Screen share: top strip of the stream flickers / shows another frame #589

Open
opened 2026-11-15 05:34:45 +00:00 by charles · 12 comments
Owner

Report (Charles, 2026-11-15). Watching a screen share sent from Flaxative, a strip at the top of the stream (roughly the top fifth) flickers: sometimes black, and it looks out of step with the rest of the picture, "as if I'm seeing 1/5 of the next frame too quickly". It looks the same as the artifact seen with the Electron app. A screenshot never shows it: the screenshot always contains a complete frame.

Viewer machine. CachyOS, KDE Plasma on Wayland, NVIDIA RTX 5090 (driver 620.70), two 4K monitors at 120 Hz and 160 Hz, scale 1.45, VRR "Automatic", KWin AllowTearing=false. Sender's OS not stated.

What that suggests. With tearing off in the compositor, a strip of another frame should not come from the display. A transient mix of two frames inside one presented image points at a race on the frame's pixels somewhere between the source and the texture:

  1. Viewer: the decoder/converter writing into a buffer the UI is still reading or uploading (buffer reuse in the receive path in fluxer-voice rtc.rs, the FrameBus, the conversion to the texture format, plane-by-plane uploads in fluxer-ui).
  2. Sender: the capture path reading a buffer the compositor is still writing (PipeWire dmabuf/SHM without waiting for the buffer to be complete, or reusing a buffer handed back too early), which would put the tear into the encoded stream for every viewer — as with the Electron app.
  3. Hardware decode (#366) producing partially decoded pictures for H.264.

To do — measure first.

  • A deterministic test: the sender publishes frames in which every row carries the frame's counter (for example a solid colour or stripe pattern derived from the frame number); the viewer checks every frame that reaches the frame bus, and every frame the widget uploads, for internal consistency (all rows from one frame, no black strip). Run it per codec, with and without the NVIDIA decoder, at a high frame rate and under CPU load.
  • The same consistency check on the capture side: frames coming out of fluxer-capture from a source that changes every refresh.
  • A debug switch that counts inconsistent/black-top frames at runtime and logs them, so it can be confirmed on the real machines.
**Report (Charles, 2026-11-15).** Watching a screen share sent from Flaxative, a strip at the top of the stream (roughly the top fifth) flickers: sometimes black, and it looks out of step with the rest of the picture, "as if I'm seeing 1/5 of the next frame too quickly". It looks the same as the artifact seen with the Electron app. A screenshot never shows it: the screenshot always contains a complete frame. **Viewer machine.** CachyOS, KDE Plasma on Wayland, NVIDIA RTX 5090 (driver 620.70), two 4K monitors at 120 Hz and 160 Hz, scale 1.45, VRR "Automatic", KWin `AllowTearing=false`. Sender's OS not stated. **What that suggests.** With tearing off in the compositor, a strip of another frame should not come from the display. A transient mix of two frames inside one presented image points at a race on the frame's pixels somewhere between the source and the texture: 1. Viewer: the decoder/converter writing into a buffer the UI is still reading or uploading (buffer reuse in the receive path in `fluxer-voice` `rtc.rs`, the `FrameBus`, the conversion to the texture format, plane-by-plane uploads in `fluxer-ui`). 2. Sender: the capture path reading a buffer the compositor is still writing (PipeWire dmabuf/SHM without waiting for the buffer to be complete, or reusing a buffer handed back too early), which would put the tear into the encoded stream for every viewer — as with the Electron app. 3. Hardware decode (#366) producing partially decoded pictures for H.264. **To do — measure first.** - A deterministic test: the sender publishes frames in which every row carries the frame's counter (for example a solid colour or stripe pattern derived from the frame number); the viewer checks every frame that reaches the frame bus, and every frame the widget uploads, for internal consistency (all rows from one frame, no black strip). Run it per codec, with and without the NVIDIA decoder, at a high frame rate and under CPU load. - The same consistency check on the capture side: frames coming out of `fluxer-capture` from a source that changes every refresh. - A debug switch that counts inconsistent/black-top frames at runtime and logs them, so it can be confirmed on the real machines.
Author
Owner

