Temporal Flow Average

You've seen it a hundred times: a Seedance or Kling clip that's gorgeous for 95% of its runtime, then a single frame pops with a color spike, or the whole thing shimmers with chroma flicker. Topaz and upscalers won't touch it, because the noise is temporal - it lives in the differences between frames, not inside any one frame. Temporal Flow Average is the heart of the FlowDenoise pack: it turns a noisy AI clip into a clean reference for everything downstream.

What it actually does

It builds a "clean" version of your clip with the classic compositor's trick: frame blending with motion compensation. Instead of naively averaging neighbors (which turns motion into smears), it estimates the optical flow - which pixel moved where - warps each neighbor to match the current frame, then averages only the matching pixels. Frames further back weigh less, so the current frame stays dominant.

Three mechanisms keep the average from destroying your image:

• Outlier rejection - after warping, a per-pixel color-similarity check (color_threshold) down-weights pixels that don't match the reference. This is the anti-ghosting mechanism.
• Occlusion detection - a forward-backward consistency check keeps pixels revealed or covered by motion from polluting the average.
• Scene detection - a vectorized per-pair MSE scan finds cut boundaries and refuses to average across them.

The flow comes from MEMFOF (the default - a strong 2025 optical-flow model) or RAFT (raft_small / raft_large, both inside torchvision, no download needed). MEMFOF's weights auto-download from HuggingFace on first use.

The inputs that matter

You can tune all ten; these are the ones a beginner actually sets:

• window_size - how many frames to average on each side (total window = 2n+1). Default 2 means five frames; 2–3 is the sweet spot for AI video.
• weight_decay - exponential falloff per frame distance; lower = more aggressive averaging. 0.7–0.8 for flicker, toward 0.6 for heavy grain removal.
• flow_model + flow_iterations - keep memfof (8 iterations is right); if you switch to RAFT, bump iterations to ~20.
• color_threshold - lower = stricter outlier rejection; this is your ghosting dial. The default 0.04 is fine, but see troubleshooting.
• batch_size - MEMFOF processes frames in chunks on the GPU. Higher = faster but more VRAM (the README suggests 8–16 on a 32GB RTX 5090 at 720p).
• precision / flow_scale - the speed knobs. bf16 is 1.5–2× faster on RTX 30/40/50 with negligible quality cost; flow_scale 0.5 computes flow at half res (~3× faster) while warping stays full-res. If it's slow, these are the time-savers.
• output_weight_map - leave it off - it allocates a full-size per-pixel tensor, tens of gigabytes of CPU RAM on long clips. The pack gated it off by default after exactly that OOM regression.

Outputs

• clean - the temporally averaged frames. The money output.
• weight_map - the per-pixel confidence map, or a 1×1×1×3 placeholder when output_weight_map is off, so downstream connections stay valid.

You rarely use clean directly. The standard pipeline is VHS_LoadVideo → TemporalFlowAverage → SelectiveDenoise: clean frames feed Selective Denoise's clean input while the original video stays as original - that's how you keep detail. Wire a copy of clean into Extract Noise to see what's being removed.

Installing it

Via ComfyUI Manager, search "ComfyUI-FlowDenoise" and install. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/AIMZ-GFX/ComfyUI-FlowDenoise.git

then restart. One dependency gotcha: memfof is not on PyPI, so pip install memfof fails with "Could not find a version" - the author has had to answer exactly this in his release thread. Use the GitHub URL:

pip install git+https://github.com/msu-video-group/memfof.git

Portable installs need the embedded Python: python_embeded\python.exe -m pip install git+.... Manager's requirements.txt / install.py handle it automatically. You'll also want Video Helper Suite, since the shipped workflow uses VHS_LoadVideo / VHS_VideoCombine.

Troubleshooting

• pip install memfof fails - you need the git+ URL above; plain pip install memfof never works.
• Ghosting on moving subjects - the most-reported complaint on the pack's launch thread; lowering color_threshold helps but doesn't fully eliminate it, and the author is working on it. A smaller window_size and lower weight_decay reduce the smearing meanwhile.
• Flicker still visible after one pass - real users report needing two passes to fully kill visible flicker. The noise visualization still shows residue after it looks clean to the eye.
• Long clips eating RAM - that's what output_weight_map gating and vectorized scene detection fixed. A DefaultCPUAllocator OOM on long clips means you're on an old version; update.
• BF16 issues on older GPUs - bf16 assumes an RTX 30/40/50-class card. If yours predates that, switch precision to fp32.

The name undersells it: this isn't just a denoiser, it's the reference-frame generator that makes the rest of the pack controllable. Get it right here and Selective Denoise becomes a simple dial.