primary technical pathways available:

1. Model Fine-Tuning (DreamBooth & LoRA): Injecting personal biometric and visual features into the weights of a base diffusion model by training it on a curated dataset of 10 to 30 photographs. This method yields the highest consistency, flexibility, and photorealism.
2. Reference-Conditioned Generation (IP-Adapter & Character Reference): Guiding an existing model using one or more reference images without modifying the underlying weights. This approach extracts identity embeddings using vision encoders (like CLIP or InsightFace) and injects them directly into the cross-attention layers of the network.
3. Post-Generation Swapping & Inpainting (Face Swapping): Generating a generic scene or portrait and subsequently mapping your facial identity onto the generated subject using latent feature alignment algorithms.

+-----------------------------------------------------------------------------+
|                      AI Likeness Generation Methods                         |
+-----------------------------------------------------------------------------+
|  Method          | Technical Basis       | Best For              | Control  |
|------------------+-----------------------+-----------------------+----------|
|  LoRA /          | Weight Fine-Tuning    | Custom scenes, high   | Maximum  |
|  DreamBooth      | (Low-Rank Adaptation) | realism, photostudios |          |
|------------------+-----------------------+-----------------------+----------|
|  IP-Adapter /    | Cross-Attention Latent| Quick iterations,     | Moderate |
|  Image Reference | Image Conditioning    | zero-training setups  |          |
|------------------+-----------------------+-----------------------+----------|
|  Face Swap       | Latent Identity       | Template matching,    | Low to   |
|  (InsightFace)   | Replacement           | static poses          | Moderate |
+-----------------------------------------------------------------------------+

Method 1: Fine-Tuning via LoRA (Low-Rank Adaptation)

Low-Rank Adaptation (LoRA) is the standard method for personal identity modeling in open-source AI ecosystems (such as Stable Diffusion XL, Flux.1, and SD 1.5). Instead of fine-tuning all billions of parameters in a base model (which is computationally prohibitive), LoRA freezes the base model weights and trains a small, separate adapter matrix—typically between 20 MB and 300 MB in size—capturing your specific facial geometry.

Step 1: Preparing a High-Fidelity Training Dataset

The quality of the generated AI photos depends directly on the diversity, clarity, and resolution of your input dataset. Models require varied visual information to differentiate between your invariant facial structures and contextual artifacts (such as lighting, background, and clothing).

• Image Count: 12 to 25 high-resolution photographs.
• Composition Distribution:
  • 50% close-up headshots (capturing eye shape, skin details, dental structure, and hairline).
  • 30% medium shots (chest-up and waist-up, showing shoulder width and neck structure).
  • 20% full-body shots (providing contextual scale and posture).
• Variation Factors:
  • Lighting: Natural daylight, soft indoor lighting, and directional shadow to prevent the model from baking a specific light source into your face.
  • Angles: Direct forward gaze, 45-degree three-quarter profile, and slight upward/downward angles. Avoid extreme profile angles unless explicitly needed.
  • Expressions: Neutral, slight smile, open-mouth smile, and serious.
  • Backgrounds & Outfits: Different locations, solid colors, outdoor settings, and varied necklines. If all photos feature the same jacket or living room, the AI will associate those elements with your identity.
• Exclusions: Avoid heavy filters, extreme wide-angle selfie distortions (which distort facial proportions), sunglasses, group photos, or low-resolution crops.

Step 2: Image Preprocessing and Cropping

Ensure images are cropped to standard resolutions compatible with the chosen architecture:

• Flux.1 / SDXL: $1024 \times 1024$ pixels (or equivalent bucketed aspect ratios such as $896 \times 1152$).
• Stable Diffusion 1.5: $512 \times 512$ or $768 \times 768$ pixels.

Many training interfaces (such as Kohya_ss, AI Toolkit, or cloud trainers) handle dynamic aspect ratio bucketing automatically, but manually removing distracting elements or bad crops ensures cleaner convergence.

Step 3: Text Tagging and Captioning

Diffusion models learn by associating visual tokens with text descriptions. During training, a rare identifier word (a "trigger word," such as ohwx or sks) is linked to your face along with a generic class noun (man, woman, person).

