👺 Save Video To HF Dataset

Video datasets don't build themselves, but they can almost build themselves. 👺 Save Video To HF Dataset takes the video files your workflow already produced and pushes them to a Hugging Face dataset repository - no manual download-and-reupload between ComfyUI and the HF website. The natural fit is a batch generation pipeline: render a bunch of clips, let this node ship them all to your dataset repo, and start a LoRA training run on the result without touching a terminal.

The clever bit is how it gets the files. It doesn't take a video tensor - it takes VHS_FILENAMES, the filename-list output from VideoHelperSuite's video combine nodes. If you've rendered with VHS, you already have that output sitting there; wire it in and the node uploads each referenced file as-is, preserving the extension (.mp4, .webm, whatever your encoder produced). That design choice means no re-encoding - it uploads the actual files, so what you trained on is byte-for-byte what you rendered. Each file gets renamed to your filename_prefix plus the original extension, which will collide on repeated runs, so keep that in mind.

Inputs and output

• filenames - a VHS_FILENAMES input, straight from VideoHelperSuite's combine-video output. This is the one that makes the node work, and it's why you need VHS installed even though the README's stated prerequisite is Kijai's comfyui-kjnodes.
• repository_id - the HF dataset repo to upload into (default nilor_dataset).
• hf_auth_token - your write-capable HF token. Default is the placeholder auth_token; don't run it as-is.
• filename_prefix - base name for the uploaded files (default nilor_save).
• No outputs - output node; the side effect is files landing in your dataset repo.

Installing it

Part of Nilor Nodes (nilor-corp/nilor-nodes):

cd ComfyUI/custom_nodes
git clone https://github.com/nilor-corp/nilor-nodes

or ComfyUI Manager → search "Nilor Nodes" → install → restart. On top of the pack, you need VideoHelperSuite installed for the VHS_FILENAMES output this node consumes - it's a separate, very popular pack you'll probably already have. huggingface_hub does the uploading and comes with the pack's dependency list (which is heavier than a couple of upload nodes warrant - boto3, fastapi, openexr and friends all install regardless). No model files.

Troubleshooting

Same auth rules as the image variant, plus one video-specific quirk. The token must be a real HF write token, the repo must already exist, and re-running the queue just adds more files with the same name - overwriting, since the name is prefix + extension with no counter or timestamp. That's the biggest difference from its image sibling and the thing to plan around: if you want a versioned dataset, give each run a distinct prefix. There's no community support for this pack, so the HfApi error is your diagnostic. And the obvious-but-worth-saying one: the node can only upload what VHS tells it about, so if your video combine didn't produce files (bad ffmpeg, zero-length render), you'll get nothing pushed and possibly a confusing empty result.