Findings (measured), PR #607

Cause found on the sender side (Linux/Wayland), hypothesis 2. libwebrtc's PipeWire capturer hands us a share of one of the two frames in its stream queue. Its PipeWire thread writes a new picture into that same frame two buffers later, even while we are still reading it: it only logs a debug warning, "Overwriting frame that is still shared" (shared_screencast_stream.cc at the pinned webrtc-a545de4, l. 698 and l. 915–917). fluxer-capture converted that frame to I420 in place (crates/fluxer-capture/src/desktop/rtc.rs, the start_capture callback). A 4K frame takes long enough to convert that the writer sometimes overtakes the read position, so the frame comes out with its top part from one picture and the rest from the picture two buffers later. That frame is then encoded, so every viewer sees it (Chromium/Electron uses the same capturer). It lasts one frame at a time, which is why screenshots miss it.

How it was measured. tools/capture-consistency/run.sh runs headless: a private D-Bus session, a private PipeWire daemon, a stand-in ScreenCast portal, and a source whose frame n is one solid colour. The real Wayland capture path runs end to end and every frame is checked. Over 20 s per case, before the fix:

  • 4K→1080p@30: 6 of 601 frames mixed
  • 4K→1080p@60: 13 of 1164
  • 4K→4K@60: 20 of 973
  • 4K→1080p@60 with 6 busy cores: 36 of 1088
  • 1080p sources: 0

In every mixed frame the two colours were two frames apart, which is the signature of the 2-frame queue. After the fix, all 7 configurations had 0 mixed frames out of 6,304.

Receive side is clean. Every received frame was checked, for AV1, H.264, VP9 and VP8, at 1080p30, 1080p60 and 1080p→720p mid-stream, with NVDEC and with LK_DISABLE_NVDEC=1, with and without CPU load. That is 56 runs, about 18,900 frames, with 0 mixed and 0 black top strips. The FrameBus and the widget upload were audited and are safe: frames are immutable Arc planes, and all three planes are written with write_texture before the draw. So hypotheses 1 and 3 are not the cause, and nothing needs to change for H.264 decoding for this issue.

Fix. The capture callback now copies the pixels, checks the copy against the frame, pulls again if the frame changed (at most 3 retries) and converts the copy. The cost is one extra copy and compare per frame. At 4K60 on this loaded machine that meant about 23 % fewer frames delivered.

Not reproduced: the black strip. No configuration produced one. The headless source uses MemFd buffers. KWin normally uses DMA-BUF, and libwebrtc reads those back without any explicit sync (it negotiates no SPA_META_SyncTimeline). That could explain the black strip on drivers without implicit sync, but it can't be tested without a compositor. Also not covered: a Windows sender (audited only: WGC fills its frames on our capture thread, so this race can't happen there) and the sender's actual OS, which isn't stated here.

To confirm on the real machines, run Flaxative with FLUXER_FRAME_CHECK=1 on both the sender and the viewer. Each frame whose top fifth is black while the rest is not gets a warn log line: side = "captured" on the sender, before encoding, and side = "received" on the viewer, after decoding. Totals are logged every 600 frames.

