dsh-v4flash-tiler

DeepSeek vision models quietly resize any image exceeding roughly 800×800 equivalent pixels, capping each image at about 384 tokens. Dense UI copy, axis labels on charts, and small icons all melt into color blobs after that shrink. dsh-v4flash-tiler solves exactly this: it automatically detects any image with a side larger than 1024px in chat, slices it into up to 9 high-fidelity tiles with 15% overlap, and annotates each tile with grid coordinates and overlap notes so the model can stitch them back into the original layout. Small images pass through untouched, and any failure gracefully falls back to the original image with an explanatory note.

The plugin offers three entry points: automatic tiling triggered by simply sending an image in chat, a CLI tool bundled under engine/, and a Python API for scripted workflows. When a message contains multiple images, each image's tiles form an isolated group and cross-image mixing is explicitly forbidden. Total tile count per message is also guarded by DSH attachment limits, with oversized images reverting to pass-through when the ceiling is hit.

If you regularly feed code screenshots, dense spreadsheets, academic figures, or any large image packed with fine text and small icons into DeepSeek vision models, this plugin lets you stop fighting the auto-resize and actually get the detail you need.