You uploaded a selfie into ChatGPT, typed "make it look cinematic," and got back something that looked like a screenshot from a phone camera with the contrast turned up. Not cinematic. Not a portrait. Just louder.
The gap between typing a style word and getting a result that actually looks like a film still isn't about your photo. It's about the instructions you gave the model. A one-word style label leaves every creative decision — the light, the lens, the color, the mood — to the AI's default guess, and that guess is almost always "even lighting, neutral color, sharp everything." Which is the opposite of cinematic.
This guide walks through one complete workflow: take a reference photo, apply a structured portrait prompt in ChatGPT, customize it, and end up with an image that looks intentionally shot rather than auto-generated. I'll use a real, tested prompt you can copy and run today — no prompt engineering background needed.
What you need before you start
A ChatGPT Plus or Team account. Image generation in ChatGPT uses OpenAI's built-in model (GPT-4o image generation). The free tier does not support image generation with reference photos.
One reference photo. A clear, well-lit selfie or portrait where your face is visible and not obstructed by hands, hair, or accessories. Front-facing or a slight three-quarter angle both work. You don't need a studio photo — a daylight selfie is fine.
Five minutes. The entire workflow from upload to finished portrait takes about that long, including generation time.
The core idea: a prompt is a lighting and camera direction, not a vibe
Here's why "make it cinematic" fails. That phrase tells the model a desired outcome but gives it no instructions on how to get there. A cinematographer creating a cinematic shot doesn't think in outcomes — they think in specific, controllable decisions:
Lens: wide or telephoto? A 35mm lens pulls you into the scene; an 85mm lens compresses the background and flatters the face.
Lighting: direction, color temperature, hardness. Golden hour light is warm and directional. Rembrandt lighting comes from a 45-degree angle and creates a triangle on the cheek.
Composition: where the subject sits in the frame, what's in the background, how shallow the depth of field is.
Color and finish: film stock (Kodak Portra is warm; Fuji is cooler), grain, contrast curve.
A good portrait prompt writes these decisions out explicitly so the model doesn't have to guess. Here's what that looks like in practice.
Step 1: Open ChatGPT and start a new chat
Create a new conversation in ChatGPT. You don't need to select a specific GPT from the store — the default ChatGPT with GPT-4o has image generation built in.
If you're unsure whether your account can generate images, type a quick test: "Generate a simple image of a red circle on a white background." If you get an image back, you're ready. If you get a text-only response explaining limitations, check your subscription tier.
Step 2: Upload your reference photo
Click the attachment icon (paperclip) in the chat input and select your reference photo. You'll see a thumbnail appear in the message composer.
One thing to check before you proceed: do you have the rights to use this photo? If it's your own selfie, you're fine. If it's someone else's photo — a celebrity, a friend, a photo you found online — you need that person's permission or a license that covers AI use. ChatGPT will also refuse to generate images of real public figures, so use your own photo or one you have clear permission to use.
Step 3: Paste the prompt and send
Here's the prompt we'll use. It's a tested recipe for a cinematic car mirror selfie — a look that went viral on social media throughout 2024 and 2025. The full recipe is available free on AI Portrait Prompts, and you can copy it directly:
A cinematic car mirror selfie portrait of a [GENDER], age [AGE], [HAIR_COLOR] hair, wearing [OUTFIT]. Shot through the side mirror of a car, golden hour lighting reflecting off the glass. Shallow depth of field, 85mm lens equivalent, f/1.8. The mirror frame creates a natural vignette. Warm tones, film grain texture, Kodak Portra 400 color science. The subject's expression is [EXPRESSION]. Background: blurred urban street with bokeh lights.
Paste this directly into the chat input, below your uploaded photo. Don't send it yet — we'll customize the variables first.
Step 4: Replace the bracketed variables with your details
The words in square brackets are the customizable parts. Replace each one with details that match you or the look you want:
| Variable | What it controls | Example values |
| --- | --- | --- |
| [GENDER] | Subject gender | woman, man |
| [AGE] | Subject age | 25, 30, 45 |
| [HAIR_COLOR] | Hair color | black, blonde, brunette |
| [OUTFIT] | Clothing description | leather jacket, white t-shirt, oversized blazer |
| [EXPRESSION] | Facial expression | confident smirk, soft smile, intense gaze |
Here's what the prompt looks like filled in for a specific person:
A cinematic car mirror selfie portrait of a woman, age 28, brunette hair, wearing a leather jacket. Shot through the side mirror of a car, golden hour lighting reflecting off the glass. Shallow depth of field, 85mm lens equivalent, f/1.8. The mirror frame creates a natural vignette. Warm tones, film grain texture, Kodak Portra 400 color science. The subject's expression is a soft smile. Background: blurred urban street with bokeh lights.
Keep the lighting, lens, and color instructions exactly as written — those are the parts that make the result cinematic. Only change the variables in brackets. This separation is the whole point: the camera and lighting direction stay stable while you customize the subject.
Step 5: Send and wait for generation
Hit send. ChatGPT will process your photo and prompt together and generate a portrait. This typically takes 15–45 seconds depending on server load.
