Orelon logoOrelon
Tarifs

Text to Video Prompts and MP4 Playback: A Working Guide

7 nov. 2026 · Par Orelon Team

Explorez les modèles vidéo IA

Parcourez quelques créations de la communauté pour trouver l’inspiration, puis ouvrez n’importe quel modèle pour continuer à créer dans Orelon.

Write text-to-video prompts that hold up, choose render settings that survive delivery, and fix any MP4 that refuses to play.

A generated clip is not a finished asset. Most people judging an AI video tool look at the render — type a sentence, wait, watch the result, decide whether the model is any good. People who actually ship video judge the handoff instead: can the file be opened on the machine that needs it, imported into the editor that will cut it, and delivered to a client without a support conversation?

That gap explains a large share of frustration. The render was fine; the pipeline around it was not. A downloaded clip that shows a black rectangle with audio playing underneath, a file that stutters on a phone but runs smoothly on a laptop, footage an editor refuses to import while a browser plays it happily — none of these are creative failures. They are format and workflow failures, and they are almost always fixable in minutes once you know what you are looking at.

This guide treats both halves of the loop as one skill. First, how a written sentence becomes a sequence of frames, and how to phrase prompts so the result survives the render. Second, the practical business of opening, checking, converting, naming, and archiving the MP4 files that come out the other end. A clip nobody can play is not footage. It is a file.

How a Sentence Becomes Frames

The model is solving for coherence, not beauty

Video generation models work in a compressed representation where images and motion are stored as numbers. Starting from noise, they iteratively refine that representation until it matches the guidance from your text. The text encoder steers every step, which means your words are not a description of a finished picture — they are a set of constraints applied repeatedly across time.

That is why a prompt that would produce a striking still image can produce a mushy video. Temporal coherence has to keep a face, a jacket, and a window frame recognizably themselves across dozens or hundreds of frames. Every element you add is another thing that must remain stable. One subject with a clear action is easy. Four characters, a moving camera, and rain are not impossible, but they compete for the same limited attention.

A useful mental model: the model is not painting a picture of your sentence, it is keeping a promise across time. The more promises you make, the more likely one of them breaks somewhere in the middle of the clip — and the break is usually in the frame you planned to use.

Motion has a budget

Three constraints shape nearly every result you will see from any text-to-video system:

  • Duration is short. Clips typically land between four and ten seconds. That is roughly one idea, one action, one camera move.
  • Complexity compounds. Fast panning plus a spinning object plus a walking character usually exceeds what the model can resolve, and something smears or warps.
  • Resolution and frame rate come from settings, not from prose. A prompt does not request 4K. Your export panel does.

Once you accept those limits, prompting becomes an exercise in narrowing rather than describing. Narrower inputs produce more usable outputs, and usable outputs are what make a sequence possible.

The Five-Slot Prompt Skeleton

Five slots, in order of importance

Instead of writing flowing prose and hoping, build every prompt from five slots. You can rearrange them, but include all five:

  1. Subject — who or what, with two or three concrete visual anchors such as age range, wardrobe, material, or species.
  2. Action — one primary verb, plus at most one small secondary motion.
  3. Setting — location, time of day, weather, and one background detail that establishes scale.
  4. Camera — shot size, angle, and movement as a single clear instruction.
  5. Light and style — light quality, color palette, and a rendering reference such as documentary, 35mm film, matte painting, or product photography.

The structure maps onto how attention gets allocated: identity, motion, environment, framing, tone. When a clip disappoints, you can usually trace the problem to a missing slot rather than a weak model. A prompt with no camera instruction gets a camera decision made for you. A prompt with no light instruction gets flat, generic illumination.

Ambiguity versus overload

"A woman walking through a city at night" leaves the model to choose the city, the decade, the lens, the color of the streetlights, and the walking speed. It will choose. It just will not choose what you imagined. Add two specific anchors — a wool coat, neon signage in a language you name, wet asphalt — and the frame narrows immediately.

The second cause of drift is overloading. Four simultaneous actions and three camera instructions get averaged into a character who half-walks, half-turns, in a shot that half-pans. Cutting an action is almost always more effective than adding adjectives.

