If You've Finished Fourth Wing and Miss Basgiath, This Fan Site Is Worth a Bookmark

 If you tore through Rebecca Yarros's The Empyrean series — Fourth Wing, Iron Flame, Onyx Storm — and closed the last page wishing there was somewhere to keep exploring the dragon vocabulary of Basgiath War College, you're not alone. The fandom has been quietly building its own corners of the internet, and one of the more thoughtful ones I've come across is FourthWingDragon.com.

It's an independent fan space — not affiliated with the author or publisher — that treats the dragon lore of Fourth Wing as something worth mapping, not just gushing over. No accounts, no sign-ups, no spammy pop-ups. Every tool runs in your browser tab and disappears when you close it.

What's actually on the site

The site is built around six sections, and the way they fit together is what makes it more than another fan blog:

• Dragon Types — A seven-by-six reference grid covering the color categories (Black, Blue, Green, Red, Orange, Brown, Iridescent) and tail labels (Daggertail, Morningstartail, Scorpiontail, Swordtail, Clubtail, Feathertail). It's a navigation tool, not a claim that every combination shows up in the books — and the site is honest about that distinction.
• Dragon Type Quiz — Twelve questions about everyday choices, mapped to seven color personality profiles. The result is a fan interpretation, not canon, and the site says so plainly.
• Signet Quiz — A separate quiz that walks you through seven creative ability themes, in the spirit of the signets cadets develop at Basgiath.
• Dragon Name Generator — Combines invented syllables with Gaelic roots to produce names you can use for fanfiction, roleplay, or just for fun.
• Dragonkind Guide — The section I appreciate most: it cleanly separates what's official from what's fan-made, and points readers back to the author's own introductions rather than paraphrasing lore.
• Threshing Guide — Context for Threshing Day, the moment cadets first bond with a dragon, written to complement the books rather than spoil them.

There's also a small character section with profiles like Aaric Graycastle, Rhiannon Matthias, and Ridoc Gamlyn — pulled from the publisher's own cadet descriptions.

Why I keep going back to it

A lot of fan sites blur the line between official lore and fan invention until you can't tell what's Rebecca Yarros and what's someone on Reddit at 2 a.m. FourthWingDragon draws that line on purpose. The reference grid is presented as a taxonomy, the quiz results are labeled as creative interpretations, and the guide section actively routes you back to the source material when you want the real thing.

That makes it useful in two different ways: as a quick reference while you're reading (especially between Iron Flame and Onyx Storm, where the dragon roster expands fast), and as a low-stakes creative toy when you just want to kill ten minutes finding out what color dragon would theoretically choose you.

Who it's for

Readers mid-series who keep forgetting which tail label belongs to which dragon. Fanfic writers who need a name that doesn't sound made-up. Anyone who finished Onyx Storm and wants one more small window into Basgiath before the next book arrives.

If that's you, the whole thing is free and lives at FourthWingDragon.com. No account, no email, no upsell — just open a tab and start exploring.

Why the old trick stopped being enough

For about twenty years, "removing vocals from a song" meant one trick: invert the phase of one stereo channel, fold both channels back to mono, and anything sitting dead center cancels itself out. It works. It is also why so many home-made karaoke tracks sound like they are being played through a wall.

Two things break. The lead vocal is not the only thing mixed to center, so the kick drum and the bass go with it. And anything that is not perfectly identical in both channels, meaning reverb, delay, doubled takes and backing harmonies, stays exactly where it was. You get a thin backing track with a ghost of the singer still in it.

Modern source separation does something categorically different. A model like Demucs is trained on material where the individual parts are already known, so it learns what a voice actually sounds like against drums, bass and everything else. Instead of cancelling a frequency band, it estimates which parts of the mix belong to which instrument and rebuilds clean files.

The practical rule: if you have a mono-ish 1960s recording and you want a rough backing track in two minutes, phase inversion is still free and still fine. If the song is a real stereo mix, and especially if you want a vocal you can actually reuse, you need a model.

Four things to check before you upload anything

Almost every free audio tool advertises "free" and buries the qualification. These four checks decide whether a tool is usable for your file, and none of them are usually on the landing page.

