Visual Scene

Juggling Clown Waterfall Scene

Four model runs tackled the same clown-juggling prompt, with the spread driven by silhouette quality, juggling rhythm, and how convincingly the three-ball waterfall loop holds together.

AnimationHTMLCharacter designMotion loopVisual review

Juggling clown four-way comparison poster

Watch episode

Published video

Juggling clown four-way comparison

The full four-way comparison and each model clip are now uploaded to YouTube. Portrait short exports are stored under videos/out/juggling-clowns-001 and will be backfilled after short-form publishing.

22s

Watch full episode Watch grid short Watch comparison short Watch best-run short Watch worst-run short

Result matrix

One task, 4 recorded outcomes

The table shows visible results first. Detailed logs and clips can be attached once real runs are recorded.

Agent	Model	Status	Score	Breakdown	Time	Cost	Failure
Claude Code	GLM	Strong	A-79	Visual authenticity (VQ)29/40 Physics and consistency (PC)20/25 Semantic restoration (SR)17/20 Composition and readability (CR)13/15	22s clip reviewed	unknown	The result is clearly stylized rather than physically grounded, and the waterfall motion still feels like a polished graphic loop more than a natural circus performance.
Claude Code	Qwen	Solid	B73	Visual authenticity (VQ)26/40 Physics and consistency (PC)19/25 Semantic restoration (SR)16/20 Composition and readability (CR)12/15	22s clip reviewed	unknown	The figure proportions feel rigid, the ball choreography is less polished than GLM, and the overall finish lands closer to a neat demo than a fully resolved character animation.
Claude Code	DeepSeek	Partial	B-68	Visual authenticity (VQ)24/40 Physics and consistency (PC)18/25 Semantic restoration (SR)14/20 Composition and readability (CR)12/15	22s clip reviewed	unknown	The juggling relation is less explicit than the top two runs, and the two-hand waterfall pattern reads more weakly because the motion emphasis sits on presentation rather than on clear arc choreography.
Claude Code	Doubao	Weak	C55	Visual authenticity (VQ)18/40 Physics and consistency (PC)12/25 Semantic restoration (SR)13/20 Composition and readability (CR)12/15	22s clip reviewed	unknown	The simplest and least refined result in the set: the body feels toy-like, the juggling paths are underdeveloped, and the animation does not convincingly sell a continuous waterfall loop.

Claude Code

GLM

A-

Status: Strong
Score: 79
Breakdown: Visual authenticity (VQ)29/40
Physics and consistency (PC)20/25
Semantic restoration (SR)17/20
Composition and readability (CR)13/15
Time: 22s clip reviewed
Failure: The result is clearly stylized rather than physically grounded, and the waterfall motion still feels like a polished graphic loop more than a natural circus performance.

Claude Code

Qwen

Status: Solid
Score: 73
Breakdown: Visual authenticity (VQ)26/40
Physics and consistency (PC)19/25
Semantic restoration (SR)16/20
Composition and readability (CR)12/15
Time: 22s clip reviewed
Failure: The figure proportions feel rigid, the ball choreography is less polished than GLM, and the overall finish lands closer to a neat demo than a fully resolved character animation.

Claude Code

DeepSeek

B-

Status: Partial
Score: 68
Breakdown: Visual authenticity (VQ)24/40
Physics and consistency (PC)18/25
Semantic restoration (SR)14/20
Composition and readability (CR)12/15
Time: 22s clip reviewed
Failure: The juggling relation is less explicit than the top two runs, and the two-hand waterfall pattern reads more weakly because the motion emphasis sits on presentation rather than on clear arc choreography.

Claude Code

Doubao

Status: Weak
Score: 55
Breakdown: Visual authenticity (VQ)18/40
Physics and consistency (PC)12/25
Semantic restoration (SR)13/20
Composition and readability (CR)12/15
Time: 22s clip reviewed
Failure: The simplest and least refined result in the set: the body feels toy-like, the juggling paths are underdeveloped, and the animation does not convincingly sell a continuous waterfall loop.

What this tests

The hidden engineering inside a simple game

Character styling

Preserves the prompt-critical clown design instead of drifting into a generic figure.

Motion choreography

Coordinates three-ball timing, hand relation, and a loop that still reads as juggling.

Prompt fidelity

Keeps the center placement, color assignments, and the required two-hand waterfall motion intact.

Temporal stability

Maintains coherent ball arcs and loop timing across the reviewed clip.

Composition readability

Keeps the clown dominant in frame so the action remains easy to parse at a glance.

Unified prompt

Create a single file containing an HTML/CSS/JavaScript animation. The main character is a stylized "clown" (purple suit, green hair, white face) standing at the center. He should continuously juggle three colored balls (red, green, and blue) with both hands, forming a waterfall-like continuous motion.