Two captioning paradigms exist:

1. Minimalist Tagging: Using only the trigger and class (photo of ohwx man). This forces the model to encode all visual features (hair color, eye color, facial hair) into the trigger token ohwx.
2. Descriptive Tagging (Recommended for flexibility): Describing the scene elements to prevent "feature leakage" into the trigger word. For example:

   A high-resolution photograph of ohwx man wearing a black crew-neck t-shirt, looking directly at the camera with a neutral expression, outdoor natural lighting, blurry park background.

By labeling the black t-shirt and the park background, the model learns to associate those visual elements with the words black crew-neck t-shirt and park background, isolating the face geometry under the ohwx token.

Step 4: Training Hyperparameters

When training via tools like Kohya_ss, RunPod, or dedicated LoRA training web applications, target these baseline settings for SDXL or Flux architectures:

• Network Rank (Dimension / Dim): 16 to 64. Higher ranks capture more micro-detail but increase file size and risk overfitting.
• Network Alpha: Typically set to half of the Rank, or matching the Rank (e.g., Dim 32 / Alpha 16 or 32).
• Learning Rate: For the UNet/Transformer, $1 \times 10^{-4}$ to $5 \times 10^{-4}$ using optimizers such as AdamW8bit or Prodigy. For the Text Encoder, $5 \times 10^{-5}$ (or freeze the text encoder during Flux training).
• Total Training Steps: Calculated as (Number of Images × Repeats × Epochs). A standard target is roughly 1,500 to 2,500 total steps.

Method 2: Zero-Training Reference Models (IP-Adapter & Reference Prompts)

If you prefer not to train a dedicated model, zero-training inference engines offer an alternative by analyzing a reference image at generation time.

Image Prompt Adapter (IP-Adapter)

IP-Adapter is an architectural extension for diffusion models that separates text cross-attention from image cross-attention. It extracts image features using a frozen image encoder (such as CLIP-ViT or OpenCLIP) and routes them into newly added cross-attention layers.

1. IP-Adapter Face / InsightFace: Uses specialized facial recognition encoders trained on biometric vectors rather than generic CLIP visual semantics.
2. Application in Workflow: In interfaces like ComfyUI or WebUI (Automatic1111/Forge), an IP-Adapter node takes one to five reference portraits of you, extracts facial landmarks and embeddings, and steers the text prompt generation toward your features.
3. Weight Adjustment: Set the IP-Adapter weight between 0.6 and 0.85. Setting it too high ($>0.9$) makes the output rigid and replicates the exact pose and lighting of the source photo; setting it too low ($<0.5$) causes the generated face to drift into a generic appearance.

Proprietary Platform Character References

Certain proprietary models feature built-in reference systems without requiring explicit script installations:

• Midjourney (--cref): Append --cref [URL of your photo] to a text prompt. You can adjust the fidelity with the character weight parameter (--cw 0 to --cw 100). Setting --cw 0 focuses strictly on preserving facial features while allowing clothing and hair to change entirely, whereas --cw 100 attempts to copy facial identity, hair, and clothing.

Method 3: Commercial and Turnkey Consumer Applications

For users seeking a simplified workflow without local GPU setups (which typically require NVIDIA GPUs with $\ge 8\text{GB}$ to $16\text{GB}$ VRAM) or technical configurations, dedicated cloud services handle dataset parsing, LoRA generation, and final rendering automatically.

+--------------------------------------------------------------------------+
|                       End-to-End Generation Pipelines                    |
+--------------------------------------------------------------------------+
|  Pipeline Stage       | Description                                      |
|-----------------------+--------------------------------------------------|
|  1. Upload Intake     | Service ingests 10-20 user selfies               |
|  2. Automated QA      | AI filters blurry, low-contrast, or cropped faces|
|  3. Cloud Fine-Tuning | Runs temporary LoRA or DreamBooth script on A100 |
|  4. Batch Generation  | Executes prompt matrices (headshots, art, styles)|
|  5. Post-Restoration  | Applies face enhancement (CodeFormer/GFPGAN)     |
+--------------------------------------------------------------------------+

When using turnkey tools, the user's primary responsibility is providing clean, diverse reference inputs. The platform executes the underlying training and inference pipelines in the background.

Prompt Engineering for Identity-Trained Models

Once a LoRA model or reference system is configured, you must construct prompts that balance stylistic creativity with likeness preservation.

