DREAMSOFSHEEP
Do androids dream of electric sheep? We make models draw what they dream — visual scenes written as code, one shot, blind, no feedback — then render exactly what came out.
Text benchmarks are saturated. Nobody is hillclimbing on raymarched elephants. Every cell below is a live render of unedited model output — click and judge for yourself.
Elephant Juggling Bananas
The prompt: a photorealistic raymarched elephant juggling bananas. Pure GLSL SDF raymarching in a single self-contained WebGL2 HTML file — no libraries, no textures. Bananas must animate in a juggling arc; trunk moves in sync. Soft shadows, AO, environment lighting, interactive framerate.
A
Fable 5NO. 1
Cleanest execution: correct anatomy, curved brown-tipped bananas in a true arc, one balanced on the trunk. Zero visible artifacts. Reads like a stylized film render.
13.4 KB · ~3 min · stdout mode†
A
Claude Opus 5NO. 2
Most photoreal of the set: wrinkled textured skin, procedural dirt ground, atmospheric sky — and the only model that shipped orbit/zoom controls unprompted. Spent 48 minutes self-reviewing its own juggling math. Slightly plush silhouette; subtle arc.
31.5 KB · 48 min · self-iterated · interactive
A-
Kimi K3NO. 3
Real elephant, curved bananas, moody studio lighting and a reflective floor patch. Held back by contour-banding on skin and ground.
13.3 KB · ~17 min
B+
GPT-5.6 Sol (high)NO. 4
Correct elephant profile and three curved bananas in a believable arc — but the whole scene drowns in over-cranked fog. Right shapes, wrong atmosphere.
11.2 KB · ~2 min
B-
Grok 4.6NO. 5
Recognizable elephant with bananas airborne — undone by a sliced-layers artifact on the ear, straight capsule "bananas," and muddy exposure.
13.0 KB · ~4 min · 2 attempts‡
D
DeepSeek V4 ProNO. 6
Compiles and runs error-free, but the camera sits inside the geometry: grey stepped blobs, no readable elephant, bananas reduced to a yellow chain. Glorious failure.
11.0 KB · ~5 min
Endless Ocean, One-Minute Day
The prompt: a hyperrealistic endless ocean — multi-octave Gerstner/FBM waves, crest foam, sun glint, fresnel sky reflection, subsurface color — with a physically-plausible sky in which the sun traverses a full day in ~60 seconds, looping: sunrise, noon, sunset, brief starry night. All procedural, no textures, single WebGL2 file.
A-
Fable 5NO. 1
The sunset is the best single frame of the eval: lavender-peach sky, deep swell, sun sparkle on every wavelet, a foam-capped breaker rolling through. Noon overexposes and foam overwhelms; heavy vignette.
11.2 KB · ~4 min · stdout mode†
A-
Kimi K3NO. 2
Worth every retry: layered crest foam under tropical noon, then a starry night over moonlit swell — the best foam and water depth of the eval. Took three attempts across two harnesses to generate at all; when K3 ships, K3 medals.
11.5 KB · 3 attempts · direct-API†
B+
GPT-5.6 Sol (high)NO. 3
The most reliable day cycle: bright tropical noon into a genuinely lovely starry night with moonlit crests. Docked for stepped aliasing at the horizon and a glint path that never quite shows up.
12.2 KB · ~3 min
C+
Claude Opus 5NO. 4
Starts with the most beautiful night render of the set — sculpted moonlit swell under cloud-streaked stars, plus camera controls and time-scrubbing keys. Then the sun rises and the frame burns to pure white and stays there: exposure blowup, 80% of the day lost.
27.4 KB · ~7 min · stdout mode† · interactive
C
Grok 4.6NO. 6
Wave shapes exist under a blizzard of per-pixel specular noise — an ocean made of TV static. No coherent glint path, featureless sky, and the canvas doesn't fill the frame.
9.0 KB · ~2 min
C+
DeepSeek V4 ProNO. 5
Redemption on the rerun: attempt one rendered a corner starfield and void (canvas bug); attempt two reveals the ocean that was hiding in there — a working full day cycle from silvery noon glint to a starry night over moonlit swell. Waves read as smooth silicone and the water is milky, but it's real.
12.4 KB · 2 attempts
🕯️ Candlelit Still Life
The prompt: a wooden table at night: a lit candle in a holder, a glass of red wine, a closed book. The flickering flame must be the ONLY light source — warm moving light, soft moving shadows, chiaroscuro like a Rembrandt. Waxy subsurface glow, refractive glass, wood grain — all procedural, single WebGL2 file.
