H3 (Hailuo 3.0) ranks #1 on Artificial Analysis' Video Editing leaderboard. Relight a shot, swap a background, add VFX, or re-voice the dialogue — while the performance you already approved stays exactly where it is. Below: real before/after pairs, 32 copy-ready prompts, and the formula that makes edits land.
Omni-Modal Video
Model
MiniMax H3 (Hailuo 3.0) — 2K with native audio
#1 on Artificial Analysis' Video Editing leaderboard · up to 9 images, 3 clips, and 3 audio references per generation
Input
Name the change, then name what must stay. The "Keep" clause is what stops H3 redrawing the whole frame.
Edit Instruction
Length
Aspect
Quality
Text to Video
Rainy neon follow shot, generated cold with native audio
Edit a Clip
Light particles added to a finished product turn
Edit a Clip
Studio backdrop replaced with a golden-hour rooftop
Every clip on this page is our own H3 render at 768P — nothing is stock. Production runs go out at 2K.
Every pair below is our own H3 output. We generated the source clip, then fed it back to H3 as Video 1 with a single edit instruction. Nothing else about the shot was touched — same subject, same motion, same framing. Pick a different edit to see the same source change in a different direction.
The prompt
In Video 1, add a soft ring of golden light particles that lifts off the bottle and drifts upward as the hand turns it, plus a faint lens bloom on the glass edge.
Keep the bottle, the cup, the hand movement, the lighting, and the camera handling exactly the same. Sound: quiet café ambience with a soft shimmer as the particles rise.
The prompt
In Video 1, replace the plain grey studio backdrop with an out-of-focus city rooftop at golden hour — warm sky, soft skyline bokeh — and match the light on her with a warm key from the right.
Keep the woman, her sweatshirt, her expression, her nod, and the framing exactly the same. Sound: soft rooftop wind and distant traffic.
The prompt
In Video 1, change the sneaker from white to deep forest green with a cream midsole, keeping the same materials, stitching, and wear.
Keep the pedestal, the shadow, the lighting, and the camera orbit exactly the same. Sound: low studio hum.
H3 is one transformer that reads text, images, video, and audio in the same pass and returns video with sound. That architecture is why it can do the thing most video models cannot: take footage as input. Feed it a clip, name a change, and it returns the same shot with only that change applied — which is a different job from generating a video, and the job most creative teams actually have.
Leaderboard positions per Artificial Analysis as of July 31, 2026 — Video Editing Elo 1130 across 5,043 blind-preference samples.
MiniMax documents one shape for edit instructions. It mostly holds — but two of these four parts behave differently from how the docs describe them, which we only found by running the edits and measuring the output.
In Video 1,
Reference every uploaded asset by modality and order — Video 1, Image 1, Audio 1. H3 resolves them positionally, so ambiguity here is what produces edits in the wrong place.
change the lighting to warm evening…
Say what is now true of the scene, not what to do to the footage. "It is evening now" works; "apply a 35mm film grade" returned a byte-for-byte no-op in our tests. H3 edits the world in the shot, not the image of it.
Keep everything else the same.
Some preservation clause is needed or H3 re-diffuses the frame. But long ones backfire: a six-item Keep list suppressed our wardrobe edit entirely, and the same edit landed once we shortened it to "Keep everything else the same." Protect the two or three things that matter, then stop.
Sound: quiet café ambience.
Audio is generated jointly with the picture, not dubbed on after. Say "unchanged" to hold the original bed, or describe the new one — H3 will mix it against the edit.
Reviews of a video model are usually somebody's taste. This is not that. Every test is one source clip plus one edit instruction, and because H3 returns identical dimensions and frame counts, the source and the result can be compared frame by frame. Three findings below contradict what the documentation says — including one that contradicts the prompt formula further up this page.
H3 regenerates the whole clip even when it changes nothing, so every edit carries some baseline drift. We fed each source back with a “change nothing at all” instruction to measure that floor. Without it the numbers are unreadable — and it is what exposed two edits that silently did nothing at all. High similarity is not a good score here; it can mean the edit never happened.
