Llama 4 Vision: How Dynamic Tiling Works

Per Meta's official Llama 4 documentation, multimodal input supports text plus up to 5 images, and image understanding is English-only. Larger images are divided into 336 x 336 pixel local tiles using <|tile_x_separator|> and <|tile_y_separator|> tokens, followed by <|image|> and a downsized 336 x 336 global image patch inside <|image_start|>...<|image_end|>.

<|begin_of_text|><|header_start|>user<|header_end|> <|image_start|><|patch|>...<|tile_x_separator|>...<|tile_y_separator|>...<|image|><|patch|>...<|image_end|>Describe the layout hierarchy and typography issues in this dashboard mockup in English.<|eot|> <|header_start|>assistant<|header_end|>