Jimeng Visual Understanding

The whole pack is about making images and videos. This node is the odd one out: it reads them. Jimeng Visual Understanding hands an image or video to ByteDance's Seed vision models and returns a text answer - describe this scene, identify this object, summarize what happens in this clip, check whether this render matches its prompt. Think of it as the quality-control node, or the bridge that lets a text-model workflow actually see your generated output.

It's the same deal as the rest of the pack - paid Volcano Ark API key, ByteDance servers, no local model - but the payment model is different in spirit: this is cheap, fast token-based inference rather than a pricey image/video render, so you can use it freely as a QC pass without watching your credits drain.

How it works

You feed up to three inputs into visual_input_1 through visual_input_3 (each accepts an image or a video), type a question, and the node uploads the file to the API and returns the model's answer. Images are cached locally and uploaded, videos too; the file_expire_seconds control (default 604800, i.e. seven days, min 24 hours) governs how long that uploaded file is valid for the API to reference. The model is doubao-seed-2-1-pro by default, with the older 2.0 variants (pro, lite, mini) still selectable for existing workflows.

The inputs that matter

• user_prompt - your question. The default is Chinese ("describe this image or video's content"), so expect to replace it unless you want Chinese answers.
• system_prompt - the model's persona; defaults to "you are Doubao, an AI developed by ByteDance." Fine to leave.
• model - the vision model, doubao-seed-2-1-pro by default.
• reasoning_mode / reasoning_effort - the "deep thinking" controls. Set reasoning_mode to enabled for hard multi-step questions (like whether a generated frame matches a prompt), or leave auto and let the model decide.
• detail - low or high; high costs more tokens.
• fps - sample rate for videos, 0.2–5.
• turns - 1 to 10. Turn this up for multi-turn conversation in one run; above 1, the model keeps the prior turns as context.
• stream - off by default; streaming output for long answers.
• seed - reproducibility.

Outputs

• content - the model's answer as a string. Wire this into a text display or another node.
• raw_json - the full API response, for when you need metadata beyond the answer.

Installing and using it

Install is the pack-standard route: ComfyUI Manager (search "ComfyUI Jimeng API") or git clone https://github.com/fkxianzhou/ComfyUI-Jimeng-API into custom_nodes, restart. The usual caveats apply - legacy UI only, content moderation on what you upload.

The genuinely useful pattern: run a generation, pipe the result into this node, and ask it to verify the output against the prompt you used. It's marked experimental in the pack and the multi-turn feature is the one that can surprise you (context accumulates across turns), but for a one-shot "what's actually in this frame" check it's hard to beat - especially since it's cheap enough to run every time.