Guides
Every section we cover, with the explainers that go with it.
After the generation
Model
-
Frame interpolation and smoother motion
Frame interpolation smooths generated video but treats subject drift and limb mutation as real motion—amplifying the very artefacts you want to hide.
-
Upscaling generated video
AI upscalers can fix soft generated video—but they also amplify latent-space flicker and hallucinate detail that was never in the source image.
Choosing between them
Model
-
Runway compared with Pika
Runway Gen-3 Alpha and Pika 2.0 offer fundamentally different control surfaces for image-to-video: director-mode camera rigs versus region-mask selective animation.
-
Veo compared with Kling
Google Veo 2 and Kling 1..6 compared on motion fidelity, prompt adherence, clip length, audio, and access paths—with dated evidence and the caveat that the answer moves.
Free and local
Model
-
The open-weights models worth your time
Stable Video Diffusion, LTX-Video, and Wan 2.1 are the open-weights image-to-video models worth your time. Here are their licences, VRAM needs, and what they produce.
-
What the free tiers actually allow
Free tiers come with watermarks, resolution caps, deprioritised queues, and commercial-use restrictions. Here is what each provider's terms actually say, dated and sourced.
Prompting for motion
Model
-
Camera moves and how to ask for them
The camera-motion terms—dolly, pan, orbit, crane, push-in—that image-to-video models interpret, and which models offer discrete camera controls versus text-prompt interpretation.
-
How to structure an image-to-video prompt
A fixed five-element image-to-video prompt template—subject, motion, camera, lighting, duration—and why the order determines what the model actually animates.
Rights, rules and disclosure
Model
-
Using generated video commercially
Which image-to-video models and tiers permit commercial use, and the specific conditions for client work, resale, and platform monetisation.
-
Who owns a generated video
Who legally owns AI-generated video output under US, UK, and EU copyright law, and what each major model's terms of service say about ownership.
Shot recipes
Model
-
A product turntable from one photograph
How to animate a product turntable spin from one photo, with the source-image rules, prompt, and diagnosis of label warping and reflective surface failures.
-
The earth zoom-out shot
How to generate an orbital earth zoom-out from a still image, with the source requirements, prompt, and diagnosis of the seam and parallax failures that expose the shot.
The models
Model
-
Google Veo
Google Veo 2 accepts a still image and generates video with high motion fidelity through VideoFX and Vertex AI. Every capability claim dated against Google DeepMind documentation.
-
Kling
Kling 1.6 adds start-and-end-frame specification to image-to-video generation, accessed through Kuaishou's platform. Every capability dated, queue times explained, failure modes documented.
-
Runway Gen-4
Runway Gen-4 exposes director-mode camera controls—pan, tilt, zoom, roll, dolly, truck—for image-to-video generation. Every capability dated, credit model explained, failure modes documented.
The source image
Model
-
Composing a still that can move
How to compose a still image with headroom, depth cues, and foreground separation so an image-to-video model has space to generate convincing motion.
-
Resolution and aspect ratio
What resolution and aspect ratio each image-to-video model expects, and the distortion and softening artefacts that appear when the source image does not match.
