Frames Concat Latent Raw

This is a video-workflow node. Frames Concat Latent Raw takes two latents and joins them end to end along the frame dimension - clip A followed by clip B - in latent space, without decoding to pixels and re-encoding. The "Raw" in the name is the promise: it's a straight concatenation, no blending, no interpolation, no VAE round-trip. Frames in, longer frames out.

Where this fits: RES4LYF leans hard into video, especially Wan, and the author's own claim to fame includes generating 601 frames in a single shot on a 4090 with the pack's Wan sliding-window nodes. Once you're working with sequences of latent frames, you often want to assemble them - extend a clip, splice two generations, build a longer timeline out of pieces. Doing that in latent space keeps you off the VAE, which is both faster and cleaner: every encode/decode cycle costs quality, so avoiding one is free quality.

How it works

Both inputs are LATENT batches shaped as frames. The node concatenates them so frames_0 comes first and frames_1 follows, producing one continuous latent sequence. Because it's raw, the two pieces are simply placed adjacent - the node isn't smoothing the transition or matching them up. Whatever is at the tail of the first and the head of the second is exactly what meets at the join.

The inputs and output that matter

Refreshingly, there are only two inputs and no knobs:

• frames_0 (LATENT) - the first clip's latents. These land at the front.
• frames_1 (LATENT) - the second clip's latents, appended after.

Output is a single latent (LATENT) - the combined sequence. Feed it into a VAE Decode to see the joined result, or into another sampler pass / another concat if you're building something longer.

How to install it

ComfyUI Manager: search RES4LYF, install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/ClownsharkBatwing/RES4LYF/, activate the venv, pip install -r requirements.txt (portable ComfyUI: use the embedded python's pip), restart, hard-refresh with F5. Nothing model-specific to download for this node - it's pure latent tensor handling - but you'll obviously need a video model (Wan, LTXV) elsewhere in the graph for the latents to mean anything.

Common issues

The thing that'll bite you is shape mismatch. Concatenating along frames only works if the two latents agree on everything else - same resolution, same channel layout, same latent format from the same model family. Try to join a latent from one model with a latent from another, or two clips at different resolutions, and you'll get a tensor-shape error at best, garbage at worst. Keep both sides coming from the same pipeline.

The other thing to set expectations on: because it's raw, the seam can show. If clip A and clip B weren't generated with continuity in mind - same subject, compatible motion, matching lighting - the cut at the join will look like a cut, because that's literally what it is. This node concatenates; it doesn't blend. If you need a smooth interpolated transition between two clips, that's a different job for a different tool. Use this when you want the pieces placed exactly and predictably, which is often precisely what you want for stitching batches you already control.