Essential Prompt Structure

A robust portrait prompt consists of five distinct structural blocks:

1. Subject Identifier: Trigger word + generic class (e.g., a medium shot of ohwx man).
2. Setting & Wardrobe: Environment, background elements, and garments (e.g., wearing a charcoal Italian wool suit, seated in a modern minimalist boardroom).
3. Lighting & Atmosphere: Illumination dynamics (e.g., cinematic rim lighting, diffused window light, shallow depth of field).
4. Photographic & Medium Specifics: Lens, camera type, and film characteristics (e.g., shot on 85mm f/1.4 lens, Sony A7R V, highly detailed skin pores, natural skin texture, 4k resolution).
5. Negative Prompts: Explicitly filtering out unwanted artifacts (e.g., deformed eyes, plastic skin, oversaturated, cartoon, blur, extra limbs, bad anatomy).

Example Prompt Matrices

Professional Corporate Headshot

Prompt: A professional executive portrait of ohwx woman, dressed in a navy blazer over a white blouse, clean contemporary architectural office background with soft daylight, confident warm smile, shot on Hasselblad H6D-100c, 85mm lens, f/2.8, photorealistic, natural skin micro-texture, subtle catchlights in eyes.
Negative: airbrushed, CGI, drawing, smooth plastic skin, 3d render, exaggerated makeup, distorted eyes, asymmetrical pupils.

Cinematic Editorial Scene

Prompt: Editorial fashion photography of ohwx man walking through rain-slicked Tokyo streets at twilight, illuminated by vibrant neon signs and reflections, wearing an oversized trench coat, dramatic atmospheric lighting, shallow depth of field, 35mm film still, Kodak Portra 400 grain, high fidelity.
Negative: illustration, painting, low quality, oversmooth, bad hands, mutated fingers.

Post-Processing, Inpainting, and Refinement

Initial outputs from diffusion models often contain small imperfections, such as misaligned eyes, blurred irises, unnatural teeth, or minor anatomical artifacts. Professional workflows incorporate post-generation refinement steps.

[Raw Diffusion Output] 
         │
         ▼
[Inpainting Face Detailer (ADetailer)] ──► Re-samples face at higher local resolution
         │
         ▼
[Facial Restoration (CodeFormer / GFPGAN)] ──► Reconstructs eyes, eyelashes, and teeth
         │
         ▼
[Latent Upscaling (ESRGAN / ControlNet Tile)] ──► Increases overall resolution to 4K

1. Inpainting (Canvas Redrawing)

If the body, lighting, and composition are ideal but the facial expression is slightly distorted, apply an inpainting mask over the face.

• Set Denoising Strength to 0.25 to 0.45.
• Re-run the generation using the same prompt. This forces the model to redraw only the masked facial region while maintaining the rest of the image.

2. Automated Face Detection (ADetailer)

In ComfyUI and Automatic1111, extensions like After Detailer (ADetailer) use YOLO (You Only Look Once) or MediaPipe object detection models to isolate the face bounding box automatically. The software automatically crops the face, runs an inpainting pass at higher native resolution, and pastes the enhanced face seamlessly back into the composite frame.

3. Face Restoration Networks

• CodeFormer: A transformer-based prediction network designed to restore degraded facial features. It excels at fixing distorted pupils and teeth, though high fidelity settings ($>0.6$ weight) can make the face look overly smooth or generic.
• GFPGAN: A generative adversarial network tailored for blind face restoration, useful for reconstructing fine facial hair and skin sharpness.

4. High-Resolution Upscaling

Generating an image directly at $4096 \times 4096$ pixels causes standard models to hallucinate multiple heads or duplicate bodies because diffusion models are trained on specific pixel-density ranges. The correct method for high-resolution output is Two-Stage Generation:

1. Generate at native base resolution ($1024 \times 1024$ for SDXL/Flux).
2. Pass the final image through a Latent Upscaler or specialized super-resolution model (such as 4x-UltraSharp or Nomos8k) at a low denoising strength (0.15 to 0.30) to add fine texture without altering the underlying likeness.

Comparing Methods and Trade-offs

Selecting the appropriate approach depends on available hardware, budget, technical experience, and required output volume.

