How to Make an AI Roast Video (Get Roasted on Stage From One Photo)

What You're Trying to Make

You, in the hot seat of a celebrity-style roast: a comedian at the podium delivers a savage burn written from your photo, the crowd erupts, and you react — laughing it off, jaw on the floor, or cringing through your fingers. The catch is nobody wrote a script and no comedian ever met you. The whole clip — the jokes, the delivery, your reaction — is generated from one photo.

This is the video upgrade to the "AI roast me" trend. This guide covers how to make one: the photo to use, how to steer the jokes, the prompt to write if you're rolling your own, which model to run it on, and how to post it.

Fastest way — the The Roast template does this from one photo: it reads your picture, writes two savage burns, picks the reaction that matches each one, and hands back a 12-second hot-seat roast with crowd laughter in a few minutes — no prompt writing. Want to roll your own or push the spice yourself? The full method's below. ↓

Is "AI Roast Me" Actually a Thing?

It's one of the most durable AI formats going — less a one-week spike, more an evergreen "make me laugh" category. The original version is text: upload a selfie, type "roast me," and a chatbot fires back jokes about your appearance, outfit, and vibe. There's a whole ecosystem of it — ChatGPT picture roasts, "roast me" GPTs, roast-me web tools, and the AI Roast Me filter trend on TikTok and CapCut. The tells that demand is real: dedicated roast generators rank for the keyword, the tools advertise intensity dials ("Gentle Tease" to "Absolutely Brutal") and roast styles, and the screenshots travel because they're surprisingly accurate and extremely shareable.

The natural next step is video. A text roast lives as a screenshot; a roast video — the burn spoken out loud on a stage, the crowd reacting, you in the hot seat taking it — is the format built for the For You page. Same joke, way more shareable.

The Fastest Way — Use The Roast Template on Starrd

The Roast template is live in the Starrd library. It packages every step in this guide — reading your photo, writing the burns, picking the reaction, staging the dais, and generating the crowd audio — into a single upload.

1. Pick a clear photo. Your face and outfit visible; bonus points for stuff in the background to roast.
2. Open The Roast template in the Starrd app or web library.
3. (Optional) Add director's notes. Leave it blank for a savage default, or steer it: "keep it playful," "go brutal," "roast my gym obsession."
4. Upload and tap generate. The template writes the jokes from your photo, picks a matching reaction (laughing / mock-offended / cringing), and generates a 12-second roast on Seedance 2.0 — comedian, podium, crowd, and the burns delivered out loud.

100 credits, a few minutes. No prompt writing.

The rest of this guide is for people who want to roll their own — write their own burns, change the setting, or run it on a different model.

Or, Build It Yourself — What You Need

Three things:

1. A clear photo of the subject. Face and outfit visible. A busier photo (background, props, company) gives you more material.
2. An AI video model that accepts a reference image and generates speech. Seedance 2.0 is the safest — it follows the prompt closely and generates the spoken jokes and crowd laughter as audio.
3. A platform to post on. TikTok, Reels, and Shorts.

You'll also need the jokes themselves — either write them, or have a chatbot roast the photo first and paste the best two lines into your video prompt.

Step 1 — Pick Your Reference Photo

The photo does double duty here: it's the face that ends up in the hot seat and the source material for the jokes. Pick one with stuff to work with.

Use:

• A clear, well-lit photo of one person, front or three-quarter angle, eyes open
• An outfit/expression with personality (the more "a choice," the better the burn)
• A background with detail — your room, your car, food, the people you're with

Avoid:

• Sunglasses or anything hiding your face (you lose the likeness in the hot seat)
• Low-resolution or blurry shots
• A totally plain backdrop (less to roast)
• AI-generated reference images (compounding artifacts)

Step 2 — Write the Burns (or Have AI Write Them)

This is the part that makes or breaks it. Two short, punchy lines — under ~18 words each — that hit what's actually visible.

• Punch at choices, not people. Roast the fit, the pose, the "main character" energy, the haircut, the messy desk behind you. Skip anything about race, weight, disability, or religion — that's not savage, it's just mean, and it gets clips taken down.
• Mine the whole frame. "Nice setup — is that a gaming chair or a throne for someone who peaked in 2019?" lands harder than a generic looks joke.
• Pick a spice level. Savage by default; dial to playful for an affectionate roast, or crank it for a brutal one.

No instinct for it? Upload the photo to any chatbot, ask it to "roast this photo, two short savage one-liners," and use the best two.

