Load Video FFmpeg (Topaz)

If you already pay for Topaz Video (the product formerly known as Topaz Video AI), this is the clean way to hand your generated clip to it from inside a ComfyUI graph. Topaz is still the paid baseline people benchmark open upscalers against, and this pack is its ComfyUI front end. TopazLoadVideoFFmpeg is the loading half of that deal: it turns a video in your ComfyUI/input folder into the image batch everything downstream expects - decoded with Topaz's own bundled ffmpeg, not whatever's on your PATH.

The README is blunt about what this is for: short AI-generated videos, not your wedding footage. ComfyUI moves video around as image batches, so the node has to decode everything to tensors and later re-encode, which costs more time than rendering in the Topaz GUI. Fine for a 3–5 second Hunyuan or AnimateDiff clip, slow for a 20-minute file.

How it works

Mechanically this is a fork of VideoHelperSuite's LoadVideoFFmpeg - the docstring in video_load.py says so outright. It was ported so max_duration could stay in seconds (matching start_time, which helps when you're splicing clips) and so the loader reuses the same Topaz install-detection logic as the rest of the pack. At runtime it spawns ffmpeg with -i video, parses the stream header for resolution, fps, and whether there's an alpha channel, then decodes straight to rgba64 raw frames and converts them to float32 tensors in [0,1]. Audio isn't pulled eagerly - it's a lazy load keyed off the same file, so you don't pay for a track you never use. There's a psutil-based frame cap that keeps a runaway video from eating all your RAM.

The inputs that actually matter

• video - a dropdown of every supported file in ComfyUI/input (mp4, mov, mkv, webm, gif, avi, flv, ts, m4v, mpg, mpeg). Dump your clip there and refresh.
• force_rate - target fps. 0 (default) means keep the source frame rate. Set it to match your model's native fps instead of guessing.
• custom_width / custom_height - resize before anything downstream. 0 keeps the source size. Both being 0 is the norm.
• max_duration / start_time - in seconds, 0 = off. Trim to a clip with start_time, cap length with max_duration. These are the fields the port exists for.
• meta_batch (optional) - a VHS_BatchManager, for when you're driving a batch through the VHS ecosystem.

Outputs and where they go

It returns IMAGE, frame_count, mask, audio, and video_info - the same contract as VHS, so anything that accepts a VHS video pipe can sit downstream. The one that matters: wire IMAGE into the pack's TopazVideoAI node (which runs the actual tvai_up / tvai_fi ffmpeg filters), then out through VHS_VideoCombine. The mask is the alpha channel inverted, or all zeros when there's no alpha. video_info is a VHS_VIDEOINFO dict with source fps/duration/dimensions plus what actually got loaded - handy if a later node needs to know the real frame count.

Install

Topaz Video must be installed and licensed first - this pack ships no weights, and all its model files come from your Topaz install. Then either search "TopazVideoAI" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/sh570655308/ComfyUI-TopazVideoAI.git

Restart ComfyUI. Then close the GUI, open a terminal in the Topaz install folder and run .\login (Windows) once, or the upscale node will fail licensing checks later.

Gotchas

Two README-flagged traps are worth knowing before you hit them. If you upgraded from old "Topaz Video AI", delete the stale TVAI_MODEL_DIR / TVAI_MODEL_DATA_DIR user environment variables - post-rename they can point at a model folder that no longer exists, which shows up as Model not found in the upscale node (the pack auto-detects the new path, but the leftovers cause confusion). And for this loader specifically: if Topaz's ffmpeg can't be found you'll get an ffmpeg subprocess error, and a deleted file produces Invalid video file. The No such filter: 'tvai_up' error you'll read about belongs to the upscale node - it means the ffmpeg path points at a system build without Topaz's filters, not at Topaz's own.