Editing
AI Camera Control
Most AI video tools randomize the camera. Martini gives you director-level shot planning per cut — orbit, dolly, push-in, pull-out, pan, crane — across Sora 2, Kling 3, Runway Gen-4, and Google Veo on one canvas. Plan the shot list, run each cut with the camera move you actually wrote, and ship a sequence that reads like real cinematography.
What this feature solves
Most AI video tools were trained on internet motion — handheld, random, sometimes pretty, sometimes nausea-inducing. The result is generations where the camera does whatever it wants regardless of what you asked for. Tell a generic image-to-video model to dolly in slowly and you might get a slow push, a fast zoom, a dolly-out, or a lateral pan — sometimes within the same shot. For social-first content where one motion is enough, that randomness is tolerable. For narrative work, branded campaigns, or anything cut against a real edit, randomness is the deal-breaker.
The newer cinematic-class models — Sora 2, Kling 3, Runway Gen-4, Google Veo — added explicit camera control as a core feature. You can write "slow orbit clockwise around the subject" and the model will actually orbit. But each model has its own prompt vocabulary, its own strongest moves, and its own quirks. Picking the wrong model for the shot type produces the same randomness problem; picking the right model requires running comparisons across all of them.
And then there is shot consistency across an edit. A two-minute branded film with twelve cuts needs every camera move to feel intentional and connected — not twelve random motions. Without a workflow that lets you plan the shot list and assign the right model per cut, edits land as a montage of disconnected moments rather than a real piece of cinematography.
Why Martini is different
Martini lets you plan the shot list as nodes on a canvas, then assign the strongest model per cut. Wide establishing shot with a slow push? Sora 2 in a video node. Subject-centered orbit on the hero shot? Kling 3. Editorial lateral pan? Runway Gen-4. Photoreal crane shot? Google Veo. Each cut picks its own model, the canvas treats the whole shot list as one pipeline, and the sequence builder orders the cuts into the edit. Shot planning becomes visible — every node is a cut, every prompt is a camera direction, every model is a choice.
Fanout is the comparison tool. When you do not know which model owns a particular move — say, a complex push-in around a moving subject — duplicate the video node, swap the model, and run them in parallel. The same prompt against the same source image across Sora 2, Kling 3, and Runway Gen-4 tells you in one pass which engine handles your shot best. No tab switching, no test-pay-compare loop across separate APIs.
Sequence integration finishes the cinematography workflow. Order the cuts in a sequence builder with their planned moves; preview the edit; export through NLE export to Premiere, DaVinci, or Final Cut. The shot list lives inside the same canvas as the generations and the timeline, so the edit reads as cinematography rather than a montage of single-shot generations.
Common use cases
Hero push-in for branded film opens
Plan a slow push-in toward the hero subject for the open of a branded film, with the model that holds focus and pace cleanly across the take.
Subject orbit for product reveals
Direct a smooth clockwise orbit around the product to reveal it from every angle — the cinematography move ecommerce video has been missing.
Crane and tilt for landscape and architecture shots
Plan a crane-down or tilt-up move on a landscape or architecture establishing shot to set scale before cutting into the scene.
Tracking shot for character-following sequences
Use a tracking shot to follow a character through a scene, with the camera holding a consistent distance and angle as they move.
Multi-shot sequences with intentional camera language
Plan a five-shot sequence where each cut has a deliberate move — wide push, medium orbit, close-up rack focus — for cinematography that reads as designed.
A/B testing camera direction for narrative beats
Run the same beat with three different camera approaches (push, orbit, lateral) and pick the one that delivers the emotional read you want.
Recommended model stack
sora-2
videoStrongest long-take camera coherence and the broadest motion vocabulary across the registry.
kling-3
videoCinematic camera language — pushes, orbits, and rack focus with director-level control.
runway-gen4
videoEditorial-grade camera direction with strong subject-tracking on lateral and crane moves.
google-veo
videoPhotoreal motion handling for live-action camera moves and natural-light tracking.
seedance-2
videoReference-locked camera moves that hold the subject identity across the take.
hailuo
videoFast iteration when testing camera direction variants on the same source image.
How the workflow works in Martini
- 1
1. Plan the shot list as nodes
Drop a node per planned cut on the canvas. Label each one with the intended camera move — wide push, medium orbit, close rack — so the shot list is visible before any generation runs.
- 2
2. Add source images per cut
Wire a reference image into each video node — the subject, scene, or product the camera will move around. Locked references prevent identity drift across cuts.
- 3
3. Pick the model per shot type
Sora 2 for long establishing pushes, Kling 3 for hero orbits, Runway Gen-4 for editorial lateral and crane, Veo for photoreal live-action. Match the model to the move, not the brand.
- 4
4. Write camera-only prompts
Tell the model what the camera should do — direction, speed, target. Avoid restating what the image already shows. "Slow push-in toward the subject's eyes, lens stays at 50mm" reads better than "show the person looking sad while the camera moves."
- 5
5. Fan out for tricky shots
When you are unsure which model owns a particular move, duplicate the node and swap models — Sora 2, Kling 3, Runway Gen-4 against the same reference. Pick the winner per cut.
- 6
6. Sequence and export
Order the cuts in a sequence builder with their intended pacing. Preview the camera flow across the edit. NLE export drops the timeline into your editor at clean frame rates.
