A practical AI video tutorial: build a shot list, write prompts that survive the render, keep characters consistent, and judge output before you commit.
Most people who look for an AI video tutorial want the same thing: a short clip that looks like it cost real money to produce. What they usually get instead is a folder of semi-random five-second renders, half of which melt at the edges and none of which cut together. The gap is rarely the model. It is the workflow around the model.
This tutorial walks through that workflow end to end: how to decide what you are actually making, how to break it into shots, how to write prompts that hold up, when to switch from text-to-video to image-to-video, how to keep a character recognizable across cuts, and how to judge a render in ten seconds instead of twenty minutes. Big platforms from Google's Veo line to open research demos keep raising the ceiling on raw quality, but the ceiling is not what stops most projects. Structure is.
Start with the deliverable, not the tool
Before you open any generator, write one sentence describing the finished asset. Not "a cool cinematic video" — that is a mood, not a deliverable. Something like: "A 30-second vertical teaser for a ceramic mug brand, three shots, warm morning light, no dialogue, ends on a logo card."
That sentence decides almost everything downstream. Vertical means composition changes. Thirty seconds means you need roughly five to seven generated clips at typical clip lengths, plus a title card. No dialogue means you can ignore lip-sync entirely — a major simplification. Three shots means you are not building a sequence, you are building a montage, which is far more forgiving of continuity errors.
Beginners tend to skip this step and generate thirty clips hoping one is magic. Professionals define the container first, then fill it. The container is the reason a $10 generation run can beat a $200 one.
A useful exercise: write the deliverable sentence, then list the three things a viewer must understand in the first four seconds. If you cannot name three, the piece is not ready to generate.
The four layers of any AI video workflow
Every AI video project, regardless of platform, has four layers. Problems almost always live in one specific layer, and naming it saves hours of blind re-rolling.
Layer 1: Intent
What the video is for, who sees it, where it plays, how long it runs. This layer is entirely human and costs nothing to fix. Most failed AI videos are failed intent documents.
Layer 2: Reference
Anything that pins down appearance: a still image, a style frame, a color script, a character sheet, a product photo. Text descriptions are lossy; images are dense. If a shot must look like a specific object or person, feed an image rather than describing it in adjectives.
Layer 3: Generation
The model call itself. This is the layer everyone obsesses over and the one that matters least once the first two are solid. Model choice does matter — motion coherence, camera behavior, and texture handling differ noticeably between engines — but it is a multiplier, not a substitute.
Layer 4: Assembly
Editing, pacing, sound, color, text. This is where a set of mediocre clips becomes a coherent piece, and where a set of beautiful clips becomes a boring one if the pacing is wrong.
When something looks bad, ask which layer failed. Was the intent vague? Was the reference missing? Did the model misread the prompt? Or did you simply cut too slowly? Diagnosing the layer is faster than switching tools.
Text-to-video: writing prompts that survive the render
Text-to-video is the most intuitive entry point and the easiest to misuse. A prompt is not a wish; it is a compressed brief. Long adjective lists confuse models more than they help, because every additional clause competes for the same attention budget.
A reliable prompt structure has five parts, in this order:
- Subject and action — one clear subject doing one clear thing. "A cyclist rounding a corner," not "a cyclist, athletic, determined, motion blur, urban."
- Setting and time of day — where and when. Light is the single highest-leverage detail in AI video.
- Camera — shot size and movement. "Low-angle tracking shot, slow push in." Camera language tells the model how to move, which is otherwise the most common source of drift.
- Texture and grade — film stock, grain, palette. Keep it to one clause.
- Constraints — what must not happen. "No text overlays, no crowd, no camera shake."
Here is a weak prompt and a strong one.
Weak: "Beautiful cinematic video of a woman in a city at night, neon, emotional, 4k, masterpiece."
