All videos start from a fixed reference frame. It can be different one but I've choosen a few images — your prompt controls the motion, action, and atmosphere. If the subject isn't doing what you want, its mostly likely hasn't been trained on it, and extra trainings LORAs(Low Rank Adaptation) will be needed.

• Limited Camera movement and angle
• Actions and gestures
• Lighting and mood
• Weather and environment
• Speed and energy of the scene

• Character identity and outfit
• Base scene composition

Output is locked at 480 × 480 px. Higher resolutions (720p, 1080p) are supported by the model but currently disabled — they require significantly more VRAM and generation time.

This uses a local model running on local hardware. No data is sent to OpenAI, Sora, Runway, or any external service. Generation is private and offline except shared here.

• Describe motion first, then mood
• Be specific about camera direction
• Avoid describing characters by name