I've been playing with the Waypoint model from Overworld, and want to share some prompting tips. 

Create Multiple Image Prompts from I2V

The Waypoint model performs much better if you pass in multiple images. I've had luck with more than 10 images. The images should be similar in style, and ideally come from the same video clip, within the same scene.

If you only have a single image, you can use any I2V model to generate a video, and then pull out image frames from the video as prompts. You can use ffmpeg for pulling out the frames: `ffmpeg -i video.mp4 -vf fps=2 frame_%02d.png`

Add HUD / mini maps to the Image

This tip may sound strange. Since the model is trained on video game footage, any visual elements that signal a video game will result in a more consistent generation. You can add a HUD / mini map by using an image editing model like Nano Banana or Qwen-image-edit.

You can also add things like health bar or game status. They all help make the generation more consistent.

Text Prompt Tips

1. Cover multiple dimensions

Describe visual style, art direction, atmosphere, mood, game mechanics, environment, player objectives.

2. Descriptive depth

Go beyond surface-level observations. Instead of "a bright room", describe the quality of the brightness, light sources, shadows, environmental details, etc.

Prompt template (You can paste this into a LLM to generate better text prompts)

[GENRE/PERSPECTIVE] The [clip/video/footage] [shows/depicts/displays] a [perspective] within [environment]. [ART & ATMOSPHERE] The visual style is [style], with [color details], creating [atmosphere]. [ENVIRONMENT] The environment [description], with [notable features]. [MECHANICS] The player [movement/interaction description], suggesting [gameplay systems]. [OBJECTIVES] The objective appears to be [goal], as indicated by [evidence].

Happy exploring!