Text-to-video and image-to-video may sit next to each other in an interface, but they solve different production problems. Text-to-video gives the model broad control over the opening composition; image-to-video commits it to a supplied visual state. Making this choice before buying credits can reduce avoidable rework and make rights management easier.
When text is enough
A text-only starting point is reasonable for atmospheric cutaways, abstract scenes and ideas whose exact object positions are negotiable. Describe a single action, camera motion, lighting and tone rather than a collection of unrelated shots. The benefit is fast visual discovery; the cost is less precise initial framing and potentially inconsistent character identity across separate generations. Gen-4.5’s current Runway documentation distinguishes a text-to-video mode from an image-led mode, but its earlier Gen-4 family required an input image. Version matters.
When an image is a specification
A product photo, illustration or approved composition can define colors, proportions, wardrobe and camera angle. That makes image-to-video attractive for e-commerce, stylized explainers and episodic characters. Prepare the opening frame at the intended ratio, remove irrelevant background objects and inspect logos or sensitive typography before upload. The model can still deform the image over time, so review the output frame by frame rather than assuming input pixels are locked.
Consider a ten-second clip showing a new lamp switching on. If any plausible lamp is acceptable, text-to-video may suffice. If the lamp is a specific SKU with exact geometry, begin from an approved product photo and keep the labels composited in the edit. The image does not provide an automatic product-truth guarantee once motion begins.
Reference and keyframe controls vary
A start frame, a character reference, a last frame and a composition guide are not interchangeable inputs. A start frame defines the opening state. A last frame constrains the landing. A character reference can influence identity without necessarily pinning the framing. Video reference may carry motion rather than appearance. Check the selected model documentation: Luma Ray3.14 excludes character-reference support; Seedance 2.5’s announcement describes extensive mixed-media references; Gemini Omni Flash allows interpolation with first and last images.
A low-waste experiment
Create one safe, fictional shot concept and run a text variant and an image-led variant only where both are offered on the same model and plan. Judge each against predefined acceptance criteria: subject shape, action readability, camera fidelity, facial stability and ease of editing. Record billable seconds and rejections, not just the prettiest still. Never advertise that one mode is universally superior without evidence.
Finishing
Keep source images and their ownership records with the final shot. Add accurate typography, captions and licensed sound after rendering when precision matters. A generated video is one asset inside a production workflow, not the entire finished film.
Worked production example
For an advertisement about a specific pair of headphones, a written prompt may invent the product geometry. An approved product photograph is a better opening anchor, but it still does not ensure the hinge or branding survives motion. Generate a short turn with the headband mostly stationary and inspect each frame. For a fantasy cloudscape behind the product, text generation might be sufficient. Keep the accurate object as a licensed still or composited element if the model cannot preserve its geometry, rather than using a fictionalized product as a factual sales image.