MiniMax H3 Video Interpolate (RIFE-lite)

Let's get the naming out of the way, because it's the most interesting thing about this node: the "(RIFE-lite)" in the display name is doing a lot of work. RIFE is a real, learned frame-interpolation model - the kind that estimates optical flow and genuinely invents intermediate frames. This node is not that. It's a torch.lerp - linear blending between consecutive frames. RIFE-lite is the marketing version of "plain linear interpolation," and once you know that, the node makes perfect sense: it's a cheap way to raise the effective frame rate of your H3 clip without re-running the sampler, and it's not pretending to be a quality interpolation pass.

So what's it for? You generate 65 frames at H3's default, which is ~2.7 seconds at 24 fps. Interpolate with multiplier 2 and you get ~130 frames - still short, but at 48 fps it looks smoother, and if you're exporting at 24 fps you've effectively lengthened nothing but smoothed everything. The honest use case is exactly that: get the clip to a frame count and frame rate that plays back nicely, fast and free, before export. The multiplier input (2–8, default 2) inserts that many blends between every pair of frames - 2 doubles your frame count, 8 makes it 8×.

The mechanism, honestly. For each pair of consecutive frames A and B, the node creates multiplier − 1 intermediate frames by lerp(A, B, t) for t = 1/n, 2/n, .... Zero motion estimation, zero optical flow, no model, no VRAM, runs in milliseconds. The consequences follow from that: slow, smooth motion interpolates beautifully; fast motion gets ghosting and trails, because a linear blend of a moving subject is a double-exposure, not a new pose. That's the exact trade you're making against a real RIFE/FILM pass - and the README says so, recommending you "swap in RIFE/FILM for higher-quality results" if you have them.

Inputs/outputs. Just images (required) and multiplier. One output, IMAGE, of the interpolated frames - same type as the input, so it slots right into your existing wiring between VAE Decode and VideoExport. The only output name in the schema is IMAGE (the node's RETURN_NAMES is empty, so the socket shows the type).

Install. Same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/TensorVizion/Minimax-H3-Video-Music-Nodes

or ComfyUI Manager → search the pack title → restart. Deps are just torch, torchaudio, numpy, Pillow. No ffmpeg needed here - that's the Export node's job.

Where it bites. Set your expectations by what it is: if you interpolate a fast action shot 8×, it will look like a morphing smear, and no setting fixes that because there's no motion model to fix it with. Use modest multipliers (2–3×) for smoothing, and reach for a proper interpolation pack when the shot has real movement. Also note the node returns early if you feed it a single frame - fine, not a bug. And one small thing the README doesn't shout about: it's a linear blend in whatever color space your frames are in, so it can slightly wash out bright motion. For a tool that costs zero VRAM and zero generation time, that's a fair price. Just don't expect RIFE - the real one - from the "RIFE-lite" label.