DEV Community: FunGen

How to Make a Photo Dance with AI: What the Photo and the Dance Clip Need

FunGen — Sun, 08 Nov 2026 09:46:03 +0000

An AI dance generator takes two inputs: a still photo of a person and a short dance clip. It returns a new video in which the person from the photo performs the moves from the clip. The dancer supplies the motion and the photo supplies the look, so you don't need to film anyone dancing.

Most bad results come from those two inputs, not from the model. Here's what each one needs.

How it works under the hood

These tools are built on motion control models. The model reads the body pose frame by frame from the reference clip, then redraws the person from your photo in each of those poses. It also tries to keep the face, hair and outfit consistent across frames.

Two things follow from that:

The model can only animate what it can see. Anything missing from the photo, such as legs that are cropped out or a hand behind the back, has to be invented.
It can only follow motion it can read. A blurry, crowded or fast-cut clip gives it a noisy pose to copy.

The photo: one person, head to feet

Do:

Use one person, standing, with the whole body in frame from head to feet.
Pick a simple background. A busy one bleeds into the figure when it moves.
Use even light and a sharp image, with the face clearly visible.

Avoid:

Waist-up crops. This is the most common mistake. If the photo stops at the waist, the model has to invent legs and feet, and those are the parts that most often come out wrong.
Group photos. The model has to guess which person should dance.
Sitting or lying poses. A dance needs a standing body to start from.
Loose, flowing clothes and long hair are fine, but expect them to change more from frame to frame than fitted clothes do.

The dance clip: short, simple, one dancer

If you upload your own reference instead of using a ready-made one:

One dancer, whole body in frame for the entire clip.
No cuts. A single continuous shot. A jump cut looks to the model like the body teleporting.
Plain background and steady camera. No zooms or pans.
Short. 3–10 seconds is typical. Short clips are cheaper to test and drift less.
Moderate moves. Side steps, sways, bounces and arm swings transfer well. Spins, floor work and hands crossing in front of the face are much harder.

Before you upload, also check that the clip's framing matches your photo. A full-body photo paired with a full-body dance gives the model the easiest job.

What to expect from the result

Be realistic:

The moves follow the timing of the clip, but they won't match it exactly.
Faces, hands and outfit details can shift during fast moves. Check every result before you post it.
It's slow. On FunGen's dance page, three example renders took about 5½–7 minutes each, and each one was the first attempt.
There may be no sound. The ready-made dances there are silent, so the results are too. Add music afterwards in an editor.

What it costs

Motion control tools usually bill per second of output. As an example, the AI dance video generator in FunGen uses the Kling 2.6 motion control model. It charges 15 credits per second in Standard mode and 30 in Pro. Its three ready-made dances are about 6.6 seconds each, so one comes to about 99 credits on Standard. It shows the estimate before it generates and charges only when the video is delivered.

Eligible new accounts get 100 free credits, which covers one ready-made dance on Standard.

A quick checklist

Full-body photo, one person, standing, plain background.
Short, single-shot dance clip with one dancer and a steady camera.
Start on the standard tier, check the result, then move up if it's worth it.
Add music afterwards and label the video as AI-generated where the platform asks.

One rule that isn't technical

Only use a photo of yourself or of an adult who has clearly agreed to it. Don't make it look like a real person did something they didn't, and don't pass the result off as real footage.

How to Make a \"Santa Caught on Camera\" Photo of Your Own Living Room

FunGen — Tue, 03 Nov 2026 04:25:51 +0000

Every December, "Santa caught on camera" clips do the rounds: a grainy shot from a shelf camera, the tree lights on, and a figure in red setting a present down. You can now make one of your own living room with an AI image model. The model does the Santa part. Whether the result looks believable mostly comes down to the photo you start with.

Here's how to take that photo, and what to check before you show it to anyone.

1. Shoot the room the way a camera would see it

Real home-camera footage has a look. It's shot from high up, wide and slightly off to one side. If you take your photo from eye level in the middle of the room, it will look like a staged picture instead of a caught moment.