| Edit | Family | SSIM | Control floor | Result |
|---|---|---|---|---|
| Add VFX particles | Variant | 0.975 | 0.978 | Pass |
| Insert a product into frame | Variant | 0.929 | 0.951 | Pass |
| Swap product colourway | Variant | 0.910 | — | Pass |
| Remove a prop | Variant | 0.896 | — | Pass |
| Add a lit device screen | Variant | 0.936 | 0.974 | Pass |
| Change wardrobe | Variant | 0.951 | 0.951 | No-op → passed when shortened |
| Re-voice into Spanish | Localization | 0.966 | 0.974 | Pass |
| Replace the background | Localization | 0.869 | 0.951 | Pass |
| Relight day to evening | Polish | 0.854 | 0.978 | Pass |
| Burn in captions | Polish | 0.963 | 0.974 | Pass |
| Regrade to film stock | Polish | 0.973 | 0.973 | No-op → partial when reworded |
| Add rain and wet ground | Polish | 0.962 | 0.973 | Partial — wet ground, no rain |
| Remove signage text | Compliance | 0.959 | 0.963 | Pass |
| Rewrite signage text | Compliance | 0.961 | 0.963 | Pass |
| Change camera viewpoint | Hard | 0.955 | 0.973 | Partial — subject turned, camera did not |
768P, 5-second clips, one run per scenario. Video models are stochastic, so single runs cannot separate a capability limit from an unlucky seed — treat the two no-ops as strong signals, not proofs. The product-pedestal tests predate the control runs and have no floor.
Asking for a colour-science operation — "regrade to warm 35mm film: lifted blacks, halation, fine grain" — produced a clip that differed from the source by *less* than our do-nothing control did. Re-phrasing the identical intent as a fact about the world — "the scene is now lit by hazy warm late-afternoon sun" — put sunlight down the street. Post-production vocabulary is the thing that fails; describe what changed in the world instead.
MiniMax's own guidance is to name what must stay. It is real advice — drop the clause and H3 re-diffuses the frame. But we hit the opposite failure: a six-item preservation list paired with an over-specified change ("a structured cream linen blazer over a white tee") returned a byte-for-byte no-op. The same edit, written as one short sentence, landed cleanly and still held her face, hair, and backdrop.
Every write-up of H3 repeats that small on-screen text is a weak spot, because MiniMax says so. At signage scale it held up fine for us: the lettering was rewritten in the same paint style, layout, and position, and separately removed to a blank board with the wood grain and shadow intact. Fine print and exact logo reproduction may still fail — we did not test those — but "H3 cannot do text" is too strong.
29 of these operate on footage you already have — relighting, VFX, background swaps, object edits, re-voicing, on-screen typography, reframing. Each one names its change and its preservation list, so you can paste it, swap the nouns for your own shot, and run it. The full library lives in the H3 prompt library.
Dress up a plain product cutdown without a compositor.
In Video 1, add a soft ring of golden light particles that lifts off the bottle and drifts upward as the hand turns it, plus a faint lens bloom on the glass edge. Keep the bottle, the cup, the hand movement, the lighting, and the camera handling exactly the same. Sound: quiet café ambience with a soft shimmer as the particles rise.
Inputs: 1 source clip (2–15s)
Sports and apparel hooks — punch up the first 2 seconds.
In Video 1, at the moment the subject plants their foot, add a single-frame white flash and a radial dust burst that expands outward from the contact point and settles over the next second. Keep the subject, their motion, the wardrobe, the background, and the camera move exactly the same. Sound: a low impact thud under the existing audio.
Inputs: 1 source clip with a clear action beat
App install ads — imply the product is running without a screen recording.
In Video 1, turn on the phone screen so it emits a cool blue glow that spills onto the hands and the underside of the face, with a soft falloff on the surrounding surfaces. Keep the phone, the hands, the pose, the room, and the camera framing exactly the same. Sound: unchanged.
Inputs: 1 source clip featuring a device
Seasonal variants of one shoot — sell rain gear from dry footage.
In Video 1, add steady rain falling through the frame, wet reflective ground with small splashes, and damp highlights on the subject's shoulders. Keep the subject, their movement, their wardrobe, the buildings, and the camera path exactly the same. Sound: rain on pavement layered under the existing audio.
Inputs: 1 source clip shot outdoors
Energy-drink and fitness creative — make ordinary motion feel kinetic.
In Video 1, add soft luminous motion trails that follow the subject's hands and trail off after roughly half a second, tinted to match the key light. Keep the subject, the speed of their motion, the wardrobe, the background, and the camera exactly the same. Sound: unchanged.
Inputs: 1 source clip with visible hand or body motion
Two dayparts from one shoot — test morning vs evening creative.
In Video 1, change the lighting from bright overcast daylight to warm evening: a deep blue window outside, a low amber lamp raking across the table, and soft warm falloff on the bottle. Keep the bottle, the cup, the hand movement, the table, and the camera handling exactly the same. Sound: quiet café ambience with faint evening street noise.
Inputs: 1 source clip (2–15s)
Rescue a grey-day shoot without a reshoot.