Strong: "Medium shot of a woman in a rain-soaked jacket standing under a store awning at night, neon reflections on wet pavement, static camera, 35mm film grain, cool blue and magenta palette, no crowd, no camera shake."
The strong version is not longer because longer is better. It is longer because every clause removes a decision the model would otherwise make for you — and models make worse decisions than you do.
Two practical habits. First, change one variable at a time when iterating. If you alter the camera and the lighting and the palette in the same pass, you learn nothing about which one broke the shot. Second, keep a prompt log. A simple text file with prompt, model, seed if available, and a one-line verdict is enough. After twenty renders you will see patterns in your own phrasing that no tutorial can teach you. Our prompt library is a reasonable starting point for structure, but your own log will always be more useful.
Image-to-video, video-to-video, and when each earns its keep
Text-to-video is fast and cheap for exploration but weak at fidelity. Once you know what a shot should look like, stop describing it and start showing it.
Image-to-video takes a still and animates it. This is the highest-leverage mode for product work, character work, and anything where a specific object must remain recognizable. Generate or photograph a hero frame, then animate subtle motion: a slow push, a light shift, a small gesture. The constraint of a fixed first frame is a feature, not a limitation — it eliminates most composition drift.
Video-to-video takes existing footage and restyles or extends it. It is the right choice when you have real footage that is boring, badly lit, or needs a different look, and the wrong choice when you have no footage at all. It is also the mode most people ignore, which is a shame: restyling a ten-second phone clip is often faster than generating a matching shot from scratch.
Keyframe interpolation — giving a model a start frame and an end frame — is the third option and the most underused. If you need a specific transition, such as a closed box becoming open, defining both ends constrains the model far more than any adjective.
A rough decision rule: use text-to-video to explore, image-to-video to lock, video-to-video to rescue, and keyframes to connect. Mixing modes inside one project is normal and usually produces better results than forcing a single mode to do everything.
Keeping characters and style consistent across shots
Consistency is the hardest part of AI video and the part most tutorials skip. There are four levers, in order of effectiveness.
A character reference sheet. Generate or collect four to six images of the same character from different angles in consistent lighting. Reuse them as image inputs for every shot. Do not rely on a text description of a face; language cannot carry that much specificity.
A locked style block. Write one paragraph describing palette, lens, grain, and lighting, and paste it verbatim into every prompt. Change subject and action between shots; never change the style block. If you paraphrase it, the grade will drift and your edit will look assembled from different films.
Consistent shot sizes. Alternating between extreme wide and extreme close-up hides drift; staying in medium and medium-wide exposes it. Choose a range and stay inside it. Counterintuitively, restraint makes AI sequences look more expensive.
A single seed or reference image per scene. Where the platform exposes it, reuse seeds to stabilize texture, and swap only the motion instructions.
Even with all four, expect drift. Plan for it in the edit: cut on motion, keep shots short, and use insert shots — hands, objects, environment details — as connective tissue. Inserts are cheap, forgiving, and make sequences feel intentional.
If you want to compare how different engines handle the same character reference, our alternative comparisons are built for exactly that question.
A practical shot list you can copy
Here is a shot list for the mug teaser from earlier. It assumes clips of about five seconds.
| # | Shot | Mode | Prompt core | Purpose |
|---|---|---|---|---|
| 1 | Close-up, steam rising from mug | Image-to-video | Locked hero frame, gentle steam motion, warm window light | Hook |
| 2 | Medium, hands wrapping around mug | Text-to-video | Hands, ceramic texture, morning light, static camera | Human warmth |
| 3 | Insert, glaze detail rotating | Keyframes | Start and end angle defined | Product texture |
| 4 | Wide, mug on wooden table, room out of focus | Image-to-video | Slow push in, shallow depth | Context |
| 5 | Logo card | Editor | Static, 12 frames | Close |
Five clips, one of which is not generated at all. The total generation work is four renders plus retries — a twenty-minute job if the references are ready, an afternoon if they are not.