Step 3 — Write the Prompt

A roast clip lives on the burn landing, then the reaction. Structure it as setup → joke → reaction → bigger joke → crowd explosion. Copy this and drop your two burns in:

Glossy multi-camera comedy-special broadcast on a celebrity-roast dais. The subject sits in the "hot seat" — a plush leather chair with a name placard, a drink in hand — while a sharp-suited comedian at a podium delivers savage burns. A laughing crowd and celebrity panel sit in soft focus behind. Warm spotlights, theater haze, lens flares, rich warm color grading.[00:00-00:03] Wide establishing shot of the dais. The subject lounges in the hot seat looking smug, sipping the drink. The comedian steps to the podium and taps the mic. Slow push in.[00:03-00:06] Medium shot on the comedian, leaning into the mic with a grin. Comedian says: "[YOUR BURN #1 — savage, under 18 words]." Subtle rack focus from the comedian to the subject behind.[00:06-00:09] Cut to a close-up of the subject reacting — [match it: head-back laugh / jaw-drop, hand over mouth / cringing through fingers]. The crowd starts to erupt, a panelist points and laughs. Handheld energy.[00:09-00:12] The comedian lands the closer. Comedian delivers: "[YOUR BURN #2 — the punchline]." The crowd explodes in laughter and applause; the subject shakes their head, mouths "wow." Pull back to a wide of the roaring room.Audio: the comedian's amplified voice, mic presence, crowd murmur building to roaring laughter and applause, the occasional "ohhh!", an upbeat comedy-special sting. Generate audio.Natural skin texture, real stage-light color, believable broadcast capture. Keep the subject's face consistent throughout.

The non-negotiable elements:

• The joke, spoken out loud. A roast video where you can't hear the burn is just someone sitting in a chair. Keep the dialogue lines and "Generate audio" in.
• A reaction that matches the joke. Savage burn → jaw-drop or cringe. Silly burn → laugh-along. A mismatched reaction reads as fake instantly.
• Crowd that erupts. The room reacting is what tells the viewer the joke landed. Vague crowds kill it.
• You in the hot seat. Drink, name placard, smug-then-roasted — the dais framing is what makes it a roast and not just stand-up.

Step 4 — Pick a Model

• Seedance 2.0 — what Starrd runs on. Strongest prompt adherence, native 12s, and it generates the spoken jokes and crowd laughter as audio. Safest pick for this format (the spoken burn is the whole point).
• Kling 3.0 — great realistic on-camera performance and expression; strong for the reaction beat. Check its speech support for the spoken lines.
• Runway Gen-4 — solid all-rounder; you may need extra language to push crowd reaction and energy.
• Veo 3.1 — capable with synced audio; reinforce the crowd-laughter and spoken-dialogue cues.

No preference? Start with Seedance 2.0. (See our Seedance vs Kling vs Veo breakdown.)

Step 5 — Generate and Iterate

First generations rarely nail it. Common failures and fixes:

You can't hear the joke / no dialogue. Reassert it: 'Comedian says: "..."' as an explicit spoken line, and keep "Generate audio" at the end.

The reaction doesn't match the burn. Spell it out in the same segment: savage line → "jaw drops, hand over mouth"; silly line → "head tipped back laughing."

The crowd is dead. Add: "crowd erupts in roaring laughter and applause, panelists pointing and cracking up."

It reads as stand-up, not a roast. Re-anchor the dais: "subject seated in the hot seat with a name placard and a drink, comedian at a separate podium."

The face drifts from the reference. Use a clearer front-facing reference, or weight the reference image more heavily if the model supports it.

Budget 3-5 generations before a keeper.

Step 6 — Post It

Lead with the burn. Cut in right before the joke lands so the punchline + reaction hit in the first 2-3 seconds — that's the scroll-stopper.

Caption it as a moment. "I let an AI roast me and it was too accurate 💀" travels further than "I made this with AI." (Within each platform's required AI-disclosure labels.)

Tag the victim. If you roasted a friend, posting it tagged (or dropping it in the group chat) is half the format's reach.

Common Mistakes That Tank Your Video

1. A muted or unspoken joke. The burn has to be heard — keep the dialogue and audio in.
2. A reaction that doesn't fit the joke. Mismatched emotion is the #1 tell that it's fake.
3. Generic looks-only burns. "You're ugly" is boring. Roast the specifics — the fit, the pose, the room.
4. Crossing the line. Protected-trait jokes aren't savage, they're just mean — and they get clips removed.
5. No hot-seat framing. Without the dais, placard, and crowd, it's just stand-up.

Related Reading

• Seedance 2.0 Prompt Guide — the full framework for writing AI video prompts that don't look like AI video
• Seedance vs Kling vs Veo — which model to pick and why
• How to Make an AI DJ Video of Yourself — another "put yourself somewhere you weren't" performance format
• Viral AI Video Trends (2026) — the running roundup of what's worth making right now