**Findings (measured), PR #607** **Cause found on the sender side (Linux/Wayland), hypothesis 2.** libwebrtc's PipeWire capturer hands us a share of one of the two frames in its stream queue. Its PipeWire thread writes a new picture into that same frame two buffers later, even while we are still reading it: it only logs a debug warning, "Overwriting frame that is still shared" (`shared_screencast_stream.cc` at the pinned `webrtc-a545de4`, l. 698 and l. 915–917). `fluxer-capture` converted that frame to I420 in place (`crates/fluxer-capture/src/desktop/rtc.rs`, the `start_capture` callback). A 4K frame takes long enough to convert that the writer sometimes overtakes the read position, so the frame comes out with its top part from one picture and the rest from the picture two buffers later. That frame is then encoded, so every viewer sees it (Chromium/Electron uses the same capturer). It lasts one frame at a time, which is why screenshots miss it. **How it was measured.** `tools/capture-consistency/run.sh` runs headless: a private D-Bus session, a private PipeWire daemon, a stand-in ScreenCast portal, and a source whose frame n is one solid colour. The real Wayland capture path runs end to end and every frame is checked. Over 20 s per case, before the fix: - 4K→1080p@30: 6 of 601 frames mixed - 4K→1080p@60: 13 of 1164 - 4K→4K@60: 20 of 973 - 4K→1080p@60 with 6 busy cores: 36 of 1088 - 1080p sources: 0 In every mixed frame the two colours were two frames apart, which is the signature of the 2-frame queue. After the fix, all 7 configurations had 0 mixed frames out of 6,304. **Receive side is clean.** Every received frame was checked, for AV1, H.264, VP9 and VP8, at 1080p30, 1080p60 and 1080p→720p mid-stream, with NVDEC and with `LK_DISABLE_NVDEC=1`, with and without CPU load. That is 56 runs, about 18,900 frames, with 0 mixed and 0 black top strips. The FrameBus and the widget upload were audited and are safe: frames are immutable `Arc` planes, and all three planes are written with `write_texture` before the draw. So hypotheses 1 and 3 are not the cause, and nothing needs to change for H.264 decoding for this issue. **Fix.** The capture callback now copies the pixels, checks the copy against the frame, pulls again if the frame changed (at most 3 retries) and converts the copy. The cost is one extra copy and compare per frame. At 4K60 on this loaded machine that meant about 23 % fewer frames delivered. **Not reproduced: the black strip.** No configuration produced one. The headless source uses MemFd buffers. KWin normally uses DMA-BUF, and libwebrtc reads those back without any explicit sync (it negotiates no `SPA_META_SyncTimeline`). That could explain the black strip on drivers without implicit sync, but it can't be tested without a compositor. Also not covered: a Windows sender (audited only: WGC fills its frames on our capture thread, so this race can't happen there) and the sender's actual OS, which isn't stated here. **To confirm on the real machines,** run Flaxative with `FLUXER_FRAME_CHECK=1` on both the sender and the viewer. Each frame whose top fifth is black while the rest is not gets a `warn` log line: `side = "captured"` on the sender, before encoding, and `side = "received"` on the viewer, after decoding. Totals are logged every 600 frames.
Author
Owner

Reopened 2026-11-16. Charles on nightly 2026.1116.341 (the capture fix of PR #607 has been in every nightly since 2026.1115.307): "The black bar still shows up during the streams". So #607 did not fix what he sees, or fixed only one of several causes. It was never confirmed in real use.

Not yet known, and needed before another fix: who was streaming and with which client (Flaxative, the Electron app or the web app), who saw the bar (the streamer in their own preview, or the viewers) and in which client, and whether it is at the top edge only. His log of that session shows him sharing from Flaxative on Linux (SystemPicker, 1920x1046 then 1920x1080, 60 fps, H264, hardware=true, encoder Auto) and no stream-watching lines, so the bar was presumably seen by viewers or in his own preview.

If the streamer was Flaxative with the hardware encoder, the capture copy fixed in #607 is no longer the only suspect: the hardware encode path (frame upload / surface reuse before the encoder has read it) and the viewer-side decode and upload are candidates. Use FLUXER_FRAME_CHECK=1 and a recorded stream to locate the stage.

**Reopened 2026-11-16.** Charles on nightly 2026.1116.341 (the capture fix of PR #607 has been in every nightly since 2026.1115.307): "The black bar still shows up during the streams". So #607 did not fix what he sees, or fixed only one of several causes. It was never confirmed in real use. Not yet known, and needed before another fix: who was streaming and with which client (Flaxative, the Electron app or the web app), who saw the bar (the streamer in their own preview, or the viewers) and in which client, and whether it is at the top edge only. His log of that session shows him sharing from Flaxative on Linux (SystemPicker, 1920x1046 then 1920x1080, 60 fps, H264, hardware=true, encoder Auto) and no stream-watching lines, so the bar was presumably seen by viewers or in his own preview. If the streamer was Flaxative with the hardware encoder, the capture copy fixed in #607 is no longer the only suspect: the hardware encode path (frame upload / surface reuse before the encoder has read it) and the viewer-side decode and upload are candidates. Use FLUXER_FRAME_CHECK=1 and a recorded stream to locate the stage.
charles reopened this issue 2026-11-16 06:34:26 +00:00
Author
Owner

2026-11-16, more from Charles: every viewer sees the bar on his shares (Flaxative on Linux, H264, hardware encoder). A stream he watched from someone else's machine, in Flaxative, showed no bar. So the viewer side of Flaxative displays a clean stream cleanly, and the defect is tied to the sending side on his machine (capture, conversion or hardware encode), possibly intermittent. Diagnosis continues; no fix has shipped since #607.

