What Are Runaway Games
Branching video where the story bends to a player's choice is real. But nothing generates on demand — you build a pre-generated tree, not a live pipeline.
TL;DR
Nothing generates on demand, so a branching video experience is a pre-generated tree rather than a live pipeline: the AI work happens in production and the game just plays the right file. Size the tree on paper first, since branching multiplies fast, and let branches reconverge to keep the count down. Chain by feeding each clip's final frame in as the next clip's start image, use first-to-last frame on Seedance 2.0 or 2.0 Fast where branches must rejoin, and carry the character on one saved reference image.
Most people searching this mean one of two things: the Runway models, or branching video where the story bends to a player's choices.
The second is a real thing you can build. But not the way it is usually described.
Nothing generates on demand
This is the load-bearing correction, and getting it wrong wastes weeks.
A generation is not instant. It takes time, it costs credits, and it is quoted before it runs. So a game where the next clip synthesizes live in response to a player's click is not what you are building. The player would sit and wait, and every click would spend money.
What you build instead is a pre-generated tree. You produce every branch in advance, store the files, and the game just plays the right one. The AI work happens during production, not during play.
That reframe changes the whole approach. You are not designing a generation pipeline, you are designing a shot list with forks in it, and the number of clips is finite and knowable before you start.
Count the tree before you generate it
Branching gets expensive faster than people expect, because it multiplies.
Three choices at three depths is 27 end states and 39 clips. Three at four depths is 120 clips. Two branches at four depths is 30, which is a weekend. Three at five is 363, which is not.
Two things keep it manageable:
Reconverge. Let branches rejoin. Three choices that lead back to the same next scene costs four clips, not nine, and players rarely notice.
Vary the middle, not the ends. Shared opening, shared resolution, branching only through the middle act. Most of the felt agency comes from a small number of forks.
Decide the shape on paper first. The generation is the cheap part; regenerating a tree you sized wrong is not.
Chaining clips so the world holds
The mechanism is real and it is simple: the last frame of one clip becomes the first frame of the next.
Export the final frame, upload it as the starting image in image to video, and the next clip begins exactly where the previous one ended. Thirty-six models support image-to-video, so you have plenty of choice.
At a fork, the same start frame feeds several different prompts. That is what makes a branch point read as a branch point: one moment, several continuations, all visually continuous with what came before.
First-to-last frame on Seedance 2.0 or 2.0 Fast is stronger still where branches need to reconverge, because you fix both ends. Give each branch the same closing frame and they all land in the same place, which makes rejoining seamless.
The Long Video Generator chains automatically for the linear stretches between forks, so you are not hand-managing frames across a whole act.
Keeping the character the same
Chaining holds the world. It does not fully hold a face over many generations.
Thirteen models accept a character reference through reference to video. Build one reference image, save the exact file, and reuse it at every branch rather than regenerating something similar each time.
Seedance 2.0 and 2.0 Fast take up to nine references, which is the most identity control available and worth it for a character the player will see hundreds of times.
Wan 2.7 is the only video model with a seed. For a set of branches that should feel like variations of one moment rather than unrelated takes, that is the only genuine repeatability on offer.
Sound
Generate ambience per location rather than per clip. One loop for the forest, one for the interior, and the game crossfades on location change instead of on every clip boundary. Far fewer assets, and no audible seam at each cut.
Dialogue is a separate pass. Generate the silent clip first, then run it through lip sync with the recorded line.
Delivery
Download each clip and store it with your game, keyed by branch. Everything is a standard video file, so any engine or a plain HTML video element plays it.
Preload the next branch's clips during the current one. That is the difference between a game that feels responsive and one that stutters at every choice, and it is entirely a playback concern rather than a generation one.
What is not real
No guidance scale, no advanced settings panel. There is no panel of diffusion controls to lower a value in.
No camera path control. You cannot lock pan speed or tilt angle on Seedance or anything else. Camera behaviour comes from the prompt wording.
No seed on most models. Wan 2.7 only, so "store every generation seed" is not something you can do elsewhere.
No motion blur or sharpness parameter. Prompt tokens like "sharp motion" are just prompt words, not settings.
No real-time or interactive generation. Covered above, and it is the assumption most worth discarding early.
The short version
Design the tree on paper, keep it small, and let branches reconverge. Pre-generate everything. Chain by feeding each clip's last frame into the next, use first-to-last frame where branches must rejoin, and carry the character on a single saved reference image.
Start in Text to Video. The catalog is at Models, and Runway Gen 4.5 and Gen-3 Turbo are both in it if the models were what you were after.
Frequently Asked Questions
Can clips generate live as the player makes choices?▾
No, and this is the assumption worth discarding early. A generation takes time and costs credits, so a player clicking a choice would wait and every click would spend money. You pre-generate every branch in advance and the game plays the right file, which means the AI work happens during production rather than during play.
How many clips does a branching story actually need?▾
More than people expect, because it multiplies. Three choices at three depths is 39 clips, three at four depths is 120, and three at five is 363. Two branches at four depths is 30, which is a weekend of work. Decide the shape on paper first, since regenerating a tree you sized wrong is the expensive part.
How do I stop branches from exploding in number?▾
Let them reconverge. Three choices leading back to the same next scene costs four clips rather than nine, and players rarely notice. Vary the middle rather than the ends, with a shared opening and shared resolution, since most of the felt agency comes from a small number of forks.
How do I chain clips so the world stays continuous?▾
Export the final frame of one clip and upload it as the starting image for the next, which 36 image-to-video models support. At a fork the same start frame feeds several different prompts, which is what makes a branch point read as continuous with what came before.
What is the best way to make branches rejoin cleanly?▾
First-to-last frame on Seedance 2.0 or 2.0 Fast, the only two models that support it. You fix both ends, so giving each branch the same closing frame lands them all in the same place and makes reconvergence seamless.
How do I keep the character consistent across many clips?▾
Chaining holds the world but not fully a face over many generations. Thirteen models accept a character reference, and you should build one reference image and reuse that exact file at every branch. Seedance 2.0 and 2.0 Fast take up to nine references, the most identity control available.
Can I lock camera movement or set a guidance scale?▾
Neither exists. There is no camera path control, no pan speed or tilt angle to lock, and no guidance scale or advanced settings panel of diffusion controls. Camera behaviour comes entirely from prompt wording, and tokens like "sharp motion" are prompt words rather than settings.
How should I handle sound across branches?▾
Generate ambience per location rather than per clip, so one loop covers the forest and another the interior, and crossfade on location change instead of at every cut. That means far fewer assets and no audible seam at each boundary. Dialogue is a separate pass: generate the silent clip first, then run lip sync with the recorded line.
