5.2Generative video (Runway, Pika, Sora, Veo)
The essentials in 30 seconds
- The central challenge: temporal consistency — objects, characters and lighting stable from one frame to the next, far harder than a beautiful still image.
- Image-to-video first: animate an already validated image (framing, style, composition) describing only the movement — far more controllable than text-to-video.
- Pro lucidity: short sequences, limited control, possible drift — aim for mood shots and loops integrated into a classic edit.
Running case — Sonia animates her client's banner
Sonia's client in Algiers wants “an AI video” for social media. Sonia, a freelance graphic designer, refuses to promise a complete film: she selects a visual already validated by the guidelines (section 5.1), animates it in image-to-video mode with a simple slow travelling shot, generates several takes and keeps only one — integrated into a classic edit. Sequel to case 5.1 — Sonia moves from still visual to animated shot without betraying the guidelines.
Sequel to case 5.1 — Sonia animates a visual validated by the guidelines, in short shots integrated into the edit.
1. The challenge: temporal consistency
Generative video is progressing very fast, on the same principle as images — but with one major extra difficulty: temporal consistency. Each frame being beautiful is not enough; objects, characters and lighting must stay stable from one frame to the next. Main players: Runway, Pika, Sora, Veo — a fast-moving ecosystem, to check regularly.
Sonia warns her client: what impresses in a demo (a spectacular short shot) does not guarantee a coherent thirty-second film. Being lucid about this limit is what separates a professional from a promise seller.
2. Image-to-video: the controllable approach
Two modes dominate. Text-to-video generates from a description — flexible but hard to control, with the model deciding everything at once. Image-to-video starts from an existing image that is already validated and animates it: framing, style and composition are fixed, you only ask for the movement — travelling, panning, slow zoom, simple action. That is the professional approach.
Tips: describe only the movement, prefer it simple and slow (wide and fast degrades consistency), pick a favourable starting image (water, smoke, fabric, clouds animate better than a frozen scene), and generate several takes — selection makes quality.
3. Limits and realistic uses
Current limits to know before promising: short sequences, limited precise control, consistency drift (faces deforming, objects appearing), and significant cost per second multiplied by the takes. Sonia prices her takes to the client instead of silently giving them away.
Realistic uses: short mood shots, animations of existing visuals, loops for social media, elements integrated into a classic edit — rather than a whole film in one block. The professional practice is hybrid: generate elements, keep control of the story in the edit. And as the field moves very fast, retest regularly.
🔎 Concrete example — Sonia animates a validated visual
A single variable: the movement.
Validated generated image (framing, style, composition controlled).
Image-to-video mode: describe ONLY the movement (slight travelling shot).
→ several takes; integrate the shot into a classic edit.Reading: the style stays on-guidelines since the starting image is validated.
Lesson: simple, slow movements — consistency degrades otherwise.
Section vocabulary
- Temporal consistency
- The stability of objects, characters and lighting from one frame to the next of the sequence.
Simple example: Sonia's character face must not deform between frames.
What it is not: the beauty of a single isolated frame.
Reflex: does the shot hold stable for its whole duration?
- Text-to-video
- Generating a sequence from a description: the model decides everything at once.
Simple example: Sonia tests a shot idea with no starting image, for exploration.
What it is not: a controlled, consistent production method.
Reflex: am I exploring or producing?
- Image-to-video
- Animating an already validated existing image, asking only for the movement.
Simple example: Sonia animates the banner validated by the guidelines with a slow travelling shot.
What it is not: redescribing everything, scene and movement, on every take.
Reflex: is my starting image already validated?
- Camera movement
- The move requested in the video prompt: travelling, panning, slow zoom.
Simple example: Sonia asks for a slight forward travelling shot on the banner.
What it is not: a complex, fast choreography.
Reflex: is my movement simple and slow?
- Hybrid editing
- Assembling generated shots, still images and real footage in editing software.
Simple example: Sonia integrates her animated shot into the client's video edit.
What it is not: a whole film generated in one block with no editing.
Reflex: who keeps control of the story?
In practice — Animate an existing visual
- Select a generated image you have validated (framing, style and composition already controlled).
- Use a service in image-to-video mode and describe ONLY the wanted movement (slight forward travelling, wind in the leaves).
- Generate several takes: as with images, the result is probabilistic and selection makes quality.
- Integrate the resulting shot into a classic edit rather than trying to generate everything in one block.
Ready-to-say script — framing a client video request
“For your video, I suggest the most reliable approach: we start from a visual already validated by your guidelines, I animate it as a short shot with a simple, slow movement, and I integrate it into a classic edit. I will not promise you a complete film generated in one block — consistency from one shot to the next does not allow it yet — but quality mood shots and loops in your brand style.”
To say to the client to calibrate the promise before any generative-video quote.
Copy-paste template — AI video shot sheet
AI VIDEO SHOT SHEET
STARTING IMAGE: ___ (on-guidelines? YES / NO)
MODE: image-to-video · MOVEMENT: ___ (simple and slow)
TAKES: ___ generations · KEPT: no. ___
DURATION: ___ s · INTEGRATION: edit ___
CHECK: stable consistency? YES / NO · AI disclosure planned? YES / NOSolo / small budget
Sonia alone: a single visual animated in a loop for the client's socials — one shot is enough to test the effect.
Small business / small team
Three takes generated, only one kept as a team: collective selection avoids publishing a shaky shot.
SME in Algiers
In Algiers, Sonia sets a viewing session with the client before integrating the shots into the final edit.
Starting team
Pair: one describes the movement, the other checks consistency frame by frame before validation.
✓ Best practices
- Start from a validated image (image-to-video) to keep style control.
- Describe only the movement; prefer it simple and slow.
- Integrate shots into a classic edit and retest tools regularly.
✗ Common pitfalls
- Promising a long, coherent video generated in one block.
- Generating everything in text-to-video: hard to control and costly in takes.
- Ignoring that the cost per second rises fast with the takes.
Key takeaways
- The challenge specific to video is TEMPORAL CONSISTENCY (stability from one frame to the next).
- Image-to-video (animating a validated image) is far more controllable than text-to-video in production.
- Current limits: short sequences, limited control, consistency drift, significant cost per second.
- Realistic uses: mood shots, visual animations, social-media loops integrated into a classic edit.
48h challenge — apply tomorrow
Take 1 validated visual of Sonia, animate it in image-to-video with a simple, slow movement: does the shot hold stable for its whole duration?
- I can explain temporal consistency in one sentence.
- I know why image-to-video beats text-to-video in production.
- My short shot is integrated into an edit, not promised as a film.
Frequently asked questions
Can you really produce a professional video entirely with AI?
Not yet end-to-end for most uses: what works is short shots, moods, visual animations and loops integrated into an edit. Duration, continuity, precise control and synchronisation stay hard — hence the hybrid practice, retested regularly as the field progresses very fast.
Is it better to start from text or from an image to generate video?
Start from an image whenever possible: in image-to-video the model handles only one variable, the movement, and consistency with your guidelines is guaranteed. Describe only the movement, simple and slow, on a favourable starting image — and generate several takes.
Which approach is the most controllable in generative video?