Stand where a camera would sit. A bookshelf, the top of a doorway or a corner near the ceiling all work. Hold the phone high and tilt it down a little.
Go wide. Use the 0.5× lens if your phone has one. You want the floor, the tree and a bit of the wall around them in the frame.
Leave space for Santa. He needs somewhere to stand, so keep a clear patch of floor next to the tree or the fireplace.

2. Light it like Christmas Eve

Night shots with the tree lights on look the most convincing. They also hide the small inconsistencies that give an AI edit away.

Turn the main lights off and the tree lights, lamps or fireplace on.
Keep the phone steady, against a doorframe if you need to. Blur reads as "fake" faster than darkness does.
Avoid a window behind the tree. Strong backlight makes it harder for the model to light Santa the same way as the room.

3. Keep people out of the shot

The photo should be the room and nothing else: no kids on the sofa, no reflection of you in the TV. It reads better, because Santa is the only person in the scene, and it's safer, because you aren't feeding a photo of your family into an AI service. Some tools block photos with children in them for exactly that reason.

4. Generate, then check the details

Upload the photo to an AI tool built for this. It should place Santa at a believable size, match the room's lighting and leave your furniture where it is. Then zoom in before you share:

Did anything in the room change? Models are told to keep the layout and the camera angle, but small details like ornaments, cushions or picture frames can shift. Kids notice their own house.
Is Santa the right size? Compare him with the doorframe or the tree.
Does the light match? If the room is lit warm by the tree, Santa shouldn't look like he's standing under office lights.

If something's off, run it again. Results change from run to run, and the second or third try is often the keeper.

5. Optional: turn the photo into a few seconds of video

A still photo works fine for a card or the family group chat. A short clip feels more like real camera footage. Some tools animate the photo you just made: Santa sets the present down, maybe glances at the camera. Expect a few seconds of clip, not a scene, and expect the exact motion to vary between attempts.

Video costs a lot more than a photo, so get the photo right first and only animate the one you're happy with.

A worked example

For a concrete idea of time and cost, the Santa in My House tool on FunGen works in two steps. The Santa photo costs about 2 credits. Turning it into a video is optional, at 11 credits per second: 44 credits for 4 seconds or 88 for 8. The video prompt asks Santa to set the present down and make a "shh" gesture at the camera. The page shows the cost before each step, and it asks you not to upload photos of children.

Label it, and keep it fun

The result is AI-generated, so say so where a platform asks you to. For the kids it's a Christmas morning surprise, and nobody needs a lecture on image models over their cereal. But if you post it publicly, a small "made with AI" saves a lot of comment-section arguments.

AI Dubbing: When Voice-Only Is Enough and When You Need Lip Sync

FunGen — Mon, 02 Nov 2026 05:17:03 +0000

AI dubbing translates the speech in a video and replaces it with a new voice track in another language. Most tools now offer two versions of the job:

Voice only swaps the audio and leaves the picture as it is.
Voice + lip sync also changes the mouth so it matches the new words.

The second sounds strictly better, but it isn't always the right choice. It costs more, takes longer and has limits of its own. Here's how to choose.

What changes between the two

Voice only is an audio job. The pipeline transcribes the original speech, translates it, generates new speech (often trying to keep the original speaker's voice) and mixes it back over the background sound. The video frames aren't touched.

Voice + lip sync adds a video job on top. The model regenerates the mouth region frame by frame to match the new audio. That's where the extra time and cost come from.

As one concrete example, FunGen's dubbing tool charges 3 credits per second for voice only (up to 5 minutes) and 11 credits per second for voice + lip sync (up to 120 seconds). In their published tests on 7–8 second clips, voice only finished in about 17–29 seconds and lip sync took about 2–7 minutes. The exact numbers differ between tools, but that ratio is typical: lip sync is a several-times-bigger job.

When voice only is enough

The speaker's mouth isn't the focus. Screen recordings, product demos with a small face-cam, B-roll with narration and slideshow-style tutorials all fall in this group. Nobody is watching the lips.
The video is long. A 20-minute course module is a lot of lip-sync seconds. Dub the voice, and save lip sync for the intro where the instructor is on camera.
You're testing a market. If you don't yet know whether Spanish or Portuguese viewers will watch, voice only lets you try several languages cheaply before you commit.
There are several speakers at once. Lip sync works best on one clear face. Panels and interviews with crosstalk often come out better with voice only.

