A user reports frustration with Gemma 4's default image resolution settings, noting that the model struggles to decipher smaller text and larger compositional elements compared to competitors like Qwen 3.6.

  • The user attempted to adjust parameters in LlamaCpp (`--image-min-tokens 560 --image-max-tokens 2240`) to improve performance.
  • Applying these specific token limits caused the server to crash and quit rather than improving vision capabilities.
  • The user seeks a method to increase image resolution for Gemma 12b to function as a comprehensive assistant.