Skip to content

readme_captions.md

Latest commit

 

History

History
41 lines (29 loc) · 2.52 KB

readme_captions.md

File metadata and controls

41 lines (29 loc) · 2.52 KB

Captions tab

The Captions tab is used to generate captions for images and videos.

Screenshot Captions Tab

General Use

  • First select a folder with images or videos to process at the top of the window.

  • To get started make sure to select a model from the dropdown menu and hit "Load Model". Models can be downloaded using the built-in downloads manager on the Captions tab (📥💾 button), more info in the readme_models.md file.

  • Select a "System Prompt" from the dropdown menu or use the "Custom" option to enter your own system prompt.

  • Set any other options as desired (they all have tooltips to explain what they do).

  • Hit "TEST FIRST IMAGE" or "TEST RANDOM IMAGE" to test on a single image.

  • Hit "START PROCESSING" to process all images in the selected folder.

In your output folder you will find a .txt file for each image/video with the generated caption.

Optimizing Speed

  • Although a single image can be a bit slow (ranging from a few seconds to tens of seconds)
  • Increasing the "Batch Size" will allow multiple images to be processed at once, experiment by raising this value to as high as your VRAM allows
  • Lowering precision and/or selecting a different attention implementation and/or torch compile may increase speed (only supported for non-GGUF models)
  • Some GGUF models may be faster as well, but require manual downloading and installation of the llama-cpp-python package, see above for instructions. This is for more advanced users.

Customizing System Prompt

  • System prompt files can be found as text files in the "prompts" folder
  • The app checks these at startup
  • You can select a predefined prompt from the list
  • Or you can use the "Custom" setting and type your system prompt in the app
  • Or create/edit the text files in the "prompts" folder

Model Families

VisionCaptioner supports two Vision-Language Model families:

  • Qwen-VL (Qwen2.5-VL, Qwen3-VL) — controlled by the "Max Resolution" dropdown.
  • Google Gemma 4 (E2B, E4B, 26B-A4B, 31B) — controlled by the "Vision Tokens (Gemma)" dropdown, which sets the soft visual token budget per image (70, 140, 280, 560, or 1120). Higher = more detail, slower, more VRAM. The "Max Resolution" setting is ignored for Gemma.

The correct settings panel becomes active automatically when you pick a model from the dropdown. Gemma 4's built-in "thinking" reasoning mode is automatically disabled for captioning.

See readme_models.md for the full list of supported models and download links.