Rewrites you can copy

Weak prompt Why it fails Rewritten
"A cool futuristic city" No subject, action, or camera "Slow aerial push over a coastal city at blue hour, glass towers with warm interior lights, low fog between buildings, wide shot, cinematic teal and amber grade"
"A dog running fast and jumping and turning" Three competing actions "A border collie sprints across wet sand toward the camera, slow motion, low-angle tracking shot, overcast light, shallow depth of field"
"Product video for my coffee brand" That is a brief, not a prompt "Macro shot of steam rising from a ceramic cup on a dark walnut table, 45-degree angle, slow dolly in, single warm key light from the left, soft shadow falloff"
"Sad person" Emotion with no visible evidence "Medium close-up of a man in his fifties sitting in a parked car, hands on the wheel, rain on the windshield, cool daylight, no eye contact with camera"

Notice the strong versions are not longer for the sake of length. Each added phrase resolves a decision the model would otherwise make on its own.

Reference frames and negative guidance

If your tool accepts negative prompts, use them surgically: text overlays, watermarks, distorted hands, extra limbs, flicker. Long negative lists backfire because they reintroduce concepts the model then notices.

Reference images are the strongest consistency lever available. Feeding a character sheet or a product photo into an image-conditioned generation keeps identity stable across shots far more reliably than describing the same face five different ways. A workable pattern: generate a still first in an AI image generator, lock the look you like, then animate that still instead of starting from text each time.

Plan the Sequence Before You Open the Tool

Write the one-line brief

Before touching a tool, write one paragraph covering audience, tone, length, platform, and the single feeling the clip should produce. This takes four minutes and saves an hour of wandering. "Thirty-second product teaser for a hiking backpack, aimed at weekend hikers, calm and confident, vertical for social, one hero shot plus three detail shots" is enough to drive every decision that follows.

Shot lists built for short clips

A shot list for AI-generated footage has one row per clip with five columns: shot number, duration, what the audience must understand, prompt draft, and status. Keep individual generations between four and eight seconds. If a beat needs twelve seconds, treat it as two shots with a cut, because cuts hide continuity problems that a single long take exposes.

Two planning rules save hours:

  • Write the voiceover or on-screen text first. Clips exist to support a line, not the other way around. Knowing the line tells you whether you need a wide establishing shot or a close-up.
  • Plan coverage, not variety. Generate the same subject at three shot sizes rather than three unrelated scenes. Wide, medium, and close versions of one moment cut together like real coverage; three different locations do not.

If you are starting from a blank page, browsing video templates can shortcut structural decisions so your judgment goes where it matters.

Settings That Decide Whether the File Survives

Before generating, confirm the parameters that determine downstream usability. It is far cheaper to set them correctly once than to discover the problem in an edit.

Setting What to choose Why it matters
Aspect ratio 16:9 for web and long-form, 9:16 for shorts, 4:5 for feeds Reframing after the render crops resolution and can cut the subject
Resolution Match the delivery target, not the maximum available Higher resolution takes longer and often gets downscaled anyway
Frame rate 24 fps for a cinematic feel, 30 fps for general web, 60 fps only if you plan slow motion Mixed frame rates cause judder after editing
Duration The shortest clip that covers the beat Longer generations drift more and are harder to re-run
Variants Three to five per prompt Selecting is faster than iterating on one output

Build one habit: review variants at full size, not in a thumbnail grid. Compression artifacts and identity drift are invisible at small preview sizes and obvious on a television or a client monitor.

Why an MP4 Won't Open on Your Computer

Container versus codec

MP4 is a container, not a codec. This single fact explains most playback confusion. A file ending in .mp4 can hold H.264, HEVC/H.265, or AV1 video, paired with AAC, MP3, or Opus audio. Your computer may handle one combination and fail on another.

Think of the container as an envelope and the codec as the letter inside. The envelope decides how tracks are labeled and interleaved; the codec decides how pixels are compressed. Editors and upload platforms care about both, in different ways.

  • H.264 in MP4 is the safest delivery format. Everything plays it, from old laptops to social platforms.
  • HEVC/H.265 roughly halves file size at similar quality but has patchy desktop support and licensing baggage.
  • AV1 is efficient and increasingly common on the web, but hardware decoding is still uneven.
  • ProRes is an editing format, not a playback format. It is enormous and not meant for sharing.