Acceptance criteria

A single HTML page renders without fatal runtime errors.
A centered clown with a purple suit, green hair, and a white face is clearly visible.
Three distinct balls remain visible and color-correct: red, green, and blue.
The clown uses both hands to maintain a readable waterfall-like juggling loop.
The animation remains visually stable long enough to support screenshot and clip review.

Scoring

Readable rubric for video and reproduction

Visual authenticity (VQ)

Judges clown silhouette quality, facial readability, costume finish, ball rendering, and overall visual polish.

Physics and consistency (PC)

Judges juggling rhythm, hand-ball relation, and whether the loop remains temporally stable without broken arcs or teleports.

Semantic restoration (SR)

Judges whether the result actually matches the prompt: centered clown, purple suit, green hair, white face, three colored balls, and a two-hand waterfall motion.

Composition and readability (CR)

Judges subject clarity, information hierarchy, and whether the juggling action is understandable instantly.

Run details

What each agent got right and wrong

GLM landscape clipWatch clip

Screenshots

Captured run

Claude Code

GLM

A-79

Visual authenticity (VQ)29/40

Physics and consistency (PC)20/25

Semantic restoration (SR)17/20

Composition and readability (CR)13/15

Status: Strong
Strength: Best overall balance of clown readability, clear ball separation, and a stable loop that most convincingly reads as intentional two-hand juggling.
Failure: The result is clearly stylized rather than physically grounded, and the waterfall motion still feels like a polished graphic loop more than a natural circus performance.

Qwen landscape clipWatch clip

Screenshots

Captured run

Claude Code

Qwen

B73

Visual authenticity (VQ)26/40

Physics and consistency (PC)19/25

Semantic restoration (SR)16/20

Composition and readability (CR)12/15

Status: Solid
Strength: Strong center framing and a readable star-lit stage make the action easy to parse, with all three balls staying visible and color-correct.
Failure: The figure proportions feel rigid, the ball choreography is less polished than GLM, and the overall finish lands closer to a neat demo than a fully resolved character animation.

DeepSeek landscape clipWatch clip

Screenshots

Captured run

Claude Code

DeepSeek

B-68

Visual authenticity (VQ)24/40

Physics and consistency (PC)18/25

Semantic restoration (SR)14/20

Composition and readability (CR)12/15

Status: Partial
Strength: Most poster-like composition in the set, with a strong vertical silhouette and a controlled stage frame that stays visually clean.
Failure: The juggling relation is less explicit than the top two runs, and the two-hand waterfall pattern reads more weakly because the motion emphasis sits on presentation rather than on clear arc choreography.

Doubao landscape clipWatch clip

Screenshots

Captured run

Claude Code

Doubao

C55

Visual authenticity (VQ)18/40

Physics and consistency (PC)12/25

Semantic restoration (SR)13/20

Composition and readability (CR)12/15

Status: Weak
Strength: The broad color blocks make the clown immediately identifiable, and the centered framing is simple enough that the viewer never loses the subject.
Failure: The simplest and least refined result in the set: the body feels toy-like, the juggling paths are underdeveloped, and the animation does not convincingly sell a continuous waterfall loop.

Postmortem

Why A Simple Juggling Scene Still Works As A Visual Arena Case

Motion quality becomes obvious fast

A three-ball juggling loop is compact enough to explain in one sentence, but weak timing, broken arcs, or unclear hand-ball relations stand out immediately.

Prompt fidelity is easy to audit

This case has concrete visual anchors: centered clown, purple suit, green hair, white face, and three colored balls. A run either preserves those anchors cleanly or it drifts.

Video matters even for simple loops

The screenshot tells you about silhouette and composition, but the clip reveals whether the balls actually maintain a stable waterfall rhythm over time.

Case package

starter/
tests/
prompt.md
task.md
scoring.md
runs.json
screenshots/
clips/
logs/

Manifest Prompt Scoring Runs JSON

Commands

npm run capture:screenshots -- <html-files...> --case juggling-clowns --output-dir images
npm run capture:videos -- <html-files...> --case juggling-clowns --output-dir videos
npm run validate

Use this draft package to keep the prompt, screenshots, local clips, and first-pass scoring aligned before external video publishing and short generation are finished.

Juggling Clown Waterfall Scene

Juggling clown four-way comparison

One task, 4 recorded outcomes

Claude Code

Claude Code

Claude Code

Claude Code

The hidden engineering inside a simple game

Character styling

Motion choreography

Prompt fidelity

Temporal stability

Composition readability

Readable rubric for video and reproduction

Visual authenticity (VQ)

Physics and consistency (PC)

Semantic restoration (SR)

Composition and readability (CR)

What each agent got right and wrong

Claude Code

Claude Code

Claude Code

Claude Code

Why A Simple Juggling Scene Still Works As A Visual Arena Case

Motion quality becomes obvious fast

Prompt fidelity is easy to audit

Video matters even for simple loops

Ready for the next visual task?