• Length limit, not file size. Most free tiers cap duration rather than megabytes. A three-and-a-half-minute single will be rejected by a two-minute cap no matter how small the file is. Decide first whether you need the whole song or just the chorus.
• Free preview versus free download. Several well-known services let you listen for free and charge to save the file. Listening is not the same product as keeping.
• Stem count. Two stems means vocals plus instrumental. Four or six stems means drums, bass, guitar and piano separated as well. If you want the drum track on its own, a two-stem tool is simply the wrong product, however good it sounds.
• Retention. Upload-based tools store your file temporarily. If the retention window is an hour, that is fine, but download before it expires instead of assuming the link will still work tomorrow.

The free route I actually use: getvocalremover.com

I keep coming back to Get Vocal Remover for one reason: it states its limits in plain numbers before you upload, instead of after you have waited ninety seconds. It runs the Demucs htdemucs model through Replicate, which is the same class of model the paid tools use, and it needs no account.

Step 0: prepare the file first

Trim to the section you actually need. A free upload is capped at two minutes and 50 MB, so exporting a 30-second chorus from your DAW takes ten seconds and saves you a rejected upload. Use a clean stereo export rather than a phone recording of a speaker, and avoid heavily clipped masters, because clipping gives the model less to work with.

Step 1: upload

Open the vocal remover, drop in an MP3, WAV, M4A or AAC file, and press split. Everything happens server-side, so this is an upload, not on-device processing, and it is worth knowing that before you send anything sensitive.

Step 2: wait for the split

Separation takes seconds to about a minute on a short file. The site produces two files, not four: an instrumental, which is drums, bass and other instruments kept together, and a vocal stem, which is the acapella-style version. The two stay time-aligned, which matters more than it sounds. You can switch between them without hunting for where the chorus starts.

Step 3: preview, then download

You can listen to both stems as many times as you like. The free allowance is three splits per day per IP, with the first successful split of the day including both downloads. Later free splits on the same day unlock the files with Credits instead of running the model again, so a re-download does not cost a second separation.

Step 4: take the file out of the building

Uploads and results are deleted automatically after roughly an hour. That is short-lived storage, not a no-upload promise, and it is the reason you should download immediately rather than plan to come back later.

Where this tool stops being the right answer

Being clear about the edges is more useful than another paragraph of praise:

• Two stems only. Drums, bass, guitar and piano stay inside the instrumental. If the model can estimate more, this product does not expose it, and that is documented openly on its about page.
• MP3 output. Fine for practice, not the format you want for a release or a mastering chain.
• Artifacts on difficult material. A centered lead vocal over a clean stereo master gives the most convincing result. Dense effects, doubled vocals, live rooms and loud cymbals can leave small traces of music in the vocal file, or a little breath in the instrumental. If your result still has the singer in it, the guide on why vocals are still present explains which parts are fixable and which are baked in.
• No lyrics, no key change, no arrangement editing. It is a splitter, not a karaoke studio. If you want lyrics on screen, that happens in a player or an editor afterwards.

Three ways to do this, ranked by effort

Method	Setup	Result on a centered vocal	What breaks	Best for
Phase inversion in Audacity	Five minutes, free, desktop install	Mediocre	Loses kick and bass, keeps reverb and doubles	Rough diagnostics on mono-ish old recordings
AI two-stem online (this tool)	About a minute, no install, no account	Good on clean stereo masters	MP3 only, two stems, some cymbal bleed	Practice backing tracks, one-off previews
Paid multi-stem or DAW plugin	Install plus a licence	Best, and per instrument	Cost, learning curve, still not the original session	Release work, remix and sampling

Five uses that actually justify the effort

1. Karaoke and cover practice. Keep the instrumental, drop the original, sing over a backing you already know. Use the vocal file to check entrances and phrasing rather than guessing.
2. Remix and arrangement tests. Solo the vocal against a new kick and bass before committing to a rebuild. If the phrase does not sit, you find out in a minute instead of an afternoon.
3. DJ transitions. Preview a blend with the instrumental, then bring the acapella in for a call-and-response moment. Same timeline on both files makes cueing straightforward.
4. Teaching and analysis. Compare original and instrumental to hear how the bass supports the kick, or how a guitar fills the space after a chorus line. Solo the vocal to study breath and phrasing.
5. Voice checks on spoken audio. A podcast or interview with a music bed can be inspected quickly. It is not a dialogue editor, and it will not remove every trace, but it tells you what you are working with.

The paragraph most tutorials skip

Removing a vocal does not give you rights to anything. The recording, the composition and the performance stay somebody else's, and an instrumental is still a derivative of a copyrighted track. For private practice, a classroom exercise or a rehearsal, you are on safe ground with material you own or are licensed to use. For anything public, meaning a YouTube upload, a paid gig, an ad or a release, clear the rights first, because automated content matching does not care that the singer is missing. The tool is explicit about this, and it is the one limitation that no amount of model quality will fix.

