The virtual camera button was on. A frame grab from the loopback had a face in it. Discord’s preview was a flat grey rectangle with the right name in the dropdown.

The Situation

We wanted a camera, through OBS, into Discord on a PipeWire desktop. The body was on HDMI into a cheap UVC capture dongle. Discord is the browser, not a native client. The expected path was the one every guide repeats: load v4l2loopback, click Start Virtual Camera, pick that camera in the call.

That path is real. It is also three different failures wearing one grey box.

The Button Was Not the Bug

OBS 32 did not draw Start Virtual Camera until v4l2loopback was loaded. That part of the guide is correct. After the module came up with exclusive_caps=1 on /dev/video10, the button appeared, and the kernel device announced a single capture format:

'YUYV' (YUYV 4:2:2)
    Size: Discrete 1920x1080
        Interval: Discrete 0.017s (60.000 fps)

exclusive_caps=1 is why Chromium will list the device at all. It is also why the device vanishes from the camera menu the moment nothing is writing to it. Query the format while Virtual Camera is stopped and the ioctl fails:

VIDIOC_G_FMT: failed: Invalid argument

The node is an output until a producer opens it, then it flips to capture. Discord’s dropdown is a photograph of that moment. A hard refresh while OBS is not writing will never grow an OBS-Virtual-Camera row. We spent a while looking for a hidden menu item. The menu was telling the truth.

Listed, and Still Grey

Once OBS was writing, Discord did list OBS-Virtual-Camera. The preview stayed grey. A grab from /dev/video10 at the same time was not grey. The pixels had a person in them. The browser had not failed to find the device. It had failed to start a stream, and the placeholder for “no frames” is the same colour as a dead camera.

The log line, from the PipeWire camera path Brave uses by default, was:

[pipewire] Stream state: "error" (error: no more input formats)

PipeWire 1.6.9’s V4L2 source does not convert. The loopback was offering one thing, YUY2 at whatever size OBS had last written. The browser’s PipeWire capturer asked for something else, ran out of formats, and disconnected. The node went back to suspended. No file descriptor on the browser. Grey box, correct name.

A second camera we published into the PipeWire graph, media.role=Camera, showed up in wpctl status and never showed up in Discord. Discord’s list on this browser is V4L2 card names: OBS-Virtual-Camera, and the dongle as USB Camera or USB Video because the device has an empty product string. A client-provided PipeWire source is invisible there. Restarting the browser does not invent a row for it.

The flag that makes Brave fall back to the V4L2 capturer, which can actually read YUY2, is:

--disable-features=WebRtcPipeWireCamera

It only applies to a fresh browser process. A second launch hands the URL to the running singleton, and the old process keeps the PipeWire path. Quit the browser. Then open Discord.

Ten Frames a Second Was a Format, Not a Ceiling

Full HD “only doing 10 fps” was the other wrong story, and it came from the same dongle.

lsusb said bcdUSB 2.00. The device sat on a 480 Mbit bus. That sounds like a hard cap, and for uncompressed video it is. The capture formats told a different story:

'MJPG'  1920x1080   60, 50, 30, 20, 10 fps
'YUYV'  1920x1080   10, 5 fps
'YUYV'  1280x720    25, 10 fps

A three-second read of the MJPEG mode, with nothing else holding the device, came back at the advertised rate:

Stream #0:0: Video: mjpeg, yuvj422p, 1920x1080, 60 fps
frame=  180 fps= 60  time=00:00:03.00  speed=0.992x

OBS’s PipeWire camera source did not pick that mode. It negotiated the uncompressed one:

[pipewire] Negotiated format:
    Format: 4 (Spa:Enum:VideoFormat:YUY2)
    Size: 1280x720
    Framerate: 25/1

Asking it for an encoded 1920x1080 stream produced the same error Discord hit: no more input formats. The beta source does not have a way to say MJPEG and mean it.

