WanVideo Add LucyEdit Latents

Lucy Edit is Decart.ai's open-weight video editing model - the one that got pitched, half-jokingly and half-seriously, as "open source Nano Banana for video" when it dropped in late 2025. It does instruction-based video editing: describe the change, feed it a source clip, get an edited clip back, no mask required. Decart shipped their own ComfyUI node pack for it (lucy-edit-comfyui), but kijai also built native support straight into WanVideoWrapper, and this node is the connector piece - it's how you fold Lucy Edit's editing signal into a WanVideoWrapper image-embeds pipeline instead of running a separate, parallel graph.

What it does

This is a small, single-purpose node: it takes your existing WanVideoWrapper image embeds and merges in a set of extra latents shaped for Lucy Edit's editing conditioning, then hands back an updated embeds object. Practically, that means: build your Wan image embeds as normal (source video encoded to latents, whatever base conditioning you're using), then route them through this node along with the Lucy Edit latents before they reach your sampler. It doesn't do the editing itself - the actual denoising and instruction-following happens in the sampler downstream, using the model loaded for Lucy Edit. This node's job is purely to get the extra conditioning into the embeds object in the shape the sampler expects.

The inputs and output that matter

Two required inputs, both already-prepared objects rather than raw media: embeds (type WANVIDIMAGE_EMBEDS), your existing image-embeds bundle, and extra_latents (type LATENT), the Lucy Edit-specific latents you're adding. There's nothing to tune on this node itself - no strength slider, no percent window - the conditioning strength lives wherever you built those latents upstream.

The output is a single image_embeds (WANVIDIMAGE_EMBEDS), the merged result, which feeds straight into your Wan sampler alongside your text embeds and model.

How to install it

Through ComfyUI Manager, search WanVideoWrapper and install; manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper

then restart. Lucy Edit's own weights (decart-ai/Lucy-Edit-Dev on HuggingFace) aren't part of the node pack - download them separately into your Wan diffusion model folder like any other checkpoint this wrapper loads. Since this node only merges an already-built latent object, most of the actual setup work - getting Lucy Edit's model loaded and your source video encoded correctly - happens elsewhere in the graph, not here.

Common issues & troubleshooting

If you're not running a Lucy Edit model, you don't need this node. It's specific glue for one editing model's conditioning shape. Bolting extra_latents from something unrelated onto your embeds won't do anything useful and will likely just confuse the sampler.

A blank or mismatched-shape extra_latents input is the most likely failure mode. Since there's no validation UI here - it's a straight merge - if the latents you feed in weren't actually built for Lucy Edit (wrong resolution, wrong frame count relative to your source video), expect either an outright error downstream or an edit that just doesn't follow your instruction. Trace back to whatever node produced extra_latents before assuming this merge node is at fault.

Community reception on Lucy Edit itself was mixed on the marketing, not the tech. The "Nano Banana for video" framing drew some eye-rolling when it launched - people pointed out it has nothing to do with Google's model - but the open weights themselves landed as a real, welcome release. Go in expecting a genuine instruction-editing model, not a video Nano Banana clone, and the results read better.