Privacy, Security, and Data Governance

Creating synthetic images of real people involves biometric data processing, creating legal, security, and ethical considerations:

• Data Sovereignty and Storage: When using third-party commercial platforms, check the vendor's data retention policies. Verify whether uploaded source images and trained weight checkpoints are deleted after generation or retained to train underlying foundational models.
• Local Execution vs. Cloud Processing: Running open-source models (via ComfyUI or Automatic1111) on local hardware guarantees that raw photographs and derived LoRA files never leave your private machine.
• Biometric Consent and Identity Safety: Generating AI imagery of other individuals without explicit, informed consent violates the Terms of Service of major AI platforms and can violate privacy, publicity, and personality rights depending on jurisdiction (such as GDPR in Europe or state-level right-of-publicity laws in the US).
• Deepfake Safeguards and Watermarking: Models increasingly embed metadata (such as C2PA standards or SynthID) into generated outputs to distinguish synthetic portraits from authentic photography, ensuring transparency in digital distribution.

Creating a convincing AI portrait of yourself

To create AI images of yourself, you provide an image generator with one or more clear reference photos, then either use an image-to-image, reference-image, or personalized model feature to produce new portraits from written prompts. The best method depends on whether you want a quick stylized image, a realistic professional headshot, or a highly consistent likeness across many scenes.

The process is not simply “upload a selfie and press generate.” Good results depend on the quality and variety of reference images, the amount of control a tool gives over identity, and prompt details describing pose, lighting, clothing, setting, and style. It also involves privacy decisions: a face is sensitive personal data, and uploaded images may be retained or used differently depending on the service and its settings.

Only create or share AI images of people with their permission. Do not use a person’s likeness to deceive, impersonate, harass, create intimate content, or imply that they did or endorsed something they did not.

Choose the right creation method

Most AI image tools offer one or more of the following approaches. Names vary by platform, but the underlying techniques are similar.

For most people, a reference-image feature is the practical starting point. If you need many images that all look recognizably like you—such as a set of illustrated social graphics or a portfolio of fictional scenes—a personalized profile may be more effective. Training a custom model is usually unnecessary unless you understand the tool’s data handling and need unusually fine control.

Prepare reference photos that preserve your likeness

AI systems infer facial features, body proportions, hair, skin tone, and visual context from the material you provide. A blurry or heavily filtered selfie gives the system less reliable information than a small, varied collection of natural, well-lit photographs.

What to include

If a tool accepts multiple reference images, choose approximately 8 to 20 recent photos unless it specifies a different range. Use only images you own or are authorized to use. Aim for variety without changing your appearance so much that the system cannot determine what is consistent.

Useful images usually include:

• A clear, front-facing photo with your full face visible.
• Three-quarter and side views.
• Neutral, smiling, and relaxed expressions.
• A mixture of close-up headshots and upper-body images.
• Indoor and outdoor lighting, provided your face is still visible.
• Your usual hairstyle, facial hair, glasses, or other lasting features.

Use sharp images with reasonable lighting. Your face should occupy a meaningful portion of the frame, without another person partially blocking it. Remove photos with extreme beauty filters, novelty lenses, intense shadows, low resolution, motion blur, or text over your face. Avoid relying solely on one highly posed image; that often causes the AI to reproduce the same angle and expression.

What not to upload

Do not include sensitive documents, photos revealing a home address or workplace security details, images of minors without appropriate authority and safeguards, or photographs of other identifiable people unless the service explicitly supports multi-person references and everyone has consented. Consider also whether backgrounds reveal information you would not want stored, such as family photos, medical settings, vehicle registration plates, or location metadata.

If you want the system to learn only your appearance, crop references to remove unnecessary background and other people. Before uploading, check whether the service allows you to delete source images and generated outputs later.

A practical workflow for making AI pictures of yourself

The exact interface differs between image generators, but the following workflow works across most tools.

1. Define the purpose before generating

Decide what the image is for. A professional profile image, an anime-style avatar, a fashion concept, and a fictional travel portrait need different levels of realism and different prompt wording.

Be especially deliberate if the result might be interpreted as a real photograph. An AI image of yourself in a business setting may be fine as a stylized profile asset, but it should not be presented as documentary evidence, a real credential, or proof that you attended an event.