Notice that the shot list includes purpose. Every shot earns its place by doing a job. If you cannot name the job, delete the shot. This single rule prevents the most common failure mode in AI video: technically impressive clips that add nothing.
Judging output: a five-point check before you keep a clip
Watching a render twice and hoping is not a method. Run the same five checks every time, and you will reject bad clips in seconds.
- First frame legibility. Pause at frame one. Is the subject clear and correctly framed? If the first frame is ambiguous, the clip will fight your edit.
- Motion path. Watch once at normal speed without pausing. Does the motion go somewhere, or does it oscillate? Oscillating motion reads as artificial regardless of texture quality.
- Edge integrity. Check hands, hair, thin objects, and the frame borders. These fail first and are the strongest signal of whether a clip can survive a large screen.
- Lighting consistency. Does the light source stay put? A shadow that flips direction mid-clip is unfixable in the edit.
- Cut compatibility. Does the clip's final frame plausibly lead into the next shot? If not, either reorder or insert a transition.
Score each check pass or fail. Two fails means re-render; one fail means consider whether an insert shot can hide it. Keep a folder of near-misses — they are often reusable for a different cut.
Common mistakes that burn render time
Prompt bloat. Twelve adjectives do not equal twelve times the control. Cut every word that is not a decision.
No reference images. Describing a face or a product in text and expecting fidelity is the biggest time sink in the hobby. Generate a still first, then animate it with the AI image generator and AI video generator in sequence.
Chasing a single perfect clip. Ten variations of one shot rarely beat one good shot plus three supporting inserts. Sequences beat masterpieces.
Ignoring aspect ratio until the end. Generate in the ratio you will deliver. Reframing a 16:9 render into 9:16 crops away the composition you carefully prompted.
No sound plan. Silent AI video feels unfinished. Even a single ambient bed and one impact sound changes perceived quality more than a resolution bump.
Skipping the edit. The timeline is where AI video becomes video. Cut ruthlessly; most first assemblies are 40% too long.
Never reusing assets. Build a small library of approved style frames, character sheets, and background plates. Your second project should start faster than your first.
FAQ
Do I need a powerful computer to make AI video?
No. Generation happens on remote hardware. A mid-range laptop with a stable browser and decent upload speed is enough. Local rendering only becomes relevant if you run open models yourself, which is a different workflow with different tradeoffs.
How long should each generated clip be?
Shorter than you think. Four to six seconds covers most cutaways, and short clips hide drift. Generate long only when the motion itself is the point, such as a continuous camera move.
Is text-to-video or image-to-video better for beginners?
Start with text-to-video to explore ideas cheaply, then switch to image-to-video as soon as you know what you want. The switch usually happens within the first hour and improves results immediately.
Why does my character change between shots?
Because text descriptions cannot carry facial specificity. Use a reference sheet of four to six images and reuse it for every shot, plus a locked style block pasted verbatim into each prompt.
Can I use AI video for client work?
Often yes, but check the terms of the specific platform and be transparent with clients about the process. Many brands are comfortable with AI-assisted video as long as the result is good and the usage rights are clear.
What is the fastest way to improve?
Keep a prompt log with verdicts, and finish projects instead of collecting clips. Finishing teaches pacing, which no prompt can fix.
How many renders should I budget per finished second?
Plan for roughly two to four renders per second of finished footage when you are learning, dropping toward one as your references and prompts stabilize. Budget time, not just generations.
Where to go from here
Pick the deliverable sentence, build a shot list with a purpose per shot, prepare references before you generate anything, and judge every clip against the same five checks. That routine will outperform any model upgrade for the next year of your work.
When you are ready to put it into practice, start with a single hero frame in the AI image generator, animate it in the AI video generator, and see how much of the workflow above you can complete in one sitting. If you want a head start on structure, browse the video templates or read more breakdowns on the Orelon blog — then go make the thirty seconds you actually described.