In Video 1, relight the scene as low golden-hour sun from camera-left: long warm highlights across the subject, a soft orange rim on the shoulder and hair, and long shadows stretching to the right. Keep the subject, their movement, the wardrobe, the set dressing, and the camera framing exactly the same. Sound: unchanged.
Inputs: 1 source clip lit flat or overcast
Nightlife and beverage angles from a daytime studio shoot.
In Video 1, relight the scene with saturated nightlife colour: magenta key from the left, cyan rim from behind, and a slow pulse of the magenta in time with the beat. Keep the subject, their movement, the wardrobe, the room layout, and the camera exactly the same. Sound: a muffled club track behind the existing audio.
Inputs: 1 source clip lit neutrally
One talent shoot, one environment per market.
In Video 1, replace the plain grey studio backdrop with an out-of-focus city rooftop at golden hour — warm sky, soft skyline bokeh — and match the light on her with a warm key from the right. Keep the woman, her sweatshirt, her expression, her nod, and the framing exactly the same. Sound: soft rooftop wind and distant traffic.
Inputs: 1 source clip on a plain backdrop
One model covers generation, reference-driven shots, and editing. The editing half is where it beats everything else on the board.
Pass a finished clip and describe the change. Subject swap, relight, background replacement, add or remove an element — the rest of the shot holds.
Up to 9 images for subject and style, 3 clips for motion, and 3 audio references in one generation — cited in the prompt as Image 1, Video 1, Audio 1.
Sound is generated jointly with the picture rather than dubbed on. Ambience, effects, and dialogue land in sync with what is on screen.
Supply a voice reference and H3 delivers new lines in it, with mouth movement matched — the basis for per-market versions of one talking-head shoot.
Hand it a clip whose camera work and editing rhythm you want, and it applies that treatment to your subject instead of copying the footage.
It reads the shot as a space, not a surface — added elements pick up the scene lighting and cast their own shadows, and viewpoint changes hold the geometry.
Large typography can float in depth behind a subject or wrap onto a wall in perspective. Headline-scale works; fine print does not.
MiniMax is releasing H3 under its Community License, which makes it the strongest open-weights video model available and a viable base for fine-tuning.
Worth knowing before you build a workflow on it. These are MiniMax's own documented constraints plus what shows up consistently in testing.
Headline-scale typography renders well. Fine print, UI labels, and exact trademark reproduction do not — composite those in afterwards rather than asking H3 for them.
A busy crowd that must stay pixel-identical behind your edit is the documented failure case. Keep the preservation list to things that actually matter to the shot.
Single changes are reliable; fine detail work degrades as you pile instructions on. Run precise edits one pass at a time on the shortest source clip that contains them.
The two modes cannot be combined. If you need keyframe control over a transition, that is a separate generation from the edit pass.
Clips cap at 15 seconds and reference video caps at 15 seconds combined. H3 is a shot-level tool, not a timeline — long-form still gets assembled in an editor.
Cold text-to-video sits #2 and image-to-video #3. If nothing about your shot exists yet and quality is the only axis that matters, the models above it are still ahead.
The economics of variant testing change when a variant costs an edit instead of a shoot.
Relight a single cut into morning, golden hour, and night. Three creatives, one production day, identical performance in all of them.
Swap the backdrop per region and re-voice the dialogue per language with matching mouth movement. Localisation without a reshoot.
Change the colourway, the finish, or the hero product itself while the pedestal, the lighting, and the camera orbit stay locked.
A cut is already converting. Add VFX, a kinetic headline, or a caption package to it instead of rerolling and losing what worked.
Extend a landscape master into vertical rather than cropping it, so Reels, Shorts, and feed all get a composed frame.
Remove the distracting prop, insert the product that was not on set, turn on the device screen. Salvage footage you already paid for.
H3 gives you the edit. Soku AI turns it into a testing loop — batch the variants, ship them as ads, and learn which change was worth making.
Point Soku at a cut that works and generate the edited variants — dayparts, markets, SKUs, hooks — as a batch instead of one prompt at a time.
Brand guidelines ride along, so every variant stays on-brand without rewriting the preservation list each run.
Reframe the same master to 9:16, 1:1, and 16:9 and push it live to Meta, Google, and TikTok without leaving the workflow.
The edit holds the subject size, so nothing important lands outside the safe area.
Track CTR, CPA, and ROAS by variant, so you find out whether the night relight or the VFX pass actually moved anything.
Winners feed the next round of edits instead of the next round of guesses.
Run MiniMax H3 in Soku AI, turn one cut into a full set of ad variants, and deploy them across Meta, Google, and TikTok.
We use essential cookies to operate and secure Soku. With your permission, we also use optional analytics and advertising cookies to measure usage and campaigns. You can change your choice at any time. Privacy Policy
Soku AI