RIFE Frame Interpolation

Want smoother slow-motion or a higher frame rate from frames you already generated? That's frame interpolation, and RIFE (Real-Time Intermediate Flow Estimation) is the standard tool for it. AUNRIFE wraps RIFE v4.7 as a single ComfyUI node, and its best trick is one most RIFE nodes get wrong: it fetches its own model weights. Point it at a batched image tensor, set a multiplier, and the rife47 or rife49 .pth file downloads from HuggingFace to ComfyUI/models/rife on first use. No separate manual model hunt, no "where do I put this file" step.

How it works, quickly: RIFE estimates optical flow between each pair of frames and synthesizes the in-between ones. The node takes your batched IMAGE (B,H,W,C), and for every adjacent pair generates multiplier - 1 intermediates - a multiplier of 2 doubles your frame count, 10 makes it ten times longer. Output is the interpolated sequence as an IMAGE tensor, ready to hand to a video combine node.

The inputs that matter: images (your frame sequence - needs at least 2 frames; with fewer it passes the input back unchanged), ckpt_name (rife47 or rife49, both run the v4.7 architecture), multiplier (INT, 2–10), and ensemble - a boolean that runs the model twice per intermediate frame and averages the results. Ensemble costs roughly double the time for modestly better quality; I'd leave it off while you're testing, then flip it on for the final render if the motion artifacts bother you.

A couple of implementation details worth knowing before you're surprised by them. It's adapted from the ComfyUI_Fill-Nodes pack (filliptm), with architecture credit to Practical-RIFE - so this isn't a reimplementation, it's the known-good integration. Frame pairs are processed sequentially behind a ComfyUI progress bar. And if inference fails for a pair - say you run out of VRAM - the node falls back to linear interpolation for that step instead of aborting the whole run. That's the good kind of failure mode, but it's also how you get a subtly softer-looking segment if you push the batch too hard; check for OOM warnings in the console if one stretch looks less sharp.

Where this slots in: feed it frames from a video helper (VHS) or an AnimateDiff/sequence output, interpolate, then recombine to video. It's not a video-length magic wand - it's a frame doubler, and a good one.

Install: ComfyUI Manager (search "AUN"), or cd ComfyUI/custom_nodes && git clone https://github.com/loz2754/AUN-ComfyUI-Nodes, restart. Dependencies are the pack's standard light set; the first run also needs internet access to pull the weights (~tens of MB, one-time).

Troubleshooting:

• First run appears stuck → it's downloading rife47; watch the console for the progress line.
• Output looks identical to input → you gave it fewer than 2 frames.
• A middle segment looks softer → likely OOM on that pair, node fell back to linear interpolation; lower the multiplier or batch size.
• Interpolation is slow on a huge batch → normal; sequential pairs + high multiplier is inherently serial.