How to combine different files to a Array[file] and send to LLM

[WilliamChen-luckbob]

Self Checks

• I have searched for existing issues search for existing issues, including closed ones.
• I confirm that I am using English to submit this report (我已阅读并同意 Language Policy).
• [FOR CHINESE USERS] 请务必使用英文提交 Issue，否则会被关闭。谢谢！:)
• Please do not modify this template :) and fill in all the required fields.

1. Is this request related to a challenge you're experiencing? Tell me about your story.

Here's a scenario:
I have a template image serving as the base image. I want the user inputs to be any image and a prompt, with the goal of replacing and filling a specified position in the template image with a creative pattern.

Therefore, in the workflow, I intend to set the base image as content loaded by default (without requiring user input). During the workflow processing, I need to combine the user-provided image with the base image into a single image list, which will be passed to the LLM as visual input.

However, I cannot find any solution that allows the LLM to use the combined image as visual input—it seems the LLM can only reference user inputs directly.

2. Additional context or comments

you can see, only user inputs variables available here, python file results are ignored, I have no idea at all...

[dosubot[bot]]

Dify can technically send multiple images (Array[File]) to an LLM node if you aggregate them into a single array variable and wire that variable into the node’s vision config. However, the workflow UI only allows you to select user input variables (like sys.files or userinput.files) for visual input, and files generated by code nodes aren’t automatically available for selection unless you explicitly add them as File objects in the variable pool with the correct type and selector. This is a known limitation—only files uploaded via the UI are reliably passed to the LLM node as visual input, not files produced or combined in intermediate code steps source.

If you need to combine a base image (template) and a user-uploaded image, the recommended pattern is to upload both images via the UI or API, aggregate them into an array (Array[File]), and reference that array in the LLM node’s vision config. If you want to inject files from a code node, you must upload them using the /v1/files/upload endpoint, construct proper File objects, and add them to the variable pool so the LLM node can fetch them source.

Workarounds:

• Upload the base image as a user input or via API at workflow start, so it appears in sys.files or userinput.files.
• Use a code node to aggregate both images into an array, but ensure the output is of type Array[File] and is registered in the variable pool.
• Reference sys.files directly in the LLM node’s vision config if code nodes aren’t available or type compatibility is an issue source.
• If you need to pass images between nodes, always use the full File object structure, not just file IDs or base64 source.

Currently, there’s no built-in way to pass files generated by Python code nodes as visual input to the LLM node in the UI. You may want to monitor future updates or contribute to feature requests for more flexible file variable handling source.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}