When lip sync is worth paying for

Talking-head content. Explainers, founder videos, UGC-style ads and avatar videos are all close-ups of a face. A mismatch between lips and words is obvious within a second, and viewers read it as low quality.
Short-form social video. TikTok and Reels clips are short, so the per-second premium adds up to little, and the face usually fills the frame.
Ads. If the video is meant to convert, the uncanny look of unsynced lips is a real cost.

Things that go wrong either way

Length drift. Languages don't take the same time to say the same thing. Spanish and German often run longer than English. Lip-synced output may come out slightly longer than the original. In FunGen's own tests, 7–8 second English clips came out about 1.3 seconds longer in Spanish. If your video is cut tightly to music, leave room for this or re-edit afterwards.

Messy source audio. Background music, crosstalk and strong accents hurt the transcription step, and every later step inherits the mistake. A clean, single-speaker source is the biggest quality lever you have.

On-screen text. Dubbing doesn't translate captions, titles or text burned into the picture. Plan to redo those separately.

A practical workflow

Cut the source to the version you actually want to publish before dubbing. You pay per second.
Dub one short section in each target language with voice only, and check the translation with a native speaker if you can.
For the languages that pass, decide clip by clip: lip sync for on-camera segments, voice only for everything else.
Re-check timing against the music and on-screen text.

If you want to try both modes side by side, the AI dubbing tool in FunGen covers 32 languages and shows the full price before it starts.

One rule matters more than any setting: only dub videos you have the rights to, and only recreate voices of people who agreed to it.

Why Your AI Lip Sync Looks Fake (and the Input Fixes That Help)

FunGen — Sun, 01 Nov 2026 02:55:16 +0000

AI lip sync takes one still image and an audio track and turns them into a talking video. The model is rarely the reason a result looks off. Most of the problems come from the inputs: the photo, the script and the audio.

This post starts from the symptom. Find the one you're seeing, then fix the input behind it.

Symptom: the mouth barely moves

Likely cause: the face is too small in the frame.

In a wide shot or a group photo, the mouth is only a few dozen pixels across, and the model has very little to work with.

Fix: crop to waist-up or closer, so the face takes up a decent share of the image. Use at least 300 px on each side, and bigger if you have it.

Symptom: teeth or lips look smeared

Likely cause: a blurry or low-resolution source photo.

The model has to invent the inside of the mouth, and it does that badly when the lips themselves are already soft.

Fix: start from a sharper photo. If the tool offers a higher-resolution output tier, try it for close-ups.

Symptom: strange expressions or a stiff face

Likely cause: a full profile, or strong side light that puts half the face in shadow.

Fix: use a face that is turned toward the camera (straight on or slightly angled) and lit evenly. Also check that nothing covers the mouth: hands, microphones, masks and hair all cause trouble.

The face doesn't have to be a real person. Illustrations, 3D characters, mascots and pets work too, as long as there's a clear face with a visible mouth.

Symptom: the face looks frozen between words

Likely cause: a very short or monotone line.

Fix: give the face something to do.

Write the way people talk, with contractions and short sentences.
Use commas and full stops for pauses. A face that rests for a beat between sentences reads as more human than one that talks nonstop.

Symptom: the voice and lips disagree on numbers or acronyms

Likely cause: the text-to-speech voice reads "2026" or "GPU" one way, and the lips follow the audio.

Fix: spell things out the way they should be said ("twenty twenty-six", "G P U"), so what you hear and what you see match.

Symptom: lips drift out of sync with uploaded audio

Likely cause: music, background noise or a second voice in the recording. The model tries to animate all of it.

Fix: upload a clean, voice-only take with one speaker. If you want music, add it in the edit afterwards.

Symptom: it looks fine, except at the start or the end

Artifacts tend to show up in the first and last second of a clip. Watch those parts closely. Trimming half a second often saves a take.

A quick review routine

Watch every result once with the sound on, and check three things:

Hard consonants. The lips should close on "p", "b" and "m".
The first and last second. This is where artifacts usually appear.
The eyes. Natural blinking matters as much as the mouth.

