AI Before and After Video Generator

Before-and-after is one of the strongest video formats because it puts the proof on screen. It is also the format most likely to mislead, because a generated “after” can look better than anything actually achievable. This guide covers two methods that stay honest, the prompts that work, and the transitions that read clearly.

Two methods, and when to use each

Method 1 — Two real photos, one transition. You have genuine before and after images. The model animates each one and the transition carries the change.

• Use for: renovations, restoration, cleaning, staging, weight loss, hair, skincare with real photography.
• Why it is the right default: both states are real, so nothing is misrepresented.

Method 2 — One real photo, one generated end state. You have a real starting image and want to preview a possible outcome.

• Use for: concept previews, design proposals, pitch visuals.
• The rule: label it. A generated end state is a visualisation, not a result. Never publish it as a completed outcome.

The workflow

1. Shoot the pair properly. Same camera position, same lens, same time of day, same framing. If the two photos do not align, no prompt will save the transition.
2. Match exposure and white balance before generating. A transition between two differently lit frames reads as a glitch.
3. Animate each state separately. Gentle motion in the before, gentle motion in the after. Do not ask one generation to perform the change.
4. Do the change in the edit. Wipe, dissolve, or match cut. The transition is editing work, not generation work.
5. Label the timeline. On-screen text for elapsed time, and for method 2, a clear “visualisation” label.

Model picks

Model	Role	Example cost	Use it for
Framepack	Longer image animation	14 credits (2s · 720p)	Holding a single before or after photo with subtle motion
FLUX 3 Video	Image-led video continuity	170 credits (5s · 720p)	Keeping one reference image coherent across a longer take
Seedance 2.5	4–30s, extendable	144 credits (4s · 480p)	Longer reveals and vertical reframes
Nano Banana Pro Edit	High-quality multi-image editing	image → image	Aligning and matching the before/after pair before animating
MiniMax H3	High-impact cinematic short video	24 credits (3s · 480p)	Cheapest model here to prove a shot before you scale it

Credit figures read from the VidVid pricing page on 2026-09-21 as published example configurations. Exact cost is shown on the Generate button and changes as models update. See current pricing

Prompts that work for before-and-after

PROMPT — Before state, subtle life
The room from the reference photo, unchanged, very slow push-in, flat even lighting, dust motes drifting slightly, nothing appears or is removed, 4 seconds, 16:9, locked-off camera.

PROMPT — After state, same framing
The same room and camera position, renovated surfaces as shown in the reference image, gentle warm light, very slow push-in, geometry identical to the before shot, 4 seconds, 16:9, locked-off camera.

PROMPT — Transition plate
Slow wipe of light travelling left to right across the frame, neutral grey backdrop, 1 second, 16:9.

PROMPT — Vertical social cut
Same pair reframed vertical, centred on the key change, faster pacing, 3 seconds each, 9:16.

Test cards

Card A — Alignment test. Input: your real before and after photos. Model: MiniMax H3. Settings: 5s, 768P. Checkpoint: do the two clips cut together without a visible jump? Output: the full H3 run for this card — file, prompt and settings — is published below.

Card B — Continuity test. Input: after photo. Model: MiniMax H3. Settings: 5s, 768P, 9:16. Checkpoint: does anything appear or vanish mid-shot? Output: the full H3 run for this card — file, prompt and settings — is published below.

Card C — Vertical cut. Input: after photo. Model: MiniMax H3. Settings: 4s, 768P, 1:1. Checkpoint: is the change readable at thumbnail size? Output: the full H3 run for this card — file, prompt and settings — is published below.

The three MiniMax H3 runs behind these cards

Same three cards, run once each on MiniMax H3 through the official API at 768P. Files, prompts and settings are published below so you can repeat them exactly. These are bench runs on a reference image and generic prompts, not on your catalogue — read them as evidence of turnaround and frame stability, not as a verdict on your own asset.

Open this clip in a new window →

Prompt we sent (as run): A slow macro push-in on a matte ceramic object, realistic reflections, black studio background, premium commercial lighting, the object keeps its exact shape, no deformation, 5 seconds.

Open this clip in a new window →

Prompt we sent (as run): Vertical 9:16 product hook, energetic but controlled camera move, subject fully visible in the first second, bright accent colour, shallow depth of field, object shape stays fixed.

