Roop Face Swap (Video)

Swap a face into a video file instead of a still image. That's the whole pitch, and it's the node that usually gets people to install this pack in the first place. Give it a source face and a path to an mp4, and it runs the swap across every frame, preserving the original FPS and frame count, then writes a new video file.

The honest headline: it's slow. The README says it plainly - video swapping works best with shorter clips. Every frame goes through the full detection-and-swap pipeline as its own subprocess against roop, so a 10-second clip at 30fps is 300 separate swaps. A short clip is a great demo. A feature film is a long lunch, or several.

How it works

Same wrapper mechanism as the rest of the pack: the node saves your source tensor and shells out to roop's run.py, this time with video flags:

python run.py -s source.png -t /path/to/video.mp4 -o swapped_video.mp4 \
  --keep-fps --keep-frames --execution-provider cuda --frame-processor face_swapper

--keep-fps and --keep-frames do what they say - your output keeps the original frame rate and every frame, so you don't get a weirdly sped-up or decimated result.

The inputs that matter

• source_image - the face to impose on the video.
• target_video_path - the video file, as a string. Absolute path, and the node does not check it exists up front; roop errors out inside the subprocess if it's wrong.
• output_name - filename of the result. It's written to roop_dir/temp_io, with auto-incrementing (_1, _2…) so you never clobber a previous run.
• many_faces - swap every face that appears in the frame.
• roop_dir - your roop checkout. Default /content/roop; update it or nothing runs.

The catch: nothing comes out of the node

Read that carefully: this node has no output. RETURN_TYPES is empty. It runs, prints progress to the console, and the finished video lands at roop_dir/temp_io/<output_name>. There's no preview node to attach, no tensor to wire anywhere - you have to go find the file. That's a weird experience if you're used to ComfyUI handing you everything as a node connection.

It's also the reason the webhook nodes exist. The natural workflow for this pack is: swap the video on a Colab or remote box, then use RoopSendWebhookFile to push roop_dir/temp_io/<output_name> to your own server or an automation endpoint. On a local machine, just open the folder.

Gotchas worth knowing

• NSFW filter still applies per-frame. If roop decides a frame is unspeakable, that frame gets skipped, and you can end up with a flicker or a missing swap in the middle of your clip. Check the console for [Warning] Roop did not produce video output if the file never appears at all.
• CUDA is mandatory. The node hard-codes --execution-provider cuda. CPU-only machines can't run it, and on a shared T4 it will be slow - this is the node where you notice the 128x128 quality ceiling and the per-frame cost at the same time.
• First run still downloads the model. roop fetches inswapper_128.onnx the first time you run any swap. Let it finish before assuming a hang.
• Output quality is 128x128-flavored. Fine for short social clips; don't expect broadcast-grade results.

Install is the same as the rest of the pack: clone glitchinthemetrix16/ComfyUI-Roop into custom_nodes, restart, and make sure roop_dir points at a real roop checkout (Furkan Gözükara's fork per the README).