Example workflow
A boutique creative agency is producing a 90-second branded film for a luxury watch and needs cinematic camera language across eight cuts. The director plans the shot list on the canvas: 1) wide establishing crane up over the city, 2) medium push-in to the storefront, 3) close-up orbit around the watch, 4) macro rack-focus to the dial, 5) lateral track of the model walking, 6) over-the-shoulder push to the wrist reveal, 7) wide pull-out to skyline, 8) close-up still on the logo. They assign Veo to the cityscape crane, Kling 3 to the orbit and rack, Runway Gen-4 to the lateral track and OTS push, Sora 2 to the establishing wide and pull-out. Each video node runs its planned move with the locked reference. The sequence builder orders the eight cuts. NLE export drops the timeline into Premiere ready to grade. The film reads as designed cinematography because the camera moves were planned, not generated by accident.
Tips and common mistakes
Tips
- Plan the shot list before generating. Decide every camera move on paper or on the canvas first — generations run faster when the brief is concrete.
- Match the model to the move. Each cinematic-class model has a strongest motion type — do not pick favorites.
- Write camera-only prompts. The model already sees the reference image — tell it what the lens should do, not what is in the frame.
- Use locked reference images for orbits, pushes, and pulls around a subject. Without a reference, the subject will drift mid-move.
- Run hero shots through two or three models in parallel. The winning camera move is rarely the winning subject fidelity.
Common mistakes
- Asking for vague motion ("dynamic camera," "exciting move") and expecting a specific result. Cinematic models reward specific direction.
- Picking one model for every shot in the edit. Different moves belong to different engines — that is the whole point of fanout.
- Writing prompts that re-describe the subject. The reference is the subject; the prompt is the cinematography.
- Trying to pack three moves into one ten-second clip. Plan one move per cut and let the edit do the work.
- Skipping the sequence preview. The cut reveals camera-flow problems that single shots hide — preview before exporting.
Related how-to guides
Related models and tools
Tool
AI Camera Control
Camera movement and angle control for AI video on Martini.
Tool
AI Video Frame Extraction
Extract frames from video for reference and image-to-video workflows.
Provider
OpenAI
OpenAI's GPT Image and Sora video model workflows available on Martini.
Provider
Kling
Kling 3, O3, and Avatar video model workflows on Martini.
Provider
Runway
Runway's Gen4, Aleph, and image model workflows on Martini.
Provider
Google's Veo video, Imagen image, and Nano Banana model workflows on Martini.
Related features
Multi-Shot AI Video — Build Connected Scenes, Not Isolated Clips
Plan, generate, and sequence multi-shot AI video on Martini — keep characters, style, and motion consistent across shots.
AI Image to Video — Animate Stills Into Production-Ready Shots
Turn still images into production-ready video shots on Martini's canvas — multi-model, reference-aware, NLE-export ready.
AI Video Workflow — Node-Based Production From Concept to Final Sequence
Build node-based AI video production pipelines on Martini's canvas — from concept and storyboard to final NLE-ready sequence.
AI Storyboard Generator — Plan Shots, Generate Frames, Then Animate
Plan shots, generate storyboard frames, and convert frames into video on Martini's canvas.
AI Lip Sync — Sync Voice and Dialogue to Portraits and Video
Sync voiceovers, dialogue, and music to portraits and video on Martini using lip-sync models.
AI Video Editing — Transform and Extend Existing Clips
Restyle, replace, extend, and transform existing clips on Martini's canvas — Runway Aleph, Kling O3, Wan, Seedance 2 chained into a real edit.
AI Video Upscaler — Polish AI Video to 4K on Martini
Improve AI video resolution and polish outputs on Martini's canvas.
AI Image Upscaler — Upscale Keyframes and Stills on Martini
Upscale keyframes, products, and still assets before video generation on Martini.
AI Background Remover — Cutout Subjects on Martini
Prepare product, character, and compositing assets with AI background removal on Martini.
Related docs
Related reading
Comparisons
Frequently asked questions
Which model has the best camera control overall?
Sora 2 has the broadest motion vocabulary and the strongest long-take coherence — best for establishing shots and complex multi-axis moves. Kling 3 leads on subject-centered camera language (orbits, hero pushes, rack focus). Runway Gen-4 owns editorial lateral and crane work. Google Veo wins photoreal natural-light tracking. The right answer is per-shot, not per-project.
Can I specify the lens length or focal distance?
Sora 2, Kling 3, and Runway Gen-4 respond well to lens-length and focal-distance prompts ("50mm prime, slow push-in, eye-level"). For exact technical control, layer the spec into the camera-direction prompt — the cinematic-class models read those cues and adjust the apparent depth and field of view.
How do I keep the subject identity stable during an orbit or push?
Pin a high-resolution reference image and feed it into the video node alongside the camera-move prompt. Seedance 2 and Kling 3 are strongest at holding subject identity through complex moves. For very long takes, chain shorter clips with the same reference instead of relying on one continuous generation.
Can I direct multiple camera moves in one shot?
Most models perform best with one clear move per generation. For combination moves (push-then-orbit, dolly-then-tilt), generate each move as a separate cut and stitch them in the sequence builder. The edit reads better and the individual generations stay clean.
How does this compare to manual cinematography in Blender or Unreal?
Manual scene cinematography in 3D engines gives you complete control but requires a full asset pipeline. AI camera control gives you director-level shot planning over generated footage — much faster for ad work, social, and rapid pre-viz, but with less precision than an Unreal sequencer for final-pixel cinema.
Will the camera moves match what an editor expects in an NLE?
Yes — generated footage exports at clean frame rates and codecs through NLE export, and the camera moves preserve their intended timing in the timeline. Editors can trim, transition, and grade as they would with any other live-action plate.
Build it on the canvas
Open Martini and wire this workflow up in minutes. Free to start — no card required.