2. Review privacy and usage controls

Before uploading, read the provider’s current policy and settings for:

• Whether uploads are private by default or visible in a public gallery.
• Whether prompts, reference images, or outputs may be used to train future systems.
• How long files are retained and whether deletion is available.
• Who owns or can commercially use the generated image.
• Whether the service restricts realistic people, public figures, or sensitive uses.

Terms and capabilities can change. A reputable-looking brand name alone does not establish that your images will remain private. If privacy is essential, prefer a tool with clear controls, use the least personal data necessary, and consider local or organization-approved software where appropriate.

3. Upload references or create an identity profile

Select the tool’s reference, character, identity, image prompt, or personalization option. Upload the prepared photos and label the profile only with information you are comfortable storing. A name is seldom needed; a neutral label such as portrait-reference-1 is generally sufficient.

If the tool offers a strength or identity-preservation setting, begin at a middle value. A very high setting may copy the reference’s exact pose, lighting, and expression rather than create a new image. A very low setting may produce an attractive portrait that no longer looks like you.

4. Write a prompt that separates identity from the scene

Your reference images communicate who should appear. The prompt should explain what the resulting image should look like. Strong prompts typically specify subject, composition, environment, lighting, wardrobe, mood, medium, and details to avoid.

A useful template is:

[subject based on the supplied reference], [camera framing and pose],
[wardrobe and setting], [lighting], [style or medium], [mood],
[composition and quality details]

For example:

A natural-looking editorial portrait of the person in the reference images,
three-quarter view, wearing a charcoal sweater in a quiet modern library,
soft window light, shallow depth of field, warm but neutral colors,
relaxed confident expression, realistic skin texture, head-and-shoulders composition

For a deliberately non-photographic result:

Use the person in the reference images as the character, illustrated as a
watercolor travel-journal portrait, standing beside a rainy coastal train station,
navy raincoat, loose expressive brushwork, textured paper, muted blue and ochre palette

Be precise about visual goals, but avoid piling on contradictory directions such as “dramatic harsh noon sunlight” and “soft evenly lit studio portrait.” When a prompt contains too many unrelated ideas, the model may follow only some of them.

5. Generate several candidates and compare likeness

Generate a small batch rather than expecting the first output to be final. Compare candidates against your references for stable traits: face shape, eye spacing, nose, hairline, skin tone, glasses, freckles, facial hair, and age impression. Also inspect the overall image at full size.

An image can look plausible at thumbnail size while containing problems such as malformed hands, implausible earrings, asymmetrical glasses, duplicated teeth, distorted text, or a background that does not make physical sense. Do not assume the image is accurate just because it is photorealistic.

6. Revise one variable at a time

When a result is close, change the most important problem first. For example:

• If it does not look enough like you, increase reference strength slightly or use clearer, more varied reference photos.
• If the likeness is good but the pose is wrong, retain the identity reference and revise the framing and pose language.
• If the output looks over-smoothed, request natural skin texture and reduce beauty-oriented wording.
• If your hairstyle or glasses disappear, describe them explicitly and include a reference where they are clearly visible.
• If the scene is excellent but your face is distorted, generate again with a simpler scene or use a localized editing feature for the face.

Changing every setting at once makes it hard to learn what improved or worsened the result. Keep notes on the prompt and reference combination that worked well, especially when generating a larger series.

Prompting for common types of self-portraits

The phrases below are starting points, not commands that guarantee a particular outcome. Adapt them to the tool’s syntax and content rules.

Professional-looking headshots

A usable professional AI portrait usually benefits from restraint. Describe credible clothing, clean lighting, and a simple setting rather than an exaggerated luxury office or impossibly flawless skin.

Professional but approachable portrait of the person in the reference images,
chest-up composition, looking toward camera, simple dark blue jacket over a plain shirt,
soft studio lighting, neutral light-gray backdrop, natural smile, realistic proportions,
subtle retouching only

If the image is for employment, a speaker profile, or a public-facing organization, disclose AI generation where context makes that material to readers. Follow the organization’s own image and authenticity policies.

Social avatars and stylized art

Stylized formats can reduce pressure for a perfectly photographic likeness while preserving recognizable features.

The person in the reference images as a clean vector illustration,
head-and-shoulders avatar, round glasses and shoulder-length dark hair retained,
bright coral background, simple geometric shapes, friendly expression,
crisp lines, no text or logo

Specify stylistic qualities rather than naming a living artist if a platform’s policies discourage artist-name imitation. For example, use “bold graphic poster style,” “ink-and-wash illustration,” or “retro science-fiction paperback cover aesthetic.”

