RIFE Frame Interpolation

If ComfyUI just threw you a "missing node pack: GACLove/ComfyUI-VFI" error after loading someone's video workflow, you're in the right place - that's basically the only way people stumble onto this pack. It's a small two-node repo built for one job: smoothing out choppy AI-generated video by inventing frames between the ones you already have. RIFE Interpolation is the node that actually does the interpolating; its sibling, Calculate Loaded FPS, is a tiny helper for getting your input fps right before it hits this node.

This is the same job every RIFE-flavored node in ComfyUI does - Wan, LTX, Hunyuan and AnimateDiff all generate video at a modest native frame rate to keep compute sane, and interpolation is the standard back-half step that turns a stuttery 16-24fps clip into something that plays smooth (the KB's own Wan production notes spell out that exact pipeline: generate, interpolate, then upscale). What's different about GACLove's version is the interface. Most RIFE packs make you set a "multiplier" - double the frames, quadruple the frames - and do the fps math yourself. This node just asks for your actual source_fps and target_fps and works out how many in-between frames it needs. That's friendlier for a beginner, and it doesn't require target_fps to be a clean multiple of source_fps either - 24 to 60 is a perfectly normal request here.

How it works

Same mechanism as any RIFE model: it looks at pairs of consecutive frames, estimates the optical flow between them - where each part of the image is moving - and synthesizes new frames along that flow instead of just cross-fading pixels. That's why it handles pans and steady motion cleanly, and why it can smear on fast, chaotic action or anything that gets occluded and un-occluded between frames. This pack ships with exactly one RIFE checkpoint (flownet.pkl) - there's no menu of model versions to pick between the way the bigger Frame-Interpolation-style packs offer. One model, one job.

The inputs that matter

images is your frame batch in, straight from a VAE decode or a video loader. source_fps and target_fps are the whole point of the node - set them to what you actually have and what you actually want. scale is your quality/speed dial: the tooltip is blunt about the trade-off - lower values (0.25-0.5) process faster but may reduce quality, higher values (up to 4) push for better results if you've got the compute to spare. Of the optional inputs, batch_size trades VRAM for throughput - raise it if you have headroom, lower it if you're OOMing - and use_fp16, on by default, halves memory use and speeds things up, but its own tooltip is explicit: it requires a CUDA GPU. The single images output is your interpolated sequence, ready to feed a Video Combine or preview node.

Installing it

Easiest path: search "ComfyUI-VFI" in ComfyUI Manager and install from there. By hand, it's the standard drill - cd ComfyUI/custom_nodes && git clone https://github.com/GACLove/ComfyUI-VFI, then pip install -r requirements.txt from inside your ComfyUI environment, then restart. One thing worth flagging: the README's own install snippet has a copy-paste placeholder left in it - it shows git clone https://github.com/your-username/ComfyUI-VFI.git, which isn't a real URL. Use the one above instead.

The RIFE model is supposed to download automatically the first time you run the node, but the README doesn't say from where - if your network can't reach it, or the auto-download just doesn't fire, grab flownet.pkl from the official RIFE repo yourself and drop it in ComfyUI/models/rife/ (or ComfyUI-VFI/rife/train_log/ inside the pack itself - either location works, per the README).

Where people get stuck

Because this pack is thin - no bundled model picker, no compile step, no ensemble toggle - most of the friction is upstream of the node itself: getting it installed at all if you've never touched git before, which is exactly the scenario that brings most people here in the first place (it tends to show up as a "missing node" error inside someone else's shared workflow, not something people go looking for on its own). Beyond installation, make sure use_fp16 is off if you're not on an NVIDIA card - the tooltip's CUDA requirement is your warning sign, and leaving it on without CUDA is a likely source of a crash rather than just a slowdown. And treat scale the way you would any RIFE scale factor: it's a speed/VRAM knob, not a magic quality switch - pushing it to 4 on a long clip mostly just costs you time.