Test short, then go long

Test a new photo with a 5–15 second line before you write a full script. Tools usually bill by the second.

For example, the AI lip sync tool in FunGen charges 11 credits per second at 720p and 21 at 1080p, and accepts up to 60 seconds of audio (MP3, WAV or M4A). It shows the cost before it generates.

Eligible new accounts start with 100 free credits. That isn't quite enough for a 10-second 720p clip, which is 110 credits, but it covers a shorter test.

One rule that isn't technical

Only animate photos you have the rights to. Don't make real people say things they didn't say without their permission. That rule heads off most legal and platform problems before they start.

From Flat Art to a Printable Mesh: What Image-to-3D Actually Has to Guess

FunGen — Wed, 21 Oct 2026 14:15:53 +0000

Hand an image-to-3D model a picture of a mascot and you get back a mesh with a front, a
side and a back. Only one of those was in the picture.

That is the whole problem in one sentence. A photo or an illustration is a projection:
depth has been thrown away, and everything facing away from the camera was never
recorded. Every image-to-3D system is, at its core, a very well-trained guesser about
the missing half. Once you see where it has to guess, you can predict which inputs will
come out clean and which will come out melted.

Guess 1: depth

From a single view, a large thing far away and a small thing up close look identical.
Models resolve this with priors: they have seen a lot of cars, a lot of cats and a lot
of bananas, and they assume your object is shaped like the typical one.

That works beautifully for common objects and badly for unusual ones. A generic sports
car converts well because the model has a strong idea of what a sports car's side
profile looks like. A one-off character with an oddly proportioned head is where the
prior pulls the shape toward "average" and you lose the thing that made it yours.

Practical rule: the further your subject is from a common category, the more the
input should show its shape — a three-quarter view beats a flat front view, because it
gives the model two faces of the object to triangulate from instead of one.

Guess 2: the back

The back is pure invention. For symmetric subjects this is mostly fine, because the
model mirrors what it sees. For asymmetric ones — a character holding something in one
hand, a logo on one side of a shirt — the back is a coin flip.

Flat illustration styles make this harder, not easier. Thick outlines and cel shading
read as "drawn", and the model has to decide whether that black line is an edge in the
geometry or just ink. Soft shading with a visible light direction is a much stronger
depth cue than a bold outline.

Guess 3: what is one object

A sticker with a drop shadow, a character standing on a patch of grass, a mascot with a
speech bubble: the model has to decide what belongs to the object. Backgrounds that
touch the subject tend to get fused into the mesh as a slab or a lump.

The cheapest fix happens before generation: a clean cut-out on a plain background,
subject centred, nothing touching its silhouette.

The render is not evidence about the mesh

This is the part people underestimate. A preview render can look flawless while the
geometry underneath is unusable for anything physical:

What you see in the render	What can be wrong underneath
Smooth, detailed surface	Very high triangle count, painful to slice or rig
Crisp painted details	Detail lives only in the texture, not in the geometry
One solid figure	Several disconnected shells, or holes that are not watertight
Sensible proportions	No real-world scale; it imports at an arbitrary size

For a game prop or a listing render, texture-only detail is fine. For a resin print it
is not: paint does not survive slicing, and anything thinner than your printer's minimum
wall either breaks or never prints. If a print is the goal, judge the mesh, not the
picture — open the file, check it is one closed shell, scale it, and look at the thin
parts.

What converts well

Putting the guesses together, the inputs that consistently come out clean are:

Common silhouettes: vehicles, fruit, toys, chunky mascots.
Three-quarter views with soft shading and a visible light direction.
Isolated subjects on a plain background.
Thick features: stubby limbs, solid bases, no wires, no hair strands.

And the ones that disappoint: thin, spiky, translucent or reflective subjects, busy
scenes, and flat-shaded line art viewed dead-on.

The short version

Image-to-3D is not reconstruction, it is informed guessing about the half of the object
nobody photographed. Feed it inputs that make the guess easy — a common shape, a
three-quarter angle, a clean cut-out — and judge the output by the mesh file rather than
the preview render.