Fictional scenes and editorial concepts

AI can place your likeness in a scene that did not occur. This is useful for conceptual art, storytelling, or design mockups, but it creates an image that should not be represented as a record of reality.

Cinematic fictional scene featuring the person from the reference images,
walking through a futuristic greenhouse at dusk, lightweight explorer jacket,
wide environmental portrait, soft blue and amber lighting, believable scale,
realistic but clearly imaginative atmosphere

Avoid prompts that create false evidence of travel, employment, political activity, emergencies, arrests, medical treatment, or other consequential events.

Improve consistency without making images look artificial

A recurring challenge is balancing fidelity to your appearance with creative variation. The controls called reference weight, image strength, character strength, or similar settings influence this balance, but their definitions are not standardized.

Use this general pattern:

1. Establish identity first. Generate a straightforward portrait with simple lighting and clothing until the result resembles you.
2. Reuse the successful setup. Keep the same profile, reference set, seed where supported, and core identity wording.
3. Vary the scene gradually. Change wardrobe, location, or artistic medium one at a time.
4. Use composition references carefully. A separate pose or scene reference can help control framing, but it may compete with your identity reference. Make clear which image controls the person and which controls the composition, if the tool provides separate slots.
5. Edit rather than regenerate everything. Inpainting or localized editing can replace a shirt, background, or unwanted object while preserving a successful face—though it may still alter surrounding features.

Consistency will never be perfect. Many systems approximate a face rather than store a precise, stable 3D representation of it. Results may vary with prompt wording, aspect ratio, lighting, model updates, and the presence of other faces in a scene.

Common problems and how to address them

Do not attempt to “fix” a poor likeness solely by adding many physical traits to a prompt. Reference quality and identity controls generally matter more than a long list of facial descriptions, and excessive detail can produce strange results.

Privacy, consent, and authenticity

Creating an AI image of yourself is usually lower risk than creating one of another person, but it still raises important issues. Your face can be used for recognition, impersonation, or identity-based profiling. Once a realistic image is posted publicly, other people may copy, alter, or misrepresent it.

Consider these safeguards:

• Use a unique password and available account security features for the generation service.
• Do not upload more reference images than the task requires.
• Delete training sets or identity profiles you no longer need, if the service permits it.
• Keep original photographs and generated sensitive portraits out of public project spaces.
• Review outputs before sharing for accidental private details, misleading uniforms or affiliations, recognizable bystanders, and realistic-looking fabricated events.
• Label an image as AI-generated or AI-assisted when viewers might reasonably mistake it for an authentic photograph, especially in professional, journalistic, political, educational, or commercial contexts.

A generated portrait may also be subject to platform rules, contracts, workplace policies, advertising standards, or local laws. For commercial campaigns, regulated industries, official materials, or anything involving identity verification, obtain appropriate legal, compliance, and brand review rather than relying on general guidance.

Editing and publishing the finished image

After selecting an image, ordinary image editing may improve presentation. Cropping, resizing, color correction, background cleanup, and accessible alt text are often more useful than another round of generative changes. Do not use editing to conceal a material deception.

For profile uses, export the dimensions required by the destination platform and keep a higher-resolution private original. If you use an image on a site or document, write alt text that describes its visible content rather than claiming it is a literal photograph. For example:

AI-generated illustrated portrait of a person with dark hair and round glasses against a coral background.

If an image matters as evidence, identification, an official record, or a representation of a real event, an AI-generated portrait is generally inappropriate. Its visual realism does not make it a reliable record.

When not to use AI self-portraits

AI portraits are useful for creative exploration, non-documentary visuals, avatar concepts, and some clearly disclosed marketing or design work. They are a poor fit when trust depends on factual depiction. Avoid using them for identification cards, verification photographs, legal or insurance evidence, medical documentation, dating profiles where authenticity is expected, job applications that require a current true photograph, or claims about real experiences.

The central distinction is simple: an AI image can be an expressive representation of you, but it is not automatically a truthful photograph of you at a particular time and place. Treating that distinction seriously protects both your privacy and the trust of people who see the image.