What you should see: An image that looks like it was shot through a car side mirror during golden hour, with warm tones, a soft blurred background, and film-like color. Your reference photo's facial features should be recognizable but transformed — same person, different light, different mood, different camera language entirely.
What you should not see: A flat, even-lit selfie with the contrast boosted. If you get that, the model ignored part of your prompt. Resend the message — occasionally the first generation doesn't fully apply the lighting instructions.
Step 6: Evaluate and refine
Rarely does the first generation hit exactly right. Here's how to diagnose common problems and fix them by adjusting the prompt, not by re-rolling blindly.
The lighting looks flat, not golden hour. The model may have prioritized your reference photo's existing lighting over the prompt's instructions. Add emphasis: "Strong warm golden hour light streaming from camera-left, casting long shadows." Naming a direction (camera-left, camera-right) gives the model less room to default.
The face doesn't look like you. This is the hardest problem. AI image generators balance two forces: fidelity to your reference photo and adherence to the prompt's style instructions. If the style instructions are too aggressive (too many adjective layers, too much "cinematic dramatic moody intense"), they overwhelm identity. Try simplifying: remove adjectives and keep only the concrete camera and lighting instructions. Generate again and compare.
The background is wrong or distracting. Be more specific about what you want: "Background: out-of-focus city street at dusk, no people, no text, no signs." Negative descriptions (what not to include) help ChatGPT's image model filter unwanted elements.
The overall look is too clean, too digital. The "film grain texture, Kodak Portra 400 color science" instruction is what adds the analog warmth. If it's missing, check that you pasted the full prompt — that line is easy to accidentally trim when copying.
Step 7: Download and use your portrait
Once you have a generation you're happy with, click the image to open it full-size, then use the download button (downward arrow icon) to save it.
For social media profiles, crop to a square or circle before uploading — the full image may have a landscape aspect ratio that doesn't fit a profile frame. For Instagram or other feeds, you can use the image as-is.
On commercial use: If you plan to use the generated portrait in a client project, marketing material, or paid work, you need to check two things. First, OpenAI's usage terms for generated images (currently, paying ChatGPT users have broad commercial rights to images they generate). Second, if the photo isn't yours, you need permission from the person depicted. The prompt itself doesn't grant or override any rights — it's the tools and source materials that carry the licensing weight.
Why this approach works (and copy-paste from a list doesn't)
You can find thousands of "cinematic portrait prompts" floating around social media. Most are keyword dumps — long lists of style words with no structure: "cinematic, dramatic, moody, film grain, 8k, hyperrealistic, professional photography, bokeh, depth of field." These produce unpredictable results because they give the model a pile of adjectives with no instruction on how to balance them.
The structured prompt works differently. It separates the stable layer — lens, lighting direction, color science, composition — from the variable layer — subject, outfit, expression. This separation is why the same prompt produces consistent, recognizable results across different reference photos. You're not re-inventing the lighting direction every time; you're swapping the subject into an existing visual framework.
That's also why a collection of tested, structured prompts is more valuable than a list of keywords. Each prompt is a pre-built camera and lighting setup you can drop your own subject into. The free prompts on AI Portrait Prompts follow this structure, and the premium collection extends it across cinematic, fashion, social media, and artistic categories — each with the same split between stable direction and customizable variables.
Where to go next
Once you've run the car mirror selfie prompt and understand how the variable replacement works, you can apply the same process to any structured portrait prompt:
Fashion editorial prompts swap the golden-hour car-mirror setup for studio lighting and designer styling references.
Professional headshot prompts use cleaner Rembrandt lighting and neutral backgrounds for LinkedIn-ready results.
Artistic prompts introduce double exposure, watercolor, or painterly treatments that go beyond photography.
The workflow stays the same every time: upload a reference photo, paste the prompt, replace the bracketed variables, generate, refine. The prompt handles the camera work so you don't have to.
If you want more tested recipes with before-and-after previews so you can see exactly what each prompt produces before running it, browse the full collection. Five prompts are free to copy; the rest are available in the premium collection with a commercial use license.
Quick troubleshooting
"ChatGPT says it can't generate images." Your account likely doesn't have image generation access. This requires ChatGPT Plus or Team. Check your subscription at chatgpt.com/settings.
"The result doesn't look like the before-and-after examples." The single biggest factor is your reference photo. A photo shot in flat indoor lighting gives the model less to work with than one with existing directional light. Try a photo taken near a window during daytime.
"The prompt was cut off when I pasted it." Some chat interfaces truncate very long pasted text. If the prompt appears cut off, paste it in two parts or use a plain text editor to copy the full text cleanly, then paste into ChatGPT.
"Can I use this with Midjourney or DALL-E instead of ChatGPT?" Yes, with adjustments. Midjourney uses a different prompt syntax (parameters like --ar 3:2 --v 6 instead of natural language descriptions) and doesn't accept reference photos the same way. DALL-E inside ChatGPT works similarly to the workflow above. The prompt collection includes compatibility labels so you know which tool each recipe was built for.
