AI model evaluations

eval forge.

Putting AI models to the test.
Explore the prompts. Compare the results.

The tests

001 / Three.js

Test #1 — Orbital Relic

Creative codingCollapseExpand

An interactive 3D artifact from a lost civilization, built in a single HTML file. The same prompt, given to three models.

The original prompt

Same prompt · All three models
You are a world-class creative technologist and Three.js expert. Your task is to create a complete, self-contained, production-quality interactive 3D experience in a single HTML file.
Core Concept
Build an immersive “Orbital Relic” — a procedurally generated, fully interactive 3D artifact that feels like a living museum piece from a lost civilization. The experience must combine:

A central floating relic (a complex, multi-layered geometric structure made of nested, rotating rings, crystalline cores, and energy filaments).
An orbiting planetary system of 5–7 smaller celestial bodies that react to the user’s actions.
Atmospheric particle systems, volumetric-style lighting, and subtle post-processing.
A sense of discovery, scale, and quiet wonder.

Technical Requirements (all mandatory)

Single self-contained HTML file — no external files. Use only Three.js (latest via CDN) and optionally one additional lightweight library (e.g., GSAP or a small post-processing helper) if needed. Everything else must be pure JavaScript.
Full 3D scene with:
Perspective camera + smooth OrbitControls (with damping).
Physically-based lighting (at least one key directional light + ambient + several colored point lights that pulse).
Soft shadows and contact shadows where appropriate.
High-quality materials (MeshStandardMaterial / MeshPhysicalMaterial with transmission, iridescence, or clearcoat on the central relic).

Procedural generation:
The central relic and its orbiting bodies must be generated algorithmically (no hardcoded vertex lists for the main structures). Use mathematical functions, noise, or recursive geometry.

Interactivity (all required):
Click / tap any orbiting body → it expands into a detailed holographic information panel floating in 3D space (with title, short lore text, and animated data visualization).
Double-click the central relic → triggers a dramatic transformation sequence (rings expand, core pulses, particles erupt, camera slowly pulls back).
Mouse / touch drag to orbit, scroll to zoom (with sensible limits).
Keyboard: Spacebar freezes/unfreezes time, R resets the scene elegantly.

Animation & Atmosphere:
Continuous gentle rotation of all major elements with different periods.
Breathing / pulsing scale on the core.
Soft particle systems (dust, energy sparks, distant stars).
Subtle camera drift when idle.

Visual Polish:
Beautiful color palette (deep space blues, ethereal golds, soft cyan accents).
Responsive design — works on desktop and mobile (touch controls).
Loading screen with a minimal progress indicator that fades out elegantly.
Clean typography for any 3D or 2D UI text (use a system font stack or a single Google Font via CDN).

Performance & Quality:
Target 60 fps on modern hardware.
Use instancing or efficient geometry where possible.
Proper disposal of objects if the scene is reset.
No console errors.


Experience Goals
When the page loads, the user should feel they have discovered something ancient and alive. Interaction should feel rewarding and slightly mysterious. The piece should be beautiful enough to leave running as a digital art installation.
Output Format
Return only the complete, ready-to-run HTML file. No explanations, no markdown wrappers, no comments outside the code itself (except brief, useful ones inside the JS). The file must open directly in a browser and work immediately.
Make it exceptional. Push the limits of what a single-file Three.js experience can be.

Test #2 — Lighthouse Cove

Voxel · Three.jsCollapseExpand

Eight models, one prompt, one self-contained HTML file each: a floating voxel island with a spinning lighthouse beam, a fishing village, boats, seagulls and a night mode. Scored on looks, accuracy, prompt-following and time.

The original prompt

Same prompt · All eight models
Build a single-file HTML page using Three.js (load from cdnjs) that renders an interactive 3D voxel scene: "Lighthouse Cove".

Scene:
- A floating island with sandy beaches on one side, grassy cliffs on the other, and a jagged rocky underside with waterfalls spilling off the edge into the void
- A tall striped lighthouse on the cliff with a rotating beam of light
- A tiny fishing village: 4 to 5 colorful cottages, a wooden dock, two little sailboats bobbing in a voxel water pool
- Palm trees, bushes, flower patches, barrels, crates, a campfire with animated flickering
- 3 small floating islets nearby, each with its own mini scene (one with a lone tree, one with a stack of treasure chests)
- Seagulls (simple voxel birds) circling the lighthouse

Visuals:
- Sunset sky gradient, blocky clouds drifting, warm directional light with soft shadows, light fog
- InstancedMesh for performance, subtle per-voxel color variation so nothing looks flat
- Water with gentle animated color shimmer

Interaction:
- Drag to rotate, scroll to zoom, auto-rotate when idle
- UI overlay with a title card top left and buttons for day/night, auto-rotate, and a toggle for the lighthouse beam
- Night mode: starry sky, glowing windows, the lighthouse beam becomes the main light source

Generate everything procedurally with helper functions (addCottage(), addPalmTree(), addBoat()). Make it dense, polished, and smooth at 60fps. Output the complete working code.
At the top, add a small black bar with [Which model], [Date], [Total time to finish], and at the very right side of the black topbar, add a link saying Back to Eval Forge homepage (https://eval-forge.github.io)
Make this all into one single contained HTML file that I can download from the chat right here.