AetherScale • Video Combine

If you've copied a video workflow off the internet, you already have VHS Video Combine in it - it's the default terminal node of the genre. AetherScale • Video Combine is a drop-in replacement for that slot: an IMAGE batch (plus optional AUDIO) in, an MP4/MKV/MOV out, with NVENC, x264 or ProRes doing the encoding.

Two reasons to reach for it over the incumbent. It never holds more than a bounded chunk of frames in RAM, which matters when your batch is 1500 frames of upscaled video. And the preview widget behaves: v0.9.2 of the pack was largely a fix for this node growing taller after every execution - the most annoying bug class in the video-node space, if you've ever left one sitting in a graph overnight. The preview is now one LiteGraph DOM widget, sized from the node width and the real aspect ratio.

How the fast path works

It isn't frame-by-frame Python. Frames are packed several at a time into one contiguous RGB8 buffer in a single Torch operation, handed to FFmpeg through the buffer protocol (no separate .tobytes() copy), and a small producer thread pre-converts the next chunk while FFmpeg/NVENC consumes the current one. Memory stays flat because chunk_mb caps the chunk and pipeline_depth caps how many are in flight - the README's own numbers are chunk_mb = 64 and pipeline_depth = 2, and those defaults are fine.

Before a single frame is streamed, the node runs an NVENC preflight: a real test encode on the physical GPU, with the actual pixel format and geometry. If the encoder can't do it, it either falls back to another NVIDIA encoder (nvenc_codec_fallback, on by default) or fails in a second instead of twenty minutes in. Odd dimensions get padded where yuv420p demands it.

Inputs that matter

frame_rate - the number that decides whether your clip is the right length. If you loaded with AetherScale • Video Loader, wire its frame_rate output straight in. codec - h264_nvenc is the default and the right choice most of the time; hevc_nvenc and av1_nvenc for smaller files, libx264 when there's no NVIDIA hardware, and the prores_* ladder when you're handing off to an editor. container is mp4/mkv/mov, but pick any prores_* codec and it forces .mov with the matching prores_ks profile - ProRes ignores bitrate_mbps, preset and nvenc_gpu entirely, and 4444/XQ preserve alpha if you handed it a 4-channel image.

bitrate_mbps (default 20), preset (p3 for speed, p4 for a little more compression), pixel_format (yuv420p for anything you'll play in a browser), nvenc_gpu - leave it on auto unless you have two cards, where the multi-GPU handling is genuinely the nice bit: the chosen physical adapter is isolated for the FFmpeg child and remapped to logical CUDA GPU 0, so a PyTorch-restricted device list and nvidia-smi's physical ordinals stop disagreeing.

save_output - flip it off and the preview goes to ComfyUI's temp folder, still viewable, not littering your output dir. save_metadata and metadata_target are the interesting ones: the node walks upstream from the images link, finds the nearest seed/sampler/scheduler/model, and writes them into a <video>.metadata.json sidecar, into the container as readable tags, or both. Override any of it with the seed, sampler_name, scheduler and model_name inputs. The encode is where a workflow's provenance usually dies, and this node tries not to let it.

audio and audio_bitrate_kbps mux an AUDIO input as AAC via a temporary WAV. save_silent_copy is off by default and you almost never want it on - it just adds a second, silent file.

Outputs are video_path and stats, both strings; stats carries the encoder, routing and frame counts.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/vizart-vj/ComfyUI-AetherScale.git

or Manager → search AetherScale → install → restart. No pip packages; NVIDIA's VFX wheel is bootstrapped inside the pack rather than installed into your environment.

Where people get burned

FFmpeg must exist. AetherScale checks PATH, then the binary bundled with VideoHelperSuite and a few known VHS paths - so on a machine that already has VHS there's nothing to fix. Without it you get a plain "FFmpeg was not found" error and no output.

ProRes won't preview in the node on most browsers, because browser support for MOV/ProRes is what it is. The file is fine; open it in a player.

Windows and NVIDIA, effectively. The generic path is just FFmpeg and will work elsewhere with libx264, but the NVENC routing layer assumes nvidia-smi, and the rest of the pack is Windows-only.