The working rule: generate however you like, deliver H.264 MP4 with AAC audio unless the client specifies otherwise.

A triage checklist for files that will not open

Work through this in order. Each step eliminates a whole class of causes, so you are not guessing.

  1. Confirm the download finished. A partial file often shows a thumbnail from the first frames but stalls a few seconds in.
  2. Try a player with broad codec support. Players that ship their own decoders separate "corrupt file" from "missing decoder" in one step.
  3. Check whether the file contains video at all. Properties dialogs and media inspectors show the actual streams. Occasionally an export contains audio only.
  4. Note the codec name. HEVC and AV1 are the usual suspects on older hardware and on systems without the appropriate decoder installed.
  5. Drop the file into a browser tab. If it plays there, the file is fine and your desktop media stack is the problem.
  6. Test on a second device. Phone plays it, laptop does not: you have a local playback issue, not a broken render.

Normalizing, Converting, and Delivering

When a file plays badly, transcode rather than re-render. Re-generating wastes time and gives you different footage; converting preserves the performance you already approved.

Encoder settings that matter

  • Video codec: H.264. The universal choice for review copies and delivery.
  • Constant quality around 20-23. Lower numbers mean higher quality and larger files; 20 is a good delivery default.
  • Pixel format 4:2:0 with 8-bit color. Files encoded in other chroma formats can appear black in older players and some browsers even though they play perfectly on the machine that made them.
  • Audio: AAC at 192 kbps, 48 kHz. Standard, predictable, plays everywhere.
  • Fast start enabled. Moves the file index to the front so playback begins before the full download completes. Essential for anything streamed or embedded.
  • Frame rate matched to the source. Do not let a converter invent frames.

Also normalize loudness before delivery. Dialogue around -16 LUFS for web platforms and -14 LUFS for social is a reasonable default, and it prevents the classic complaint that your clip is quieter than everything around it.

Naming, versioning, and backups

scene02_shot03_v2_seed4481.mp4 beats output_final_final.mp4 every time. A useful convention is scene, shot, version, and any identifier that lets you reproduce the look. Keep the seed or reference frame in the filename or a sidecar note. Six months later, when a client asks for "that one with the fog," the filename is the only search engine you will have.

Back up the source renders. Delivery files get recompressed, reformatted, and sometimes deleted by platforms; the originals are the only version that can be re-graded and re-cut at full quality. Storage is cheap relative to re-generating footage you can no longer reproduce exactly.

Three Worked Examples

A product close-up for a landing page

The brief: a six-second hero loop for a skincare landing page, calm and premium, 16:9, no text on screen. The shot list has three rows — bottle in light, texture macro, hands applying.

Prompt for row one: "Amber glass serum bottle standing on a pale travertine surface, single soft window light from the upper left, shallow depth of field, slow 15-degree orbit, wide-to-medium framing, muted warm palette, product photography." Render at 24 fps, 1080p, three variants. Pick the one where the label stays legible for the full duration.

Then transcode the selection to H.264, lay a quiet ambient bed under it, and export a second 9:16 version generated natively rather than cropped in the editor — cropping a 16:9 render to vertical loses the sides of the frame and frequently the product itself.

A narrative establishing shot

The brief: an eight-second opener for a short documentary about a coastal town, overcast, observational, no characters in frame. Prompt: "Wide static shot of a granite harbor wall at dawn, small fishing boats moored in flat grey water, low cloud, distant lighthouse, fine mist, documentary photography, desaturated palette." Static camera, because movement in an establishing shot fights the cut that follows it.

Generate two variants and check the water for repeating texture patterns. Reflection artifacts are the most common giveaway in wide landscape generations, and they are much cheaper to fix by re-rolling with a different framing than by painting them out later.

A vertical social hook

The brief: a five-second hook for a fitness app, 9:16, energetic, motion-forward. Prompt: "Low-angle tracking shot of a runner's feet striking wet pavement, splashes catching backlight, camera moving with the stride, dusk, high contrast, slight motion blur." Render at 30 fps and 9:16 from the start; do not shoot wide and crop.

Because the subject is feet and pavement, identity drift is a non-issue. That is a deliberate choice: shots without faces tolerate far more generation variance, which makes them the best place to start when you are learning a new model.

