Testing 3D Water Rendering with GPT-5.6 and Fable: Impressive Prototypes but Systemic Limits

Cedric Chee shared a series of tests using GPT-5.6 Sol combined with Fable to tackle realistic 3D water rendering. This workflow demonstrates generative AI's ability to accelerate specific implementations in graphics development, while also exposing clear boundaries and performance costs when aiming for complete, physically realistic systems.

Project and Workflow

His general pipeline involves using GPT-5.6 Sol to write Three.js TSL water shaders, which are then handed over to Fable for scene integration. This was applied to two main scenarios: enhancing the water realism for the Golden Gate Bridge scene in Arena (which required him to spend a week studying modern water rendering), and recreating a realistic island and yacht scene based on Apo Island aerial photos using Blender and Fable in about two hours.

Capability Boundaries

Cedric Chee noted that GPT-5.6 Sol handles writing water shaders easily, even producing stunning demos. However, once the goal shifts to a "complete ocean rendering system," the model struggles. Specific challenges include FFT for physically realistic waves, wind-driven wave generation, and underwater effects like caustics and refraction. The AI is thus better suited for prototyping rather than delivering a complete, physically accurate ocean solution.

Performance Costs

He also highlighted a hidden cost: as visual quality improved, the FPS dropped by about 50%. This indicates that combining Sol and Fable to balance high graphics quality and rendering speed hits significant performance optimization bottlenecks.

2026-07-19 ~ 2026-07-20 · 7 related posts

1 near-duplicate retellings: cedric_chee