WAN Image_Embeds Multi-Switch

In WanVideoWrapper (Kijai's Wan pack, the one that makes Wan 2.1/2.2 run well in ComfyUI), image-to-video works by encoding a reference or start image into embeddings that condition the sampler. Those embeddings are the WANVIDIMAGE_EMBEDS type, and this node is a five-slot switch that picks between them. Same first-non-None logic as every multi-switch in this pack: wire up to five image-embeds inputs, bypass the ones you don't want, and the first live slot flows out the image_embeds output. No widget to flip.

Why switch pre-encoded embeds instead of just swapping the image upstream? Encoding is not free - it runs the image encoder every time the graph re-executes - and re-running it changes the latent you hand the sampler, which can shift your result even with the same seed. If you're doing I2V with a few candidate start frames (different compositions, different subjects, a "keep this frame" versus a "let it drift" version), you can encode each once, park them in this switch, and pick between them by bypassing - no re-encode, no rewiring, no surprise variance from a fresh encode pass.

The README's placement rule matters here more than most places: bypassable options in the early slots, always-on default last. The search starts at input1, so an always-enabled encode in slot 1 will be first-non-None forever and shadow the rest. If you want slot 2 to ever win, slot 1 has to be bypassable.

It also carries the Purge_VRAM toggle (the only required widget, default off). On an embeds switch that's mostly irrelevant - embeds are small tensors, not models - but it's there for symmetry with the model switches in the pack, and it does unload every model before the switch runs if you ever want a clean slate for the branch downstream.

Output is image_embeds, which wires into the WanVideoWrapper sampler's image-embeds input (or into the I2V conditioning chain).

Installing it

Ships in ComfyUI-RvTools_v2 - Manager → search ComfyUI-RvTools_v2, or clone https://github.com/r-vage/ComfyUI-RvTools_v2 into custom_nodes and restart. No models, no keys. But the WANVIDIMAGE_EMBEDS type only exists if ComfyUI-WanVideoWrapper is also installed - that's the implicit dependency for every WAN-type node in this pack, and it's not in RvTools' own requirements.

The pack is unmaintained (README → ComfyUI_Eclipse); the [RvTools] suffix is from v2's rename after v1's repo disappeared.

Common issues

• No WANVIDIMAGE_EMBEDS sockets anywhere - WanVideoWrapper isn't installed. Install it first, then this switch has something to route.
• Slot 1 always wins - an always-enabled source in slot 1 is never None, so the other four slots never get looked at. Keep the default in the last slot.
• "Missing Input" error - all five slots empty or bypassed. Keep one wired.