Dify Proposes Workflow Agent Image-Passing Fix to Add Multimodal Input for Vision Models

On November 6, the Dify community opened PR #45745 and added tests on November 7 to fix the workflow agent’s image input pipeline. When a user uploads an image and the selected model supports vision, the agent may still receive only a text reference to the file. The proposal converts the image content and sends it to the model through the multimodal user_prompt layer registered in the runtime. At the time of review, the PR remained open and had not yet become an official release fix. Proposed fix

The original issue was reported on a self-hosted deployment based on version 1.18.0. The reporter said that agent_v2 serialized sys.files into the workflow prompt text without creating multimodal image content; the shared client’s workflow input structure also lacked user_files, and the plain-text prompt layer in use could not carry images. Therefore, adding image conversion upstream alone would not be enough to get images to the model; the backend request structure and prompt layer also had to be connected. Issue and root-cause analysis

This distinction directly affects how agents handle screenshots and charts. The official documentation says the newer Agent node has its own capabilities and sandbox, enabling it to complete tasks autonomously within a workflow. But making a file available to the sandbox and sending an image directly to a vision model are different input paths. According to the issue report, the agent had resorted to downloading images and installing OCR tools to recognize them. The engineering implication is that when diagnosing “the model can’t see the image,” teams should inspect the content sent to the model, rather than only confirming that the upload succeeded or that vision is enabled. Agent node documentation, user report

The author said a regression test for mixed image and PDF uploads failed under the old implementation because images were not converted, and passed after the changes were applied. A runtime test also checks BinaryContent and verifies session snapshot transitions from text to image and back to text. However, some tests stubbed file conversion and model capability checks, and end-to-end execution with real file storage and model providers has not yet been completed. Teams considering adoption should track the PR’s merge and release inclusion, then verify that images are passed correctly for mixed attachments and across multiple turns. Validation scope