When the free tier is not enough

The two-minute cap is the real constraint. A full song needs paid minutes unless you work in excerpts, and if you only ever need one chorus, the free tier covers it indefinitely. If you need full-length files repeatedly, Credits are sold as one-time packs rather than a subscription: Starter at 9.90 for 50 minutes, and Pro at $19 for 120 minutes, with paid uploads up to 10 minutes, minutes that do not expire, and failed jobs refunded. Every job is rounded up to the next minute, so a 3:10 song costs four. That pricing model is the honest one for a tool you use in bursts.

Try it on one chorus

Pick a song you have permission to use, export a 45-second section, and run it through the karaoke maker path or the main uploader. Judge the result on the vocal file, not the instrumental, because the instrumental flatters almost any separator. If the vocal is clean enough for your purpose, you have your workflow. If it is not, you now know exactly which of the four checks above you need to trade up on, and you have spent zero dollars finding out.

Why the cover matters on Xiaohongshu

A Xiaohongshu cover has a demanding job. It needs to communicate the subject quickly, create a recognizable mood, and leave enough visual space for a short headline. A beautiful image is not automatically a useful cover: if the subject is unclear or the composition is too busy, the post can be difficult to understand in a crowded feed.

The strongest prompts describe the purpose of the image, not just its decoration. Instead of writing “a pretty cafe,” specify the intended audience, the focal subject, the mood, the composition, and the empty space needed for text. This makes the result easier to adapt into a real post cover.

A simple prompt formula

A practical cover prompt can follow this structure:

Subject + visual style + composition + color direction + text space + format

For example:

A cozy weekend cafe guide for young urban readers, warm editorial photography, a cup of iced coffee and an open notebook on a wooden table, soft morning sunlight, clean negative space in the upper left for a headline, vertical social media cover composition.

You can replace each part without rewriting the entire prompt. Keep the subject specific, and treat “negative space” as a design requirement rather than an afterthought.

Xiaohongshu cover prompt examples

1. Lifestyle checklist cover

A practical morning routine checklist for busy professionals, bright natural lifestyle photography, neatly arranged notebook, coffee, headphones, and a small plant on a light desk, fresh cream and sage green palette, clear negative space at the top for a short title, clean vertical cover composition.

This prompt works when the post is organized around habits, routines, or a simple list. The objects suggest the topic without making the image feel like a literal infographic.

2. Travel guide cover

A three-day city travel guide, cinematic editorial travel photography, a traveler walking through a lively historic street with colorful storefronts, soft afternoon light, balanced composition with open sky on the upper third for text, vivid but natural colors, vertical social media cover.

For travel content, a recognizable setting is more valuable than a generic “beautiful vacation” scene. If the destination is important, name it directly in the prompt and check the generated image for visual accuracy before publishing.

3. Food recommendation cover

A weekend brunch recommendation, appetizing overhead food photography, pancakes, fresh fruit, coffee, and linen napkin arranged on a warm stone table, inviting sunlight, uncluttered central composition with a small empty area for a headline, warm beige and berry color palette, vertical cover format.

Food covers benefit from a close focal point and controlled styling. Asking for too many dishes or props often makes the result look crowded and reduces the space available for typography.

4. Beauty routine cover

A beginner-friendly evening skincare routine, clean beauty editorial photography, three minimal skincare bottles beside a folded towel and a small ceramic dish, soft diffused bathroom light, pale pink and ivory palette, generous clean space on the right for text, polished vertical cover composition.

The phrase “beginner-friendly” gives the image a different feeling from a luxury campaign. Use the audience and the promise of the post to guide the visual tone.

5. Fashion outfit cover

A simple office outfit idea for early autumn, modern street-style editorial photography, full-body subject wearing a tailored blazer, relaxed trousers, and loafers, neutral charcoal, cream, and brown palette, uncluttered city background, open space near the top for a short title, vertical social media cover.

For fashion posts, specify the type of framing you need. A full-body composition is useful for outfit posts, while a closer crop may work better for accessories or makeup details.

6. Educational carousel cover

A beginner guide to personal budgeting, modern editorial illustration, notebook, calculator, simple receipts, and a small growing plant arranged on a clean light background, calm blue and yellow accent colors, strong visual hierarchy, wide empty area at the top for a clear headline, vertical cover design.