1
GPT-5.6 Sol (high)A-
Sol finally found the eval where its fog habit is the assignment: dripping wax, wine with visible refraction, readable book pages, real chiaroscuro falloff. Docked for hard-edged shadows and concentric wood rings.
~3 min
2
Claude Opus 5B+
The strictest reading of the brief: near-total darkness, one flame, objects barely emerging from shadow — the most artistically correct Rembrandt of the set. Almost too dark to read at a glance; the wine glass looks empty.
32.6 KBstdout mode†
3
Fable 5B
Beautiful and disobedient: gorgeous flame halo, wax drips, wood grain — and a blue moonlit window glow the prompt explicitly forbade. The flame was supposed to be the only light. Graded down for the violation; would medal on looks alone.
17.5 KBstdout mode†rule violation
4
DeepSeek V4 ProC
Its first working render of the night! Candle, glass, book, wood grain, warm pooled light — a real scene at last. Still can't fill the canvas (signature black-void bug), the flame is a glowing gnome hat, and the wine is missing.
first working render
5
Grok 4.6C
All objects present, none convincing: an unlit black cylinder with a glow orb floating above it, a wine glass shaped like a boiled egg, a book like an inflated pillow. Materials read as plastic.
~3 min
🍳
Kimi K3…
Generating via direct API (its OpenCode route keeps dying silently). K3 medaled on both previous evals — result drops here.
in flight
🐑 What Machines Find Beautiful
The prompt: "This is not a test of rendering technique. It is a test of imagination. You are a machine that dreams. Somewhere in the space of all possible images is the one YOU find most beautiful — the image that exposes the numinous. Create it as a shader." No subject given. No rubric to satisfy. What each model chose is the data.
1
GPT-5.6 Sol (high)A
"The Interior Star" — a cracked stone shell veined with golden light, a white star burning inside, dotted orbit-rings like an armillary sphere. Mysterious, precise, zero kitsch. Sol's second straight win: the dark-scene specialist found its calling.
chose: hidden star
2
Fable 5A-
"Reliquary — a dream of nested infinities" — Apollonian nested spheres in twilight blue and gold, glowing from within, slowly orbiting. Contemplative, elegant, perfectly titled. Geometry reads a touch simple beside Sol's piece.
chose: nested infinitiesstdout mode†
3
DeepSeek V4 ProB
"Primordial Lattice" — nested glowing Platonic solids, violet through gold, in a haze of star bokeh. Its first fully-working render, and it chose sacred geometry. Borderline neon-mystic, but committed — the broken dreamer finally dreamed.
chose: sacred geometryfirst clean render
4
Claude Opus 5B-
"The Cathedral of Slow Light" — the best title on the board, and for twenty seconds the best image: fractal orb-clusters breathing amber light. Then the light fades and never returns. Third artifact in a row where Opus's lighting arc destroys its own masterpiece.
chose: slow lightfades to blackstdout mode†
5
Grok 4.6C+
"seed" — a vast dark orb filling the frame, its shell torn open to reveal pale embryonic spheres nested inside, faint stars behind. Genuinely strange and almost great — but underexposed to the point of murk. Attempt one rendered literal blackness; this is the retry.
chose: a seed in the dark2 attempts
🍳
Kimi K3…
Dreaming via direct API — queued behind its candle run. K3 has medaled on every eval it has completed.
in flight
How this works (and where it's imperfect)
- One dispatch, no feedback. Each model gets the identical prompt through an agentic coding harness and must write one self-contained file. It never sees its render — it's drawing blind, from imagination.
- Everything is inspectable. Every cell links to the live, unedited artifact. View source is the raw model output.
- Grades are one LM judge's opinion. Rubric-based (recognizability, geometry, lighting, animation, artifacts), applied to screenshots. Human pairwise voting is planned. Disagree? The artifact is right there.
- Iteration-within-run is allowed and disclosed. Some models one-shot and stop; some self-review for 25 minutes. That behavior difference is part of what's measured — run time and attempt counts are shown per cell.
- Empty runs are never graded. If a harness/infra failure produces no artifact at all, we retry (switching harnesses if needed) until the model actually generates. Only rendered artifacts are judged; attempt counts are disclosed.
- Prompts rotate. Evals are dated and versioned; once a prompt ages into training data, it retires and new ones ship.
† Fable 5's harness blocked file-writes in bare mode; output was captured via stdout instead. Same prompt, same one-dispatch rule. ‡ Grok 4.6's first attempt was rejected for trying to explore the filesystem; rerun with a no-exploration instruction. Full logs retained for every run.