Edit Video with AI Prompts: MiniMax H3 Instruction Editing
Instruction-based AI video editing, explained: how MiniMax H3 swaps objects, changes backgrounds and rewrites dialogue through video reference — with examples.
AI video has always had one answer to a flawed shot: roll again. The take is 90% right — the light is perfect, the move is perfect, but there's a cat where the client's dog should be — and your only tool is the slot machine. Re-roll, and you lose the 90% along with the 10%.
Instruction-based editing is the category that ends that. Instead of regenerating, you hand the model your existing footage and a sentence: change this one thing, keep the rest. MiniMax shipped H3 with this as a headline capability on July 31 — and its official manual demonstrates it with disarmingly mundane examples that escalate fast. We run Pixo, a multi-model video platform where H3 is live in the Playground, and here's how prompt-based editing actually works — including how to do it today.
Re-Roll vs. Edit: Two Different Machines
The distinction is worth being precise about, because "AI video editing" gets used loosely:
| Re-rolling | Instruction editing | |
|---|---|---|
| Input | Your prompt, again | Your footage + an instruction |
| What survives | Nothing is guaranteed | Everything you didn't mention |
| Cost of a 10% fix | A full new generation, new dice | One targeted change |
| Use case | The shot is wrong | The shot is almost right |
Editing turns video generation from gambling into revision — which is the difference between a toy and a production tool, because clients don't ask for "another one," they ask for "the same one, but."
What H3's Editing Can Do: The Four-Step Escalation
The manual's own demonstration sequence is the best tour of the range:
1. Swap an object. "Replace the cat in the video with a dog." The dog inherits the cat's position, scale and motion; camera, lighting and background don't move. This is the atomic unit of instruction editing.
2. Replace the world. A green-screen shot becomes a storybook background — the environment swaps wholesale while the subject keeps their light-wrap and edges consistent with the new scene. Relighting instructions ("make it night") work the same way.
3. Rewrite the dialogue. Change what a character says, and the voice and the performance update — mouth movements follow the new line, delivery adjusts to match. This is the step no timeline editor can do, because the pixels of the mouth were never separate from the audio.
4. Change three things at once. The manual's convenience-store example stacks them: a soda can becomes a branded cola, the glowing sign changes, and the closing line is rewritten — one instruction, three edits, everything else in the shot untouched.
How to Do It Today: Video Reference + Instruction
In Pixo's Playground, editing runs through H3's reference system — there's no separate edit button, and you don't need one. The footage you want to change is the reference:
- Upload the clip you want to change. 4–15 seconds, up to 50MB. It becomes Video 1 — AI-generated or real footage, either works.
- Write the instruction against it. Address the reference by number, name the change, protect the rest.
- Generate. The output is your shot with the named change executed. Iterate at 768p, finish at 2K, export watermark-free.
Three instruction patterns, ready to adapt:
The showcase's reference files
The manual showcase this prompt adapts
Video 1 is the footage to edit. Replace the cat with a golden retriever — same position,
same movement, same scale. Keep the camera move, lighting, background and all sound
exactly as in Video 1.
The showcase's reference files
The manual showcase this prompt adapts
Video 1 is the footage to edit. Change the time of day to night: moonlight replaces
the daylight, windows glow warm, long soft shadows. The people, their movements and
the camera work stay exactly as in Video 1.
The showcase's reference files
The manual showcase this prompt adapts
Video 1 is the footage to edit. Image 1 locks the replacement product. In Video 1,
replace the drink can on the counter with the product from Image 1, and change the
final spoken line to: "Cold. Since 1987." Everything else stays unchanged.
Create AI videos with Pixo
Turn any idea into a publish-worthy video. One sentence is all it takes.
The Instruction Formula
Every reliable edit instruction has three parts — this is the pattern to memorize:
[Name the target] + [state the change] + [protect the rest]
- Name the target: "the man's jeans," "the drink can on the counter," "the final spoken line." Pronouns and vibes fail; nouns work.
- State the change: "replace with black tailored trousers" — including the end state, not just "change it."
- Protect the rest: "keep the camera move, lighting and background unchanged." Saying it explicitly is cheap insurance against helpful over-editing.
For time-scoped changes, add when: "in the first three seconds, …" The more surgical the sentence, the more surgical the edit.
Where This Fits in a Real Workflow
Instruction editing is a revision tool, and revision is most valuable where shots are almost approved: ad variants ("same spot, swap the SKU" — see e-commerce product videos), dialogue fixes after review (the line changes, the performance follows — see H3's native audio), and restyling passes (style transfer is the same reference mechanism pointed at aesthetics). Both of this year's flagships converged on editing the same day — how H3's instruction approach compares to Seedance 2.5's time-ranged and region-based tools is covered in MiniMax H3 vs Seedance 2.5.
FAQ
Can you edit an AI-generated video instead of regenerating it?
Yes — this is what instruction-based editing means. With MiniMax H3, you provide the existing video as a reference and write a plain-language instruction: swap the cat for a dog, change the background to dusk, rewrite the closing line. The named change is executed while everything you didn't mention stays stable.
How does editing work in Pixo's Playground?
Through video reference. Upload the footage you want to change as a video reference (4–15 seconds, up to 50MB) — it becomes Video 1 — then write the instruction against it: "In Video 1, replace the cat with a golden retriever; keep the camera move, lighting and background unchanged."
What can an edit instruction change?
The documented range covers swapping or removing characters and objects, changing backgrounds and lighting, and rewriting dialogue — with the performance and lip-sync adjusting to the new line. Multiple changes can ride in one instruction.
Does it work on real footage or only AI clips?
Both. The mechanism is a video reference plus an instruction, and the reference can be any clip you upload — AI-generated or shot on a phone.
How do I write a good edit instruction?
Use the three-part pattern: name the target ("the man's jeans"), state the change ("replace with black tailored trousers"), and protect the rest ("keep everything else unchanged"). Vague targets produce vague edits; named targets produce surgical ones.
How do I try it?
MiniMax H3 is live in Pixo's Playground — upload your clip as a video reference, write the instruction, and generate at 768p or 2K. New users get 200 free credits on sign-up, and plans are currently up to 55% off.
MiniMax H3 is now live in Pixo's Playground — upload footage as a video reference, write the instruction, and generate at 768p or 2K. Sign up now — new users get 200 free credits on sign-up, and plans are currently up to 55% off. New to the model? Start with what MiniMax H3 is.
Generate AI summary
From idea to finished video.
In one conversation.
Start CreatingNo credit card required • Free 200 credits
