LRW WAN Latent Guide Blend

Every other node in the WAN branch builds or measures something. This is the one that acts. LRW WAN Latent Guide Blend takes your actual WAN2.2 First-Last Frame video latent, takes a selected geodesic keyframe, and blends the keyframe into the video latent along the time axis - before it ever reaches KSampler. It's the practical injection point, the reason the whole pack exists, and the node with the most reported confusion around it.

The big idea

WAN2.2 FLF is great at hitting the last frame, but with large pose, gaze, or framing changes the middle can be abrupt - or the clip "holds" the first frame too long and transitions late. The author's fix: compute geodesic keyframes between first and last, pick one, and softly steer the intermediate video latent toward it. Direct FLF stays the main continuity path (so you still arrive at the real last frame); the blend just gives the sampler a smoother middle direction. He's said it plainly: it's a soft guide, not a magic quality booster. Its measurable effect is motion timing - motion starting earlier and distributing more evenly.

The inputs that matter

• video_latent - your WAN video latent, typically from WanFirstLastFrameToVideo. Must be 5D (B, C, T, H, W); it'll get converted if not.
• keyframe_latents - the stacked output of LRW_WanGeodesicKeyframes.
• batch_size (default 1) - the batch used upstream by the keyframe generator. Leave at 1.
• keyframe_index (default 2) - which keyframe to blend in. For 5 keyframes, 2 is the middle; match this to how many you generated.
• blend_strength (default 0.15) - how strongly the guide affects the video. README's recommended starting point: 0.15, with the VRAM guide suggesting 0.05–0.10 on 12GB and 0.10–0.20 on 16GB.
• time_schedule - all_frames (uniform), middle_focus (default, peaks mid-clip), or late_focus (heavier toward the end). The author's own tip: the FFLF first-frame anchor can overpower late_focus, so it can feel too soft - that's expected behavior, not a bug.
• normalize_guide (default true) - rescales the guide latent's norm to roughly match the video latent's. Usually leave it on; the author suggests toggling it off as a debugging step.

Outputs: latent (the blended 5D latent - this replaces your video latent's path into KSampler) and info (a string confirming shapes, selected index, and schedule).

The real-world gotcha (read this)

The most common complaint from the community - with the author in the thread agreeing - is "I enabled it and the output is identical." If your result doesn't change even at blend_strength=1.0, work through these, in order:

1. Verify KSampler's latent_image really comes from this node's latent output, not the un-blended FLF latent.
2. Confirm the shapes match (the info string tells you both).
3. Try normalize_guide=false and different keyframe_index values.
4. Check whether another node is overriding the latent downstream.

One user found the blend only visibly mattered when the high-noise KSampler's start step was raised from 0 to 1 - worth a try if you're on the two-pass WAN 2.2 setup and seeing nothing.

Install

Search comfyui-lrw-nodes in ComfyUI Manager, or git clone https://github.com/lajjadred/comfyui-lrw-nodes into ComfyUI/custom_nodes, pip install -r requirements.txt, restart. Real deps: latent-riemannian-world >= 0.3.0, torch >= 2.4. The demo workflow also wants ComfyUI-GGUF, WanFirstLastFrameToVideo, and VideoHelperSuite. Early pack builds had a registration bug that showed nodes as broken/UNKNOWN - git pull + full restart, and keep a single copy in custom_nodes. License is BSL-1.1.