AI Workflows for 3D Worlds, Games, and Video

A group of creators showcased practical AI workflows for 3D worlds, games, and video production. The common thread is not a single prompt that outputs a final result, but a pipeline where creative intent is defined first, then translated into implementable structure, and in some cases handed off to agents for continued building and revision. That makes these posts notable as examples of repeatable production workflows rather than simple demos.

Core 3D world workflow

@techartist_ described a step-by-step process for building a 3D world with AI and Three.js. The starting point is a natural-language description that goes beyond surface appearance to include mood, composition, interaction, animation, realism, and which parts should be procedurally generated versus sourced externally. Before coding, the workflow asks AI to create a reference image that is actually buildable in code, meant to anchor the visual target rather than be copied pixel by pixel.

In a second stage, AI turns the concept into detailed coding prompts covering scene structure, materials, lighting, animation, interaction, and technical constraints. It then helps identify a suitable stack based on the idea, with examples including Three.js, GLSL/TSL, GSAP, post-processing, procedural geometry, particle systems, physics, and animation. @techartist_ stressed that the tech stack should serve the creative goal, not dictate it.

Agent-based production experiments

Several posts focused on how agents can take over more of the production loop. @piotrbinkowski said a first-pass piece improved after being handed to Agent One for further refinement. In his account, the agent could spot issues that had not been explicitly pointed out, including removing a gargoyle that should not have appeared in the final scene, and could carry out many revisions once the request was clearly described. He later added that this agent-driven creation flow could continue with further tasks.

@majidmanzarpour highlighted a more automated setup: with a single prompt plus @threejs agent skills, the system could keep building a game while he was away and send progress screenshots to his phone. @miran248 shared a one-week, almost no-code experiment building a 3D procedural world that began as a 2D zombie game and shifted toward procedural generation, using combinations such as opencode + deepseek v4 pro and claude + fable 5 during the process.

Extending the logic to AI video

On the video side, @taherdhanera argued that many current AI video generation pain points are really previs problems. His proposed workflow is to first use 3D software such as Blender to resolve physical collisions, lock camera trajectories, and complete editing, then pass the result to a video model like Seedance 2.0 for final image generation. This mirrors the broader pattern seen across the cluster: establish structure, motion, and constraints first, and let generative models handle the final rendering layer.

2026-07-14 ~ 2026-07-16 · 12 related posts