Mistakes, Review Habits, and Library Hygiene

The expensive mistakes in AI video work are rarely technical. They are ordering errors — deciding something at the wrong moment, then paying to undo it.

  • Prompting a scene instead of a shot. Models generate clips. Sequence logic belongs in your edit.
  • Deciding aspect ratio last. The crop happens after the fact and takes resolution with it.
  • Chasing photorealism when style serves better. A stylized look hides the small inconsistencies that realism magnifies.
  • Re-rolling endlessly instead of changing one variable. Adjust a single slot per attempt, or you learn nothing from the result.
  • Copying prompts from lists without adapting them. Borrowed prompts carry someone else's decisions about lens, light, and pacing.
  • Delivering the raw render. Normalize codec, loudness, and filename before anything leaves your machine.
  • Skipping the playback test. Open the final file on a phone and a desktop before sending it. Two minutes now prevents an awkward email later.

Watch every selection three times with a purpose. Once for story: does the shot communicate the beat? Once for artifacts: hands, eyes, text, edges of frame, repeating textures. Once with sound off and playback speed reduced, which is when warping and flicker become visible.

Reject fast. A clip that is 80 percent right will not improve in the edit, and a sequence built from near-misses looks worse than a sequence built from simpler, cleaner shots. The instinct to rescue a flawed generation is the most expensive habit in this work.

Keep a reject bin rather than deleting. Fragments sometimes work as inserts, transitions, or texture layers, and a library of good fragments compounds the way a stock footage library does.

Within a month you will have hundreds of files, and the difference between a usable archive and a mess is decided early. A simple folder structure — project, scene number, versions — plus one plain-text log per project is enough. Note the prompt, the seed or reference image, the settings, and the reason a variant was chosen. When a client asks for a variation six weeks later, that log turns a rebuild into a tweak. If you produce at volume, keep proxy copies in H.264 at 720p for browsing; scrubbing a folder of high-bitrate masters is slow, scrubbing proxies is instant, and you only open the master once you have decided.

FAQ

Can I edit an AI-generated MP4 in a normal video editor? Yes. Once the file is H.264 in an MP4 container, every mainstream editor imports it without special steps. Editing directly from HEVC or AV1 files works in recent versions but can be sluggish on older machines, so converting first is often faster overall.

Why does my video play with sound but no picture? The video stream uses a codec your player cannot decode while the audio codec is a common one. The file is intact. Convert it to H.264 and the picture returns.

How long should a single generated clip be? Four to eight seconds covers most needs. Longer generations increase the chance of drift and reduce your ability to re-cut. If a moment needs more time, generate coverage at multiple shot sizes and edit them together.

Does a higher resolution setting improve quality? Not automatically. Upscaling a weak generation preserves the weakness at a larger size. Fix composition, lighting, and action in the prompt first, then render at the resolution your delivery target actually needs.

Why do my clips look different from each other? Identity and color drift are the default behavior. Lock a seed where the tool allows it, reuse a reference image for the subject, and keep the light and style slot identical across every prompt in a sequence.

What is the safest format to hand to a client? An H.264 MP4 at the platform's recommended resolution, AAC audio at a standard bitrate, and a clear filename. Supply a high-bitrate master only if they request one for further editing.

Do I need a separate converter? Not always. Many editors and players can export a clean H.264 copy. A dedicated transcoder is worth installing for batch work, resizing, and fast-start output, but a single clip can usually be normalized inside whatever tool you already have open.

How do I stop a sequence from feeling like disconnected clips? Repeat three things across every prompt in the sequence: the same light direction, the same color palette, and the same lens language. Continuity in AI video comes from repeated constraints far more than from matching story beats.

Start Creating with Orelon

The distance between a strong idea and a finished clip is mostly discipline: a one-line brief, a shot list, a five-slot prompt, three variants, a full-size review, and a clean transcode at the end. Get those habits in place and the software becomes the least interesting part of the process, which is exactly what you want.

Orelon is built for the cinematic half of that work, turning written ideas into moving footage you can cut into a real sequence. Open the AI video generator, draft your first shot using the five slots above, and watch the result at full size before you decide anything. When the phrasing stalls, the prompt library exists to unstick the next shot rather than replace your judgment — and the Orelon blog has more workflow detail for the parts that happen after the render.