Text to video vs image to video, per Google, Kling and Runway docs: when to use each, the quality and consistency trade-offs, prices and a workflow.
Method. Runway, Google, Kling and LTX prompting guides, manuals and price lists read November 18, 2026, with the price table and Flow's model page re-checked 19 November 2026; prices are single-generation list prices and the cost chart is our arithmetic. Vinc makes an AI film studio.
Text to video vs image to video: the short answer
In text to video vs image to video, the difference is who decides the first frame. In text-to-video the model composes the shot from your words. In image-to-video your image becomes the first frame, and the prompt only describes motion. Vendors recommend text-to-video when exact consistency doesn't matter (B-roll, plates, complex motion), and image-to-video when a character, product or composition has to match across shots. On Google, Kling, Runway and LTX the per-second video price is the same for both modes. Image-to-video adds only the cost of making the still.
What is the difference between text-to-video and image-to-video?
Text-to-video. Runway's Text to Video Prompting Guide says these models "transform written descriptions of scenes into video". Prompts need "Visual descriptions" (what we see, where, and how it looks) and "Motion descriptions" (how the scene moves).
Image-to-video. Runway's Image to Video Prompting Guide says "your image defines composition, subject matter, lighting, and style". The prompt describes "the motion, camera work, and temporal progression". Google's best practices say the same: the image is the first frame, so "prompt for motion only".
"You only have text and want to generate a video from it"
"Content generated entirely from the text input"
Image-to-Video
"Key visual moments are already defined and you need precise control over how they appear in the final video"
"The visual state at specific points on the video timeline"
Omni Reference
"You have images, videos, or other assets to reference, or only a creative direction to start with"
"The appearance, style, actions, and camera movements of the generated content"
There's also a middle ground. Veo 3.1 takes reference images of a character, object, scene or style that guide the look without fixing the first frame; Google's Veo 3.1 prompting guide calls this "ingredients to video", and Kling's Omni Reference accepts up to 10 images per task. Both also accept first and last frames: you supply both ends and the model generates the motion between.
Which gives better quality, text-to-video or image-to-video?
No vendor we read benchmarks its own text and image modes against each other. The docs do say where each mode tends to break.
Image-to-video inherits the image, flaws included. Runway warns that "Visual artifacts, such as blurry hands or faces, may be intensified once your image is transformed into a video." Google puts it positively: "A sharp, clear, and well-composed photograph yields a more coherent and higher-quality video", because "the stronger the start, the better the finish". Google recommends 720p (1280×720) or higher, in 16:9 or 9:16, per its image-to-video reference.
Image-to-video fights implied motion. Runway's FAQ notes that input images "can contain implied motion through elements like motion blur, mid-action elements and poses, or directional lines". Prompting against those cues "may require more iteration". Its example is a car still with dust clouds and motion blur that would not stay parked until the cues were removed.
Text-to-video handles complex motion more freely. Runway's FAQ: "Because Text to Video isn't constrained to a specific starting scene (like Image to Video is), it can handle complex motion sequences more effectively."
So image-to-video quality rides on the still, text-to-video on the prompt. Pick the one you can control more easily.
Which is better for character consistency?
Image-to-video, or reference images. The vendors are explicit on this.
Runway positions text-to-video for cases "when exact character or scene consistency isn't the priority—like creating B-roll, stock effects, or background plates."
Kling's image-to-video is the mode for "precise control" over the visual state at points on the timeline. Its Omni Reference asks you to "declare the facial features, hairstyles, clothing, and other traits to keep consistent" for each character image.
Google's reference images exist "to maintain a consistent aesthetic across multiple shots".
If you must stay in text-to-video, Google suggests pasting "the entire, unchanged character description into your prompt for every new scene" and reusing the same seed. Both help, but neither fixes a face the way a shared image does. Wardrobe is part of this too; see Costume and Wardrobe Continuity.
How much control does each mode give?
Text-to-video
Image-to-video (first frame)
First and last frame
Composition and framing
Described in words; model decides
Fixed by your image
Fixed at both ends
Character identity
Description plus seed (Google)
Fixed by your image
Fixed at both ends
Lighting and colour
Described
Fixed by your image
Fixed at both ends
Motion and camera
Described
Described (the whole prompt)
Described (the path between)
Where it struggles
Matching shots to each other
Motion that contradicts the image
Ends that are too far apart
Typical prompt length
Longer: Runway says text-to-video prompts "need to describe both the visual elements and the motion in detail"
Shorter, motion only
Shorter, the transition only
Runway lists four cases where an image-to-video prompt should describe visuals: "Introducing an element not present in the image", "Dramatic changes from the starting image", "Specifying transformation details", and "Specifying interactions between two (or more) elements". Otherwise, leave description to the image. Google warns that "Redundant prompts confuse the model".
The difference is the still: Google lists Imagen 4 at $0.04 per image ($0.02 for Imagen 4 Fast), Kling lists Kling Image 4.0 at $0.063 per 2K image. Image-to-video often saves money anyway, because the costly part of AI video is retries, and a still is cheap to redo.
Cost of iterating: four candidate stills vs one video attempt
4 Imagen 4 stills$0.16
4 Kling Image 4.0 stills$0.25
1 Kling 4.0 clip, 10 s$2.35
1 Veo 3.1 clip, 8 s$3.20
Our arithmetic from list prices checked 19 Nov 2026: Imagen 4 at $0.04 per image and Kling Image 4.0 at $0.063 per 2K image; Veo 3.1 at $0.40/s and Kling 4.0 at $0.235/s (1080p).
Four rejected stills cost less than a tenth of one rejected 8-second Veo take. Settle composition, wardrobe and lighting on stills and fewer video generations go to finding the shot; see What an AI Shot Really Costs.
A workflow that uses both
No vendor prescribes a full pipeline; these steps put their recommendations in order.
A shot workflow built from vendor guidance: lock characters and frames as stills, animate with motion-only prompts, use text-to-video where consistency doesn't matter.
Make character and location references. These become Veo reference images or Kling @Image assets. Name each one's role in the prompt, as Kling requires.
Animate with motion-only prompts. Refer to "the subject" rather than re-describing them. For a precise transition, add a last frame.
Use text-to-video for B-roll and plates. That is Runway's recommended use, and these shots don't need to match a face.
Bridge longer shots. Runway's guide suggests "extracting the last frame of a completed generation and using that as the image input for a new video", then trimming the shared frame in the edit.
This is also the order Vinc's studio uses, as its landing page describes it: "Approve each keyframe, then animate it."
How we compiled this
On 18 November 2026 we read the vendors' own prompting guides, FAQs and pricing pages (Runway, Google, Kling, LTX), listed under Sources. Prices are list prices for one generation, and the chart is our arithmetic. No vendor compares quality between its own modes, so we don't claim one; Vinc makes an AI film studio.
AI Music for Film: Suno, Udio and Other Generators on Licences, Ownership, Lawsuits and Cue Sheets
Whether an AI-generated track can go in your film depends on the plan you made it on, whether the terms exclude film and TV, and whether anyone can own it at all. This guide compares the published terms of Suno, Udio, ElevenLabs Music, Stable Audio and AIVA, sets out the label lawsuits as reported, and explains what US copyright law and North American PRO rules mean for your cue sheet.