OBS’s own V4L2 source can say MJPEG. We set 1920x1080, pixelformat MJPG, 60 fps. The log accepted the request and then did not get frames:

v4l2-input: Resolution: 1920x1080
v4l2-input: Pixelformat: MJPG
v4l2-input: Framerate: 60.00 fps
v4l2-input: /dev/video0: select timed out

Release the device and the same mode is 60 fps again under ffmpeg. The plugin’s open path stalls this chip. The chip is fine. “10 fps at 1080p” was YUYV. “OBS cannot do 60” was the plugin. Those are different sentences, and only one of them is about the hardware.

The Dialog That Blamed the Encoder

Starting OBS with the virtual camera forced on popped a modal: “Starting the output failed. Please check the log for details.” Under it, the stock note about NVENC and AMD drivers.

NVENC was not the failure. The same log had already loaded it:

[obs-nvenc] NVENC version: 13.1 (compiled) / 13.1 (driver), AV1 supported: true

The line that matched the modal was Failed to start virtual camera. We had set the loopback control keep_format=1 so a browser could not smash the producer’s size. That control also stops the producer. OBS could not set a format on /dev/video10, so the output failed, and the UI pasted the encoder paragraph it pastes for every failed output. Clear keep_format before OBS starts. Lock it after the virtual camera is writing, if you lock it at all.

One Opener

The preview that finally moved was not the virtual camera. The dropdown said USB Video. That is the dongle. Discord’s V4L2 capturer had opened /dev/video0 itself.

OBS, still trying to be a camera source, logged:

error set output format: -16 (Device or resource busy)

fuser had both PipeWire and the browser on that node. A UVC capture device does not share. Whoever opens it second gets busy, or a grey box, depending on which side you are looking at. The working arrangement is the boring one: if Discord is on the dongle, OBS does not also take the dongle. Hide the capture source, or don’t start it. The virtual camera is a second device only while OBS is writing to the loopback. It is not a way to sneak a second opener onto the first device.

The stretch in that direct preview was in the pixels, not in the CSS. The box was about 16:9, edge to edge, no bars, and the face was wide. The device was reporting 1280x720 YUYV. A capture chip that scales a taller HDMI picture into that rectangle will do exactly this. The camera body’s movie-mode HDMI, real 16:9, is the fix. OBS is not required for it.

Receipts

Formats, from v4l2-ctl -d /dev/video0 --list-formats-ext:

Mode Size Rates the device advertised
MJPEG 1920x1080 60, 50, 30, 20, 10
YUYV 1920x1080 10, 5
YUYV 1280x720 25, 10

What each layer actually did:

Layer What it held
Raw read, MJPEG 1920x1080 180 frames in 3.0 s, 60 fps
OBS PipeWire source YUY2 1280x720 @ 25
OBS V4L2 source, MJPEG 60 select timed out, then 60 fps again once released
Discord, PipeWire camera path no more input formats, grey preview
Discord, V4L2 path, dongle picture, device busy if OBS also opens it
Virtual camera start with keep_format=1 Failed to start virtual camera

Takeaway

A grey preview with the right name is a stream that never started, not a camera that never enumerated. On this stack the browser’s PipeWire camera path will not convert a YUYV loopback, and it will not list a PipeWire source you invented beside it. Read the V4L2 device, or don’t.

And read the format list before you call a frame rate a hardware ceiling. Ten frames at 1080p was the uncompressed path. The same chip, asked for MJPEG, held 60.

What to run

# what the dongle will actually do
v4l2-ctl -d /dev/video0 --list-formats-ext

# prove the MJPEG mode, with nothing else holding the device
ffmpeg -f v4l2 -input_format mjpeg -video_size 1920x1080 -framerate 60 \
  -i /dev/video0 -t 3 -f null -

# Brave: V4L2 capturer, then quit the browser fully and reopen
# ~/.config/brave-flags.conf
# --disable-features=WebRtcPipeWireCamera

If Discord is on the dongle, OBS does not get the dongle. That is the whole sharing policy.