How to Make a Korean Baseball AI Video: Step-by-Step Prompts (May 2026)

The Korean baseball AI trend — the one with the candid "fan cam goddess" looking surprised in the stadium stands — exploded across Threads, X, and TikTok in May 2026. By the time you're reading this, a single Threads tutorial post hit thousands of views and the HitPaw guide racked up ten million views on the example clip.

If you've tried to copy it and gotten generic-looking AI faces or a stiff zoom-in, you're hitting the same wall everyone else does. The trick is two prompts, not one — one for the static image, one for the Kling animation — and both have to fight the model's default beautification instinct.

This guide gives you both prompts in full, exactly as the community is using them, plus the one-click shortcut if you don't want to stitch the steps yourself.

What you need

• One clear photo of the person you want in the stadium. Front-facing or 3/4 profile works best. Don't use a glamour shot — the trend is about looking caught, not posed.
• A static-image AI that does image-to-image with identity preservation. ChatGPT (GPT-4 image), Gemini, or Flux Kontext all work.
• Kling 3.0 for the image-to-video animation step. Or skip to the one-click route below.

Step 1 — Generate the static image (the "anti-beauty" prompt)

The static image is where most attempts fail. The default AI face is too smooth, too symmetrical, too "influencer." The viral examples look like an actual broadcast accidentally caught a real person.

The dominant prompt structure has two parts: an identity anchor that locks the face, then an anti-beauty clause that tells the model to stop polishing.

Paste this into ChatGPT or Gemini with your reference photo:

Use the uploaded reference image as the strongest identity anchor.
The person must look like the exact same person from the reference,
not just a similar Korean face. Preserve their exact facial identity
with high priority: same face shape, same eye spacing, same nose,
same lip shape, same skin tone, same hair.

Create an ultra-realistic candid KBO baseball broadcast screenshot of
the same person accidentally caught by a live TV camera in the
spectator seats. Seated among a lively Korean baseball crowd, wearing
a clean baseball jersey, holding a cheering stick or iced drink. They
notice the camera and give a small natural smile — slightly surprised
but composed.

Important: No face reshaping, no enlarged eyes, no jaw editing, no
overly smooth or glassy skin, no influencer or photoshoot vibe. Must
feel like a real accidental broadcast capture of an ordinary
spectator that went viral online. Slight blur from live-stream
quality, compression noise, realistic skin texture, imperfect
stadium lighting.

A few things to know before you paste:

• The identity-anchor opening ("strongest identity anchor", "exact same person, not just similar") is the single most-replicated structural pattern across the trend. The Synthia/X version of this prompt is what most English-language tutorials end up rewriting.
• The anti-beauty paragraph is the contribution of The Tab's editorial — explicitly telling the model not to do its usual cosmetic pass. Without it, you get an AI beauty filter.
• Team jersey is a swappable parameter. The Fanplus community post that kicked off the Korean side of the trend recommends naming a specific KBO team. Try "LG Twins", "Lotte Giants", "Doosan Bears", "Samsung Lions", or "KIA Tigers" — the specificity makes the jersey design tighter.
• For horizontal output, append make the photo horizontal, 16:9 composition. The ezday community tutorial recommends this because vertical crops crop out the surrounding crowd, which kills the "caught by broadcast" feel.

If the first generation looks too clean, regenerate with harsher stadium lighting, more crowd visible, less makeup, broadcast camera grain appended.

Step 2 — Animate with Kling (the I2V prompt)

Upload your generated image to Kling AI and switch to image-to-video. The video prompt that's working right now is short and physical:

Realistic sports broadcast video of a person in a crowd. Camera
slowly zooms in as they are unaware of being on camera. They adjust
their hair or look around casually watching the match, or sip a
drink, all in a natural casual way. Subtle blinking, slight
handheld camera shake, shallow depth of field, SPOTV-style Korean
broadcast aesthetic.

A longer alternate that the HitPaw guide uses:

Hyper-realistic live KBO baseball broadcast video. The person slowly
turns their head, smiles slightly, waves a cheering stick, subtle
blinking, realistic skin texture, flyaway hair, mild sweat and shine
under stadium lights, slight handheld camera shake, shallow depth of
field, dramatic night stadium lighting, accidental crowd cam moment,
extremely lifelike, SPOTV broadcast aesthetic.

Why this works:

• "Camera slowly zooms in" + "unaware of being on camera" is what sells the accidental feel. The viral examples don't have the subject pose for the camera — they react to it.
• "Subtle blinking", "slight handheld camera shake" tell Kling to stop smoothing the motion. Without these, you get a too-clean dolly zoom.
• "SPOTV" references the actual Korean KBO broadcast channel. Kling has clearly trained on Korean sports broadcast footage and responds to the name.

Set duration to 5 seconds. Longer durations let Kling drift the face — and the whole point of this trend is the face stays identifiable.

Step 3 (alternative) — One-click on ClipTrend

If you don't want to run two separate tools, the Korean Baseball generator on ClipTrend wraps both steps into one upload. You paste your photo, and it returns a finished 5-second clip — same Kling 3.0 backend, same identity-anchor prompt baked in.

This is the funnel KBAT was built around. The four Kling effects on the homepage (Korean Baseball, Raid Check, Sassy Shake, Toss & Run) each deep-link into the corresponding ClipTrend template with the prompt and aspect ratio pre-tuned.

Pro tips from the community

A few things that didn't make the prompts above but separate the great clips from the okay ones:

• Match the time of day in the image to the lighting in the video. A day-game image with night-game lighting cues looks off. The strongest examples are night games — stadium lights give the face a more cinematic key light.
• Don't crop the crowd out. The viral feel is crowd as context. A tight headshot doesn't read as "broadcast caught them" — it reads as "AI portrait of a person in a jersey."
• One person per frame. Two people in the stands confuses the identity anchor and Kling smears the second face.
• For couples or groups, generate each member separately and stitch — don't try to do them in one shot.

Common mistakes

FAQ

Do I need a Korean photo for the Korean baseball trend? No. The trend is the aesthetic, not an ethnic constraint. Any face works — the identity anchor preserves whoever you upload, and the jersey + stadium do the cultural framing.

Can I use this without Kling 3.0? Kling is the dominant tool because its image-to-video preserves identity well. Runway and Luma both work but tend to drift the face more. If you're paying per generation, Kling 3.0 is currently the best value for this specific trend.

What about the 야구장 여신 (yagujang yeosin) version? "야구장 여신" — literally "baseball stadium goddess" — is the original Korean-community framing of this same aesthetic. The prompts above are the English distillation of what Korean creators were posting on Threads and Fanplus a week earlier. Same trend, same Kling backend, just credited to the original Korean fan community. See What is the Korean Baseball AI Fan Cam Trend? for the cultural backstory.

Will this trend still work in a month? The fan-cam aesthetic has been viral for years on actual Korean TV — the AI version just lets anyone make one. The visual format will keep working; what'll shift is which sport and league. KBO is the current peak, but the same prompt structure with NBA, EPL, or NFL substitutions should pick up the next wave.

Have a prompt variant that works better? The community is moving fast — tag the canonical examples and we'll update this page.