O conteúdo está em inglês; a navegação e os filtros mantêm o idioma selecionado.

AI Models Take on a Lime-Slice Pose Drawing Challenge

Ann Nguyen asked Gemini 3.7 Flash, Kimi K3, Claude Opus 5, and GPT 5.6 Sol to create a human pose integrated with a lime slice. The social-media comparison highlights differences in creative interpretation, but it is not a standardized benchmark.

Publicado em 08/09/2026, 03:16:43Última atualização 08/09/2026, 03:16:43Publicação editorial

The experiment is an informal visual prompt test rather than a controlled evaluation. Each model was given the same basic creative challenge: combine a human pose with a small lime slice and present the result as an integrated drawing.

Nguyen has shared related prompts involving a die, a sky photograph, and animal silhouettes. A separate discussion also references a banana-based comparison involving other models. These examples suggest that the models produce noticeably different compositions and creative directions when responding to object-integration tasks.

The available evidence comes primarily from social posts and reposts. There is no published methodology covering repeated trials, consistent scoring, prompt controls, or cost normalization, so the results should be treated as an entertaining showcase rather than a general ranking of image-generation quality.

Evidências das fontes 6

A verificação de fontes não garante exatidão permanente. Considere a fonte original, a data e as correções posteriores.

Fable 5 vs Grok 4.6 vs Sol vs Qwen3.8-Max (Aug 2026) | explainx.ai Blog | explainx.aiexplainx.ai · Complementar

The banana drawing test (Ann Nguyen, @ann\_nnng): gave Grok 4.6, Claude Opus 5, GPT-5.6 Sol, and Qwen3.8-Max a banana and asked each to be creative with an object-integrated drawing. Grok 4.6's standout answer treated the banana as a tiny island. A commenter's ranked preference on the thread: Opus 5 > Grok 4.6 > Qwen3.8-Max > GPT-5.6 Sol — one person's subjective call on one prompt, not a scored benchmark. Note the model set: this test used Claude Opus 5, not Fable 5. For an actual scored, cost-normalized drawing benchmark that does include Fable 5, see explainx.ai's coverage of the TryAI Canvas Arena, where GPT-5.6 Sol beat Fable 5 on drawing quality for roughly a twentieth of the cost. [...] ## Related on explainx.ai Sources: Product Compass Bug Hunt Bench v10, published August 12, 2026 (Paweł Huryn, @PawelHuryn) · RuneScape Bench, best-of-1 results · Arena.ai Code Arena WebDev leaderboard post, August 12, 2026 · Ann Nguyen (@ann\_nnng) banana drawing thread · Shirish (@shiri\_shh)