What it does • Lets you roleplay with fixed AI characters in real time. • Tracks character positions, obstacles, and state so scenes stay consistent. • Generates cinematic, frame-consistent video clips from the latest story beat. • Uses obstacle logic to maintain continuity (if a troll is on fire in scene 1, it stays on fire until the plot changes).
Tech • Story + dialogue: GPT with a structured RP prompt system. • Video: Wan 2.2 for text-to-video generation. • State tracking: Previous messages.
Why We made it Most LLM roleplay feels like a chat log. We wanted something that feels more like a movie — immersive, reactive, and visual. We are aiming for Interactive Movie than Generative Games.
Observation • Wan 2.2 consistently adheres to prompts better than Wan 2.1. • GPT-4.1 is excellent at text RP, maintaining better prompt adherence than 4o.
Current limitations • Video generation takes 30s per clip at 480p so can use once per turn rather than completely displace text. • Wan 2.2 still struggles with verbose aciton shots since it can generate only 5s video. • Character consitency forced us to only focus on main obstacles in a scene like monsters.
Future direction • Implement streaming video gen with Wan 2.2 following self-forcing, causvid papers • Character consistent workflows with comfyUI • Elaborate RP to track health, location map and more characters • More engaging Screenplay and Story lines
Demo - https://shorturl.at/L7zpR — Just pick a character and start.
Try it out here - https://www.reveriedr.com
Would love feedback on: • How to make the video feel even more like a live film set. • Ideas for better state tracking across long sessions. • Any moments where the immersion breaks for you.
Disclaimer - This post was drafted by LLM.