DreamGen

Every role-play session runs on a scenario, which acts like a game master's guide with a character sheet attached. A scenario holds the plot, setting, writing style, characters, locations, objects, opening scene, and example interactions that keep the AI consistent. You can clone a public scenario and edit your copy, build one from scratch in the editor, or use the scenario wizard. The wizard walks you through brainstorming characters, settings, and plot outlines when you do not want to start from nothing.

Instructions are a distinct interaction type in DreamGen that are not part of the story itself; they are your directions to the AI. You can describe the next scene in a sentence, set the location and time, tell a character to react a certain way, or cap how long the next reply should run. You can also change the tone and pacing on demand, asking for a slow, tense scene or a fast, action-focused one. This is the core of what DreamGen calls steering: the AI follows your lead rather than running the plot on its own.

DreamGen builds a role-play out of three interaction types: narrative for scene-setting, character messages with dialogue and action, and instructions. Behavior modes decide who drives. Take-Turns, the default, has the AI detect when it is your character's turn and stop, while it can still introduce new characters on its own. Spectate hands every character to the AI so you watch the story play out, and NPC Only keeps the AI to defined characters without turn detection. You can switch modes mid-session.

Role-plays are stored as a tree rather than a straight line. Retrying a reply keeps the original and the new version side by side, so you can flip between them with arrow buttons. You can retry with a different model, make a reply longer or shorter, or ask the AI to rewrite it in a specific way. Forking copies a session from a chosen point into an independent branch, which suits testing a major plot change without losing the original run. Pinning also lifts critical interactions higher in the AI's memory.

The assistant is a helper that can see your scenario and the recent session history. You ask it questions in a chat window, like what should happen next, three ways a conflict could resolve, or how a given character would react to a situation. It also helps with world-building, suggesting what might sit in the ruins the characters just found or how a kingdom's politics work. It is a separate voice from the story itself, aimed at unblocking you rather than writing the scene.

DreamGen runs its own first-party models, and the lineup shifts as newer ones land. Current examples include GLM 5.2, GLM 4.7, Gemma 4 31B, and Gemma 4 26B A4B. Each model carries its own context window and credit cost, so you can trade quality against credits. You can switch models mid-role-play or retry a single interaction with a different one. A built-in token counter in the scenario editor shows how much of a model's window your scenario definition eats up.

Alongside text, DreamGen offers the AniMuse image model line, aimed at the same creative, character-driven audience. The documentation for image generation was still marked as under construction, pointing users to the AniMuse guide. For anyone building visual novel content or wanting character art to match a scenario, this keeps illustration and writing in one place instead of sending you to a separate service.

Finished work does not stay locked in the platform. You can export a role-play as Markdown for sharing or as JSON to re-import later, and clone a whole session to experiment without touching the original. The Choose The Next Step feature, also called CYOA, suggests next moves below the latest reply; click one to load it into the input box, edit it, or ignore it and write your own. CYOA can be switched on for any session, even one you already started.