AI on the Lot 2026 - QwenVL - i2t
The third workflow released for AI on the Lot 2026, made available for ComfyUI 101 workshop attendees to open and follow along if they choose. A simple image-to-text prompt creator using Qwen VL, useful for generating prompts from reference images.
About
This workflow turns any image into descriptive, instruction-following text using Qwen-VL inside ComfyUI. It’s a simple, reliable chain: LoadImage brings your source frame or photo into the graph, AILab_QwenVL runs the Qwen-VL vision-language model to generate captions or answers based on your instruction, and PreviewAny displays the resulting text directly in the UI. Because AILab_QwenVL accepts both an image and a prompt, you can steer the output toward concise alt text, detailed scene breakdowns, tags, or even OCR-style transcriptions.
What you provide
- Reading the workflow
Runs in ComfyUI. This workflow can't run on this page. Open it in Comfy Cloud with everything loaded, or download it for your own ComfyUI.
Try in CloudQuestions & answers
AILab_QwenVL runs a Qwen-VL vision-language model on the input from LoadImage. It takes your instruction (prompt) plus the image, then generates text—such as a caption, list of objects, or an answer to a question about the image—which PreviewAny displays.
It depends on how AILab_QwenVL is configured in your environment. Many setups load local model weights and run fully offline. If your node is configured to call a remote backend, you may need network access or credentials. Check your node’s model selection and documentation.
Guide the model with explicit instructions (e.g., "One-sentence caption," "Return keywords separated by commas," or "Output JSON with fields: setting, subjects, actions"). If your node exposes settings like maximum tokens or temperature, use them to limit length and make results more deterministic.
Use clear, specific prompts (e.g., "Transcribe all legible text exactly, line by line"), provide sufficiently large, sharp images, and reduce temperature for factual tasks. If the scene is busy, crop to the relevant area before running LoadImage so the model focuses on what matters.