AI Kissing Video Looks Wrong? 7 Problems and How to Fix Them
Face changes, twin faces, stretched frames, extra people, stiff kisses and failed runs: what causes each AI kissing video problem and the quick fix.
Most AI kissing videos go wrong for one of six reasons: a blurry or small face, very different photos, the wrong aspect ratio, a crowded upload, a prompt that fights the built-in instructions, or a model that struggles with close contact. Each one has a quick fix, listed below. If a generation fails outright on VdoBloom, your credits are returned automatically.
A kiss is one of the hardest things to ask a video model for. Two faces get very close, touch, partly cover each other and then come apart, and the model has to remember who each person was the whole time. When it loses track, you see face swaps, merged features or a stranger at the end of the clip. Here is what causes each problem and how to fix it.
Problem 1: A face changes partway through
Cause: The model had too little detail to hold on to. Small faces, motion blur, sunglasses, heavy beauty filters and dark photos all make this worse.
Fix:
- Use a sharp photo where each face is large in the frame. Head-and-shoulders works better than a full-body shot.
- Remove sunglasses, hats with deep shadows and strong filters.
- Pick a model with steadier faces for the final run. Kling V2.5 Turbo (40 credits for 5 s) usually holds faces better than the cheapest options.
Problem 2: Both people look like the same person
Cause: The two photos look too alike, for example similar pose, light, hair and clothing, so the model blends them.
Fix: Use photos with clearly different outfits or backgrounds. Make sure both faces are roughly the same size. If one is much smaller, crop that photo closer before you upload.
Problem 3: The video looks stretched, cropped or squashed
Cause: The first frame and the video were different shapes, so the video model had to crop or stretch it.
Fix: Choose the aspect ratio before you generate: 9:16 for Reels and TikTok, 16:9 for YouTube. On VdoBloom the combined first frame is built at the same aspect ratio as the video, so they match.
Problem 4: An extra person appears
Cause: One of the photos had other people in it, or the prompt described a crowd, such as "at a party".
Fix: Use photos with one person each in two-photo mode, or one photo with just the couple. Choose a quiet setting such as a balcony, a beach at sunset or a cosy room rather than a busy place. The built-in prompt already says no one else should appear; don't work against it.
Problem 5: The kiss looks stiff, or never happens
Cause: The clip is too short for the action, or the prompt asks for too much, such as walking up, hugging, kissing and spinning, all in 5 seconds.
Fix:
- Keep one action per clip: lean in and kiss.
- Use 10 seconds for a slower, more romantic build-up.
- Try the Sweet or Romantic style, which use gentler motion than Passionate or Dramatic.
Problem 6: The people float on a black or empty background
Cause: A cut-out PNG with a transparent background was sent straight to the video model.
Fix: On VdoBloom this is handled for you. Before anything moves, a full scene is painted behind cut-out photos. If you use another tool, upload the original photo, not a cut-out.
Problem 7: The generation was blocked or failed
Cause: Prompts are screened before a job runs, and providers sometimes decline requests. Typical triggers are explicit wording, a photo that looks like a minor, or a well-known public figure.
Fix: Keep the prompt romantic, not explicit. Use photos of adults who agreed to take part. If the provider declines or the job fails, VdoBloom returns your credits automatically, so trying again costs nothing extra. Sometimes switching to another model in the picker also works.
Quick checklist before you generate
| Check | Good | Avoid |
|---|---|---|
| Faces | Large, sharp, front or three-quarter view | Tiny, blurry, full profile, sunglasses |
| Photos | One person each, or just the couple | Group shots |
| Aspect ratio | Chosen before generating | Changed after the frame is made |
| Prompt | One action, one setting, one camera move | Several actions in 5 seconds |
| Model | Cheap draft, then one premium run | Iterating only on the most expensive model |
A cheap way to iterate
Run drafts on Seedance 2 Mini at 480p (8 credits for 5 s) or Runway Gen-3 at 720p (10 credits) until both faces and the framing are right. Then run the same photos, style and prompt once on Kling V2.5 Turbo (40) or Runway Gen-4.5 (48) for the final version. Five drafts and one final run usually costs less than several attempts on a premium model.
Common questions
Why does my kissing video look fine at the start and wrong at the end?
Faces are easiest to keep in the first frame, which the model copies directly. Drift builds up over time, which is why the end looks worse. Use larger, sharper faces, or pick a steadier model.
Do I lose credits when a video fails?
No. If a generation fails, the credits go back to your balance automatically.
Can I fix one bad second without starting again?
Not inside the clip. The quickest fix is to regenerate. A second pass on the same inputs often fixes an odd hand or expression.
Ready to try again? Open the AI Kissing Video generator, or read our guide to combining two separate photos.
Ready to try it?
Create your first AI video in minutes — no credit card required.
Start Creating Free →