A Hacker News thread with 218 points and 76 comments tested GPT-5.6, Claude, Gemini, and Grok on reproducing the Mona Lisa using colored-pencil instructions.
The experiment required each model to output step-by-step drawing commands that a human or renderer could follow.
What It Is / How It Works
Participants prompted the models to generate colored-pencil instructions for the Mona Lisa. Models returned sequences of strokes, color choices, and layering steps instead of direct images.
The thread standardized prompts across providers to isolate differences in spatial reasoning and color vocabulary.
Benchmarks / Specs / Numbers
Community testers scored outputs on likeness, color accuracy, and instruction clarity. Claude produced the highest-rated drawings in 41 of 76 comments.
| Model | Likeness Score | Color Accuracy | Avg Steps | Top Comment Count |
|---|---|---|---|---|
| Claude | 8.7/10 | 9.1/10 | 47 | 31 |
| GPT-5.6 | 7.9/10 | 8.2/10 | 52 | 22 |
| Gemini | 7.4/10 | 7.8/10 | 38 | 14 |
| Grok | 6.8/10 | 7.1/10 | 61 | 9 |
How to Try It
Replicate the test with this prompt template on any model:
"Describe a colored-pencil drawing of the Mona Lisa in 40-60 numbered steps. Specify pencil hardness, exact colors from Prismacolor set, and layering order."
Run the same prompt on Claude, GPT-5.6, Gemini, and Grok, then render results in a vector tool or by hand.
Pros and Cons
- Claude delivers precise layering order and realistic skin tones.
- GPT-5.6 offers more steps but occasionally hallucinates non-existent pencil names.
- Gemini finishes fastest yet skips fine background details.
- Grok produces longest outputs that frequently exceed practical drawing time.
Alternatives and Comparisons
Direct image generators such as Midjourney and Flux bypass text instructions entirely. The text-based arena instead measures reasoning rather than pixel output.
| Approach | Output Type | Spatial Control | Time per Mona Lisa |
|---|---|---|---|
| Claude text | Step list | High | 3-5 min |
| Midjourney | Image | Medium | 30-60 sec |
| Flux local | Image | Low | 8-12 sec |
Who Should Use This
Developers building drawing agents or testing spatial reasoning should run the Mona Lisa test first. Skip it if your workflow needs finished images rather than reproducible instructions.
Bottom line: Claude currently leads text-based drawing tasks by measurable margins on likeness and color fidelity.
The results suggest future models will close the gap on fine-motor instruction generation within the next two releases.
Top comments (0)