How do LLM nodes support multimodal input?

Discussion
1

To use doubao-seed-1-6 or qwen3-omni-flash, you need to pass image_url or video_url. How can this be implemented using an LLM node?

1 Like
2

For handling image_url with LLM integration, my implementation approach is as follows (welcome more sharing and discussion):

  1. Add a “HTTP Request” node
    Use a GET request to image_url to obtain the output result, which is a list of files.

  2. Add a “List Operation” node
    Input the file list from the previous step to obtain the first_record as the output file.

  3. Add an “LLM” node
    At this point, you can reference the file from the previous node.

Note: The HTTP request in step 1 may return images in GIF format (even if the file extension is jpg/png/jpeg, etc.), which could cause the LLM to fail to parse. Therefore, an if condition should be added beforehand to prevent LLM errors.

3

image

5

It’s a bit troublesome. I’m currently using an HTTP node instead of LLM. There’s a reply downstairs mentioning that the LLM node now supports document input.

6

Which version are you using? I’m using 1.10.1 and there’s no document input option.

7

The self-referential visualization logic has been modified, and corresponding configurations have been added for the document model.

8

There are multiple ways to implement this. My approach mainly involves parsing the file first via HTTP requests, then using an LLM-VL model combined with prompts to extract content from the image. Friends below, you can try directly extracting using a model that supports this image. Both methods are worth trying, and you can evaluate them based on model costs.