Jimeng Reference to Video

This node does one thing simply: you give it one to four reference images and a prompt, it animates them into a video. No first-frame/last-frame choreography, no multimodal reference arrays - just "here are pictures, make them move." For that job it's the easiest node in the pack to understand, and for that same reason it's also the one to be careful about adopting today.

The reason to hesitate isn't the node's design, it's the model underneath. This node is hardwired to doubao-seedance-1-0-lite-i2v, and the README is upfront that the lite image-to-video model is in the retirement pipeline. The pack marks this node deprecated and points you at Seedance 2.0, whose reference-image inputs cover the same ground with a much better model. So: fine for a one-off, but don't build a long-term workflow on it.

Like everything here, it's cloud-only - a paid Volcano Ark API key is the only requirement, and ByteDance does the generating.

How it works

Mechanically it's the simplest video node in the pack: it base64-encodes whatever reference images you feed in, attaches them to the request as "reference_image" content, and submits to the fixed lite model with your prompt, duration, and resolution. It runs the same polling-and-download loop as the rest of the family. Give it no images at all and it refuses immediately - at least one reference is required, which is the point of the node.

The inputs that matter

• ref_image_1 through ref_image_4 - up to four reference images. More references generally mean more coherent character/object identity across the clip, but each one adds request size, and there's a total-size ceiling the node enforces.
• prompt - describes the motion you want (the content of the images is already in the request).
• duration - a float from 1.2 to 12 seconds.
• resolution / aspect_ratio - capped at 720p here, with aspect ratio defaulting to 16:9 (no "adaptive" option, unlike the other video nodes).
• enable_offline_inference - the cheaper "flex" service tier.
• enable_random_seed / seed, plus the standard generation_count, filename_prefix, save_last_frame_batch, and non_blocking runtime controls.

Outputs

video, last_frame (the final frame as an image), and response (task JSON) - the standard trio for this pack's video nodes.

Installing and the gotcha you'll actually hit

Install is the pack-standard path: ComfyUI Manager or

cd ComfyUI/custom_nodes
git clone https://github.com/fkxianzhou/ComfyUI-Jimeng-API

then restart. The volcengine-python-sdk[ark] dependency self-installs. The usual pack quirks apply (legacy UI only, ByteDance content moderation on prompts). The bigger, pack-specific gotcha is the deprecation: the lite model's quota is being wound down as the platform retires it, so expect errors to start appearing and treat this as a stopgap. If you need reference-image-to-video reliably, skip ahead to Seedance 2.0 - it does everything this node does, plus video and audio references, and it isn't on the chopping block.