Open this clip in a new window →

Prompt we sent (as run): Extreme macro of a brushed metal surface, slow lateral camera move, reflected light crawling across the material, no distortion, commercial product photography look.

Run	Mode	Settings	Output file	Submit → ready
A	image → video	5s · 768P	1344×768, 24 fps, 5.2 s, 726 KB	196 s
B	text → video	5s · 768P · 9:16	768×1344, 24 fps, 5.2 s, 1.0 MB	117 s
C	text → video	4s · 768P · 1:1	768×768, 24 fps, 4.5 s, 341 KB	83 s

Scope: three single runs on one model. We did not measure the credit deduction inside VidVid, we did not time any other model, and we did not score output quality — that stays a human call. VidVid’s published H3 example rate is 24 credits (3s · 480p); the live figure is on the Generate button.

Now run the same three cards on another model

A measured card only means something when the model is the single variable. Keep the input, prompt, duration and resolution identical and the next generation shows you what that model does differently. Models from the table above, with their published example costs:

• Framepack — 14 credits (2s · 720p) — Holding a single before or after photo with subtle motion
• FLUX 3 Video — 170 credits (5s · 720p) — Keeping one reference image coherent across a longer take
• Seedance 2.5 — 144 credits (4s · 480p) — Longer reveals and vertical reframes

Open the studio, load the same reference image, paste the same prompt and switch the model. One generation is enough to see the difference on your own asset. Run the comparison in the VidVid studio

Credit figures are VidVid’s published example configurations, read on 2026-09-21, and are only comparable at matched settings. None of these models was timed by us.

Swapping the model changes the bill and the result — not the workflow. The same shot on the other models in this guide: Seedance 2.5 — 144 credits (4s · 480p), FLUX 3 Video — 170 credits (5s · 720p), Framepack — 14 credits (2s · 720p). What actually differs: Framepack animates a single still cheaply, which suits a slow reveal; FLUX 3 Video keeps continuity between the before frame and the after frame; Seedance 2.5 is the flexible option when the transition needs to run longer or sit in a different aspect ratio.

The failure that ruins the format

Symptom: the transition looks impressive and viewers still do not believe it. Or worse — the “after” contains features that were never built.

Cause: the before and after were not shot from the same position, so the eye reads the jump as a trick; or a generated end state invented details.

Fix:

1. Reshoot the pair from a locked-off tripod position. This is a photography fix, not a prompt fix.
2. Match colour temperature and exposure before generating.
3. Keep generation subtle — the model should add motion, not content. Add constraints: nothing appears or is removed, geometry identical.
4. For method 2, label the output as a visualisation and never present it as a completed result.

Where the format works

• Renovation and property — strongest use case when both photos are real. See the real estate guide.
• Restoration and cleaning — high satisfaction, low ethical risk.
• Product and packaging redesign — use method 2 for concept pitches only.
• Personal transformation — only with real, consented photography and honest timelines.

FAQ

Can AI generate a before-and-after transition on its own?

It can animate both states, but the transition itself belongs in the edit. Generate two clean clips, then cut.

Should I use image-to-video or text-to-video?

Image-to-video, always. You have real before and after photos; describing them in words throws away the evidence.

How do I stop the ‘after’ from inventing details?

Short clips, explicit “nothing appears or is removed” constraints, image-to-video, and a human comparison against the real after photo before publishing.

Is it OK to publish a generated ‘after’?

Only if it is clearly labelled as a visualisation. Publishing a generated end state as a completed result is a misrepresentation.

What length works best?

Three to four seconds per state, then the transition. Long enough to register, short enough to loop.

Can I see real MiniMax H3 output before I spend anything?

Yes. Three complete H3 runs are published on this page with their prompts, settings and the MP4 files themselves — 768P, 4–5 seconds, 83–196 seconds from submit to file ready. The other models are listed with their published example cost, so you can run the same prompt and compare on your own footage.

Where to go next

• Open the VidVid studio and run Card A on a real photo pair.
• Property transformations: the real estate guide.
• Vertical versions: the vertical video guide.
• Better inputs: the AI video prompt guide.