Educational content usually needs more text space than inspirational content. Ask for visual hierarchy and a clean background so the final headline remains legible.

How to make generated covers more usable

Write the headline separately whenever possible. Image generators may produce distorted or unreadable lettering, so it is usually more reliable to generate the background first and add the final text in a design editor. Keep the headline short, use one clear promise, and make sure it describes the post rather than the picture.

Choose one dominant subject. If the prompt asks for a desk, a meal, a person, a city skyline, and several decorative objects at the same time, the image may lose its center of attention. A strong cover normally needs one main idea and only a few supporting details.

Also define the emotional temperature. Words such as “calm,” “playful,” “practical,” “premium,” or “energetic” are more useful when they match the post. A post about a quick beginner tutorial may need a clear and friendly visual language rather than a dramatic luxury aesthetic.

If you save a useful reference image from Xiaohongshu or RedNote, you can use it as a visual starting point for your own cover direction. Study its layout, color balance, and subject placement, then create an original composition instead of copying the image or its text.

FAQ

What makes a good Xiaohongshu cover prompt?

A good prompt explains the post’s subject, audience, mood, composition, color direction, and space for text. It describes a usable cover rather than only asking for a beautiful image.

Should I include the title inside the image prompt?

You can describe where the title should go, but it is usually better to add the final wording afterward. This gives you more control over spelling, typography, and mobile readability.

What aspect ratio should I use?

Use a vertical composition designed for a mobile feed. The exact export dimensions can depend on your publishing workflow, so check the platform preview and keep important subjects away from the edges.

Can I reuse one prompt for every post?

Use a consistent structure, but change the subject, audience, mood, and composition for each post. Reusing the same visual formula too closely can make an account feel repetitive.

How many prompt variations should I generate?

Start with three to five variations that change the framing or mood while keeping the topic consistent. Select the version that leaves the most useful space for your headline and remains clear at a small size.

What Are AI Portrait Prompts? A Complete Beginner's Guide

 If you've seen stunning AI-generated portraits on social media and wondered how people get those results, the answer is almost always one thing: a well-written prompt.

An AI portrait prompt is simply the text instruction you give an image generator — Midjourney, DALL-E, ChatGPT, or any other model — that tells it what kind of portrait to create. Sounds simple, but the difference between a generic AI image and a cinematic, photorealistic portrait is entirely in how you structure those words.

How prompts work

Image models don't "understand" your words the way a human does. They translate text into a set of visual patterns learned from millions of images. That's why vague prompts give vague results: type "portrait of a woman" and you get the average of every portrait of a woman the model has ever seen. Generic in, generic out.

The fix is to be specific about the things a photographer would care about:

• Lighting — "golden hour backlight," "soft studio lighting," "moody low-key light"

• Camera language — "85mm lens," "f/2.8," "shallow depth of field"

• Setting — "rainy street at night," "vintage film studio," "minimal gray background"

• Mood and color — "warm amber tones," "teal-orange grade," "film grain"

The anatomy of a good prompt

Most effective portrait prompts follow a structure like this:

[Lighting] + [subject detail] + [camera/lens language] + [mood/color grade]

Example:

"Golden hour backlight portrait, rim light glowing through hair strands, 85mm lens, f/1.4, soft bokeh background, warm amber tones, photorealistic, cinematic film still"

Every element has a job: the backlight separates the subject from the background, the lens specs control depth of field, the color direction handles the mood.

What the big generators do differently

Midjourney responds best to descriptive, style-heavy prompts and rewards camera language. DALL-E is more literal — it follows multi-part instructions precisely. ChatGPT's image model (GPT-Image) combines DALL-E's fidelity with conversational iteration: you can say "now make it golden hour" and it adjusts the image.

Common beginner mistakes

1. Over-describing the face ("beautiful, perfect skin, big eyes") — this pushes models into uncanny territory

2. Ignoring aspect ratio — portraits look dramatically better at 2:3 or 9:16 than square

3. Forgetting skin texture — adding "natural skin texture" is the single easiest realism boost

4. Writing prompts from scratch every time — tested recipes will beat your improvisation 9 times out of 10

The bottom line

You don't need to be a prompt engineer to get great AI portraits. You need a structure that works, and you need to stop guessing. Tested recipes, used consistently, will outperform spontaneous prompt-writing almost every time — and you'll see the difference the first time you compare side by side.

How to Generate Cinematic AI Portraits with ChatGPT: A Complete Beginner's Guide

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.