[GH-ISSUE #1286] Feedback sobre leitura de imagem — DeepSeek Chat #837

New issue

Closed

opened 2026-10-26 21:30:48 +00:00 by bsakaguchi · 3 comments

bsakaguchi commented

2026-10-26 21:30:48 +00:00

Owner

Originally created by @QiCriativo on GitHub (May 8, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1286

Descrição do problema:

Durante uma sessão de desenvolvimento, enviei um print de uma tela mostrando avatares de membros com fallback quebrado (círculo não arredondado, centralização incorreta). A IA analisou a imagem e:

Não identificou visualmente que os fallbacks estavam quebrados (quadrados em vez de circulares)

Sugeriu correções no componente errado (MemberCard) por várias iterações

Só resolveu quando eu insisti que o problema era visual e descrevi textualmente

Comportamento esperado: A IA deveria identificar problemas visuais em screenshots (alinhamento, forma, cores) com a mesma precisão que analisa código.

Sugestão: Melhorar o modelo de visão para identificar inconsistências de UI (border-radius ignorado, centralização incorreta, tamanhos inconsistentes entre elementos irmãos).

Ambiente: DeepSeek Chat, maio 2026.

Originally created by @QiCriativo on GitHub (May 8, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/1286 Descrição do problema: Durante uma sessão de desenvolvimento, enviei um print de uma tela mostrando avatares de membros com fallback quebrado (círculo não arredondado, centralização incorreta). A IA analisou a imagem e: Não identificou visualmente que os fallbacks estavam quebrados (quadrados em vez de circulares) Sugeriu correções no componente errado (MemberCard) por várias iterações Só resolveu quando eu insisti que o problema era visual e descrevi textualmente Comportamento esperado: A IA deveria identificar problemas visuais em screenshots (alinhamento, forma, cores) com a mesma precisão que analisa código. Sugestão: Melhorar o modelo de visão para identificar inconsistências de UI (border-radius ignorado, centralização incorreta, tamanhos inconsistentes entre elementos irmãos). Ambiente: DeepSeek Chat, maio 2026.

bsakaguchi

2026-10-26 21:30:48 +00:00

closed this issue
added the
stale

closed-as-stale
labels

bsakaguchi commented

2026-10-26 21:30:49 +00:00

Author

Owner

@qingkong66 commented on GitHub (May 8, 2026):

@QiCriativo Thank you for the detailed report.

You pointed out a real limitation: the vision model can read text in screenshots, but it does not reliably detect UI rendering issues — like a fallback avatar appearing as a square instead of a circle, or incorrect centering. It also struggled to map the visual problem to the correct component.

Workaround for now:
For UI/style issues, describe the problem in text as clearly as you can, even if it feels repetitive. For example: "The fallback avatar is a square, but it should be a circle. The image is centered incorrectly." This gives the model the signal that the image alone does not provide.

For code or error logs, copy-paste the text instead of using a screenshot — that avoids the vision model entirely and gives much cleaner results.

We've noted this as a quality observation under "vision / UI understanding". It won't go into the regular feedback pool for now, but it's kept on file.

Thanks again for bringing this up.

@qingkong66 commented on GitHub (May 8, 2026): @QiCriativo Thank you for the detailed report. You pointed out a real limitation: the vision model can read text in screenshots, but it does not reliably detect UI rendering issues — like a fallback avatar appearing as a square instead of a circle, or incorrect centering. It also struggled to map the visual problem to the correct component. Workaround for now: For UI/style issues, describe the problem in text as clearly as you can, even if it feels repetitive. For example: "The fallback avatar is a square, but it should be a circle. The image is centered incorrectly." This gives the model the signal that the image alone does not provide. For code or error logs, copy-paste the text instead of using a screenshot — that avoids the vision model entirely and gives much cleaner results. We've noted this as a quality observation under "vision / UI understanding". It won't go into the regular feedback pool for now, but it's kept on file. Thanks again for bringing this up.

bsakaguchi commented

2026-10-26 21:30:50 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Jun 16, 2026):

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

@github-actions[bot] commented on GitHub (Jun 16, 2026): This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you believe this issue is still relevant, please leave a comment to keep it open. Thank you for your contributions!

bsakaguchi commented

2026-10-26 21:30:50 +00:00

Author

Owner

@github-actions[bot] commented on GitHub (Jul 14, 2026):

false

@github-actions[bot] commented on GitHub (Jul 14, 2026): false