If you want to try that loop on your own art, FunGen has an
ai 3d image creator online free to start
with: upload a picture or generate one, and it hands back an OBJ with its texture map
rather than a viewer link, so you can check the geometry yourself.

One Prompt, Four Modalities: What a Unified Generation Agent Actually Has to Solve

FunGen — Sun, 18 Oct 2026 16:13:05 +0000

Most generation tools are one modality wide. You have a favourite for images, a
different one for video, something else for voice, and if you touch 3D at all it is
a fourth tab with its own account and its own export quirks.

The obvious fix is to put them behind one text box. That sounds like a UI decision.
It is not. Once a single prompt field has to serve image, video, voice and 3D, three
genuinely hard problems show up, and none of them are about layout.

Problem 1: the modalities disagree about what a prompt is

A good image prompt is a dense noun phrase. Adjectives, materials, lighting, lens.
Nothing about time, because there is no time.

a worn brass desk lamp, single hard key light from the left, dusty highlights

A good video prompt is a verb phrase with a camera in it. The same noun phrase
produces a static shot that technically moves.

a worn brass desk lamp; camera pushes in slowly as the filament flickers on

A good 3D prompt is neither. It is a set of geometric constraints, and half the
adjectives that improve an image actively hurt it — "dusty highlights" is a
texture instruction with no geometry behind it, and "hard key light" is
meaningless in a mesh.

a desk lamp, hollow base, flat bottom, thick shade walls, no thin wires

A voice prompt is a performance direction: pace, register, emphasis.

So a single prompt box is lying a little. The same sentence cannot be optimal in
four places at once. The design question is what to do about that, and there are
three honest answers:

Make the user rewrite it per modality. Correct, and nobody does it.
Rewrite it silently. Fast, and it hides why a result was bad.
Rewrite it and show your work. Slower, and the only option where the user learns anything.

The third is what "agent" should mean in practice: the thing plans, tells you what
it is about to do, and lets you disagree before spending the compute.

Problem 2: each modality fails in an unrelated way

The failure modes have almost nothing in common, which means one generic error
state is useless:

Modality	Typical failure	What the user needs to see
Image	Composition drifts from the brief	A/B against the prompt
Video	Frame-to-frame identity drift	Which frame it broke on
Voice	Right words, wrong performance	The specific phrase to re-read
3D	Geometry is plausible, topology is not	Mesh stats, not a render

That last row is the one people underestimate. A 3D render can look flawless while
the underlying mesh is unusable — hundreds of thousands of triangles, disconnected
shells, no consistent scale. The render is not evidence about the geometry. Any
tool that shows you only the pretty picture is answering a different question from
the one you asked.

Problem 3: chaining is where the value is, and where the state lives

The single-modality workflow is: prompt, generate, download, done.

The interesting workflow is: generate an image, keep that exact subject, put it in a
scene, animate it, add a voice over it, and separately turn the subject into a mesh.

Each arrow in that chain is a place where identity has to survive. That is a state
problem. The system has to carry a persistent handle on "this specific thing" across
model boundaries that were never designed to agree with each other. Tool-hopping
loses that handle at every step, which is why the manual version of this workflow
produces four assets that do not quite look like the same object.

Keeping the chain inside one project is not a convenience feature. It is the only
place the continuity can live.

What this implies for exports

If the chain is the point, the exports have to be ordinary files — MP4, PNG, WAV,
OBJ — and they have to be downloadable without ceremony.

A generation tool that will not hand over the file is not a tool, it is a demo. This
matters most for 3D, where the file is the entire deliverable: nobody wants a mesh
they can only look at inside someone else's viewer.

The short version

Putting four modalities behind one prompt is not a consolidation exercise. It
requires the system to translate intent per modality, surface modality-specific
failure honestly, and keep subject identity alive across model boundaries. Get those
three right and the single text box is genuinely simpler. Get them wrong and you
have built four tools that happen to share a font.

If you want to see what the chained version feels like, I have been building it as an
ai video agent that plans the shot list, shows its work before
spending credits, and hands you ordinary files at the end. New accounts get free
credits, which is about the right amount of commitment for finding out whether the
unified version is worth it.