Save Video (C2C)

Two separate problems, one node. The first: the default way to save a clip in ComfyUI is VAE Decode → Video Combine, which materialises the whole frame batch in RAM before anything gets written. The second is delivery - video_00003.mp4 is fine to post and useless the moment the shot is going into Resolve.

Save Video (C2C) is an output node that writes an IMAGE batch into sixteen formats, muxes audio, optionally applies an OCIO colour transform, and shows a browser preview with a timeline scrubber after each run.

What you get

The format list is the point: MP4 H.264, MP4 H.265 10-bit, four ProRes profiles (422 up to 4444 XQ), three DNxHR, FFV1 lossless in MKV, WebM VP9, GIF, animated WebP, 16-bit PNG sequences and half/full-float EXR sequences. Delivery-grade formats sit beside the social-media ones - no second pack for the "pro" half.

The bit that isn't in other save nodes

Stream save. Wire latent and vae instead of images and the latent is decoded straight into the file, in order, keeping the VAE's causal state - the full frame batch never exists in RAM. For long clips on a card that barely fits, that's the difference between finishing and OOMing.

With a Wan VAE there's tile_mode. On auto it tiles only when the frame wouldn't fit the GPU (the tooltip's own example is 2K on 8 GB landing on 512 px tiles), and each tile is decoded over the whole clip in one causal pass and blended on disk - so the tile is never held as a batch either. off disables it, manual takes your tile_size.

The encoder underneath is PyAV, imported lazily, and PyAV bundles FFmpeg - no separate ffmpeg binary to install.

Inputs you actually set

• format - start here, everything else follows.
• images or latent + vae - the normal path versus the streaming path.
• audio - muxed in for formats that carry it.
• video_info - a VHS_VIDEOINFO, and the trick worth knowing: leave fps at its default 24, wire video_info, and the loader's loaded_fps is used instead. Load a 30 fps source, process, save, and it stays 30 fps without you remembering the number. Set fps to anything else and yours wins.
• quality - CRF / qscale / ProRes profile depending on format, and hidden for lossless formats. If the widget won't move, you picked FFV1 or an EXR sequence.
• filename_prefix (default C2C/video), subfolder, save_output - normal output plumbing.
• naming - default, ComfyUI counter, or Folder Version. That third one is the pack's own FolderIncrementer: output lands in v001/, v002/, and a cancelled run doesn't burn a version because the version is resolved off the filesystem rather than a counter file.
• colorspace_in / colorspace_out - sRGB - Display, Linear Rec.709 (sRGB) or ACEScg, with colorspace_out defaulting to "same as input". It uses OCIO when a config is available, built-in matrices otherwise.

Outputs

filenames (STRING) is where the file went. preview (IMAGE) is what the in-node scrubber plays. It's an output node, so it always runs, with its own change-hash so an unchanged graph stays cached.

Installing it

Manager → search "CustomNodePacks", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Code2Collapse/ComfyUI-CustomNodePacks

Restart and look for [C2C] CustomNodePacks: 112 nodes loaded ... 0 failed. No line means the pack didn't import, and nothing else matters yet.

Don't run pip install -r requirements.txt on a working ComfyUI - the README says so outright and it's right: it overwrites bundled torch, torchvision, numpy and Pillow. The repo's requirements.txt is also longer than the README's table (timm, einops, huggingface_hub, ftfy, regex, cryptography, for the SAM 3 and Vault nodes). PyAV, numpy and torch are the whole bill for this node.

Where it bites

ModuleNotFoundError: av - install PyAV. Mainstream ComfyUI installs already have it for the video nodes; a stripped-down or old environment won't.

Your export looks washed out or flat. Almost always a colour-space mismatch: many graphs run linear or ACEScg internals, and saving that as display-referred sRGB gives you a milky file. Set colorspace_in to what the graph is holding and colorspace_out to sRGB - Display for delivery.

Alpha and odd sizes. Only ProRes 4444, FFV1 and VP9 carry alpha; H.264 and H.265 are 4:2:0 and can't - the same constraint pads odd width/height by a pixel with edge replicate. Animated WebP goes through Pillow rather than PyAV on purpose: PyAV's libwebp encoder segfaulted the process on an RGBA batch (PyAV 17.0.1), and inside ComfyUI a segfault kills the server.

The scrubber and non-browser codecs. The preview runs in your browser, so H.264 and VP9 play happily while ProRes and DNxHR may not. That's the browser, not a bad write - check those in your NLE.

RAM still spiking. You're on the images path. Streaming only happens when the latent goes in and the frames come out one at a time.

For a quick MP4 to post, the node you already have is fine. Add this one for the two things it really sells: long clips on limited VRAM, and a delivery-grade file when the shot is going somewhere else.