Generation, and the rest of the toolkit
Text-to-video and image-to-video are the headline. The more useful half is everything around them — the tools that take a generated or filmed clip and make it usable.
Camera controls let you specify a pan, orbit, zoom or dolly rather than describing one and hoping. That is the difference between directing a shot and rolling dice on it, and it is the feature professionals reach for first.
Motion brush lets you mark which part of a still should move and in which direction. Combined with image-to-video it is the most controllable workflow available anywhere in this category: you approve the frame, then specify the motion, rather than gambling on both at once.
Video-to-video restyles existing footage, keeping the motion and replacing the look. For teams with real footage and a stylistic problem, that is worth more than generating from nothing.
Image-to-video is the workflow that actually works
Worth stating separately because it applies across this whole category and few people arrive at it immediately.
Text-to-video asks the model to solve two hard problems simultaneously: compose a scene you will accept, and animate it plausibly. The failure rate compounds.
Generating a still first — in Midjourney, FLUX or anything else — approving it, and then animating that specific frame separates the problems. You have already accepted the composition, the colour and the subject, so the only remaining variable is motion.
Almost everyone doing serious work with these tools converges on this, and it saves more credits than any other single habit.
The four-second problem
Generated clips are short — seconds, not minutes — and that shapes everything about how the tool is used.
You are not making a video. You are making shots, and assembling them somewhere else. Continuity between clips is the hard part: the same character, the same lighting and the same room across three generations is not guaranteed, and no amount of prompting fully solves it.
This is a property of the current generation of video models rather than a Runway failing — Sora, Veo and Kling all share it. But it is the single most common reason people conclude these tools are not ready for their work, and it is better known before paying than after.
The practical consequences: plan for cuts rather than continuous action, use the same source still across related shots where possible, and treat generated footage as inserts within edited work rather than as the work itself.
Credits, and how to budget them honestly
- A browser and an account — nothing runs locally, and there is no GPU requirement.
- Credits consumed per second of generated video. This is the constraint that decides whether Runway fits a project. Higher-quality models cost more per second.
- A free tier for evaluation, exhausted quickly. Sustained work is a paid plan.
- Source footage or stills if you intend to use the editing tools rather than generate from text.
Budget by the failed attempts, not the finished shot. A usable four seconds commonly costs several generations, so a realistic estimate is three to six times the naive per-second figure. Teams that price a project on the assumption of one generation per shot lose money on it.
Where the physics still gives it away
Failures across this category are consistent and worth recognising before you present output to a client.
Hands gain or lose fingers. Objects pass through each other. Poured liquid does not behave like liquid. Text on signage becomes gibberish. A person walking away and turning back may not be the same person. Reflections do not correspond to what is being reflected.
Generated video survives a glance and rarely survives attention. That decides where it can be used: background plates, brief inserts, stylised or abstract sequences — not anything a viewer studies.
Rights, and what is unresolved
Runway grants users rights to output subject to its terms, and those terms vary by plan. Read the current version before commercial use.
The separate and unresolved question is training data, which is under dispute across this industry generally. A grant from a vendor covers what the vendor can grant. For advertising, broadcast or anything with rights clearance attached, that distinction is worth raising with whoever signs it off — before the shoot, not after.
Who this is really for
Video professionals with an existing pipeline who want AI for specific shots — an establishing image, a background plate, an effect that would otherwise need a 3D artist.
Advertising and content teams producing short-form work where a few seconds is the deliverable rather than a fragment of one.
Editors using it as a repair tool: removing an object, extending a background, cleaning a plate. That work is unglamorous and is where the credits are best spent.
It is a poor fit for anyone expecting to generate a finished narrative video, for projects that cannot absorb variable per-second costs, and for material that cannot be uploaded.
What you get that pure generators lack
- Editing tools around the generation, so output can be fixed rather than only regenerated.
- Directable camera movement and motion brushing, instead of describing motion in a prompt.
- Video-to-video restyling, which works on footage you already own.
- A real track record in professional production, which is rare here.
- Nothing to install — it runs in a browser on any machine.
What it still gets wrong
- Clip length, with unreliable continuity between generations.
- Cost escalates quietly — per-second pricing plus iteration is hard to estimate in advance.
- Physics and hands remain unreliable, and the failures are obvious to any viewer.
- Text within video is poor, so signage and captions must be added afterwards.
- Hosted only. Footage is uploaded, ruling it out for material under strict NDA.
What you keep when you stop
Downloaded clips. Projects, generation history and any settings live in Runway and do not export.
The practical habit that matters here: download everything worth keeping as you go, and keep a note of the source still and prompt for shots you may need to match later. Because continuity across generations is already unreliable, losing the exact inputs that produced a shot means you cannot approximate it again.
If Runway is not the right fit
- Sora — stronger raw generation, far less tooling around it.
- Kling AI — competitive quality, materially cheaper per generation.
- Luma Dream Machine — fast and cheap for exploring ideas before spending here.
- Descript — the better tool if the job is editing recorded video, not generating it.
Compiled from Runway’s documentation and public sources. We have not hands-on tested this tool. Last reviewed 16 August 2026.
