Prompt Relay Encode (Kijai) — deprecated

This node is deprecated, and the pack is refreshingly upfront about it: "use PromptRelayEncodeC2C with backend='kijai'." If you're building a new workflow, skip this page and read that one - the unified encoder does everything this one does with a cleaner canvas. You only land here because you opened an older workflow that still uses it, in which case you should know what it is before you rip it out.

What it was

Prompt Relay Encode (Kijai) was the WanVideoWrapper-specific encoder for Gordon Chen's per-segment prompting technique: it takes a WANVIDEOMODEL from WanVideoModelLoader plus a WANTEXTENCODER from LoadWanVideoT5TextEncoder, and produces the text embeddings for a video where each segment has its own prompt and attention is patched so the segments relay across time (the epsilon knob, default 0.001, sharpens or softens the boundaries between segments).

The inputs

The inputs are all the kijai flavor of the unified node: model and t5 in, global_prompt (the persistent base prompt), local_prompts (per-segment, |-separated), segment_lengths (comma-separated pixel-space frame counts, empty = equal segments), negative_prompt (encoded once for negative embeddings), latent_frames (default 81 - the Wan math, (pixel_frames-1)//4 + 1), and encode_device (gpu/cpu) for where the T5 encode runs. Optional relay_options accepts a RELAY_OPTIONS bundle from Prompt Relay Advanced Options. Outputs: model and text_embeds (WANVIDEOTEXTEMBEDS), straight into the sampler.

Should you migrate?

The practical story is short: if a saved workflow loads with this node, it works, and there's no rush to migrate - the pack keeps it registered precisely so old graphs don't break. But new work should use the unified PromptRelayEncodeC2C with backend=kijai, which collapses this into one node with dynamic sockets and adds the same knobs. Migration is mechanical: same model, same t5, same prompts, same latent_frames - the values map one-to-one. If you see "arrived as None" errors on this node, it's the same stale-wiring failure mode as the unified version: a disconnected upstream input, and the node refuses to guess.

Install

Install is the shared pack (clone Code2Collapse/ComfyUI-CustomNodePacks into ComfyUI/custom_nodes or ComfyUI Manager → "CustomNodePacks"). One thing to keep regardless of which encoder you use: the latent_frames number has to match your actual video length, or your per-segment prompts land at the wrong frames - that failure mode hasn't changed between the old and new nodes, because it's inherent to the technique.