2026-11-16, more from Charles: every viewer sees the bar on **his** shares (Flaxative on Linux, H264, hardware encoder). A stream he **watched** from someone else's machine, in Flaxative, showed no bar. So the viewer side of Flaxative displays a clean stream cleanly, and the defect is tied to the sending side on his machine (capture, conversion or hardware encode), possibly intermittent. Diagnosis continues; no fix has shipped since #607.
Author
Owner

PR #763 is merged. This time the defect was measured and measured gone, in a headless setup that copies the reporter's sender (DMA-BUF capture from an NVIDIA GPU, 1080p and 1046p at 60 fps, H.264 6 Mbit/s):

  • Cause: libwebrtc's PipeWire capturer writes the next picture into a buffer it already handed over; #607's copy-and-compare misses a writer that stalls mid-buffer (its glReadPixels being descheduled under CPU pressure). The frame handed on has the top 6–901 rows of a picture 2–5 frames later. It happens before conversion and encoding, only with DMA-BUF capture and a busy CPU.
  • Before the fix, under CPU pressure: 6 mixed frames of 7118 converted (6 of 6436 received by a viewer) in 120 s; 7 in 300 s with all cores loaded. After: 0 of 6676 and 0 of 7281 under the same pressure, 0 of 21,561 without load.
  • FLUXER_FRAME_CHECK=1 now also checks the frame handed to the encoder and logs per-pull decisions; FLUXER_CAPTURE_VERIFY_LATEST=0 turns the fix off for an A/B test.

Kept open until the reporter confirms in a real share on his KWin session, which was not driven here.

PR #763 is merged. This time the defect was **measured** and measured gone, in a headless setup that copies the reporter's sender (DMA-BUF capture from an NVIDIA GPU, 1080p and 1046p at 60 fps, H.264 6 Mbit/s): - Cause: libwebrtc's PipeWire capturer writes the next picture into a buffer it already handed over; #607's copy-and-compare misses a writer that stalls mid-buffer (its `glReadPixels` being descheduled under CPU pressure). The frame handed on has the top 6–901 rows of a picture 2–5 frames later. It happens before conversion and encoding, only with DMA-BUF capture and a busy CPU. - Before the fix, under CPU pressure: 6 mixed frames of 7118 converted (6 of 6436 received by a viewer) in 120 s; 7 in 300 s with all cores loaded. After: 0 of 6676 and 0 of 7281 under the same pressure, 0 of 21,561 without load. - `FLUXER_FRAME_CHECK=1` now also checks the frame handed to the encoder and logs per-pull decisions; `FLUXER_CAPTURE_VERIFY_LATEST=0` turns the fix off for an A/B test. **Kept open** until the reporter confirms in a real share on his KWin session, which was not driven here.
Author
Owner

The fix of PR #763 is in nightly 2026.1116.350. Open until the reporter confirms on a real share.

The fix of PR #763 is in nightly 2026.1116.350. Open until the reporter confirms on a real share.
Author
Owner

2026-11-17, from Charles: the share where the top of the frame was still scrambled (seen by him as a viewer, after nightly 350) came from another user, not from his machine. Not yet known: that sharer's client (Flaxative, Electron, web), version and OS.

Why it matters: the fix of PR #763 is only in Flaxative nightlies from 2026.1116.350, and only on the Linux Wayland-portal capture path (X11 and Windows skip the second look). If the sharer was on Flaxative for Windows, or on an older nightly, or on the Electron/web client (which has this artefact by itself), the fix was never in play. Charles's own shares from Linux have not been tried since the fix.

2026-11-17, from Charles: the share where the top of the frame was still scrambled (seen by him as a viewer, after nightly 350) came from **another user**, not from his machine. Not yet known: that sharer's client (Flaxative, Electron, web), version and OS. Why it matters: the fix of PR #763 is only in Flaxative nightlies from 2026.1116.350, and only on the Linux Wayland-portal capture path (X11 and Windows skip the second look). If the sharer was on Flaxative for Windows, or on an older nightly, or on the Electron/web client (which has this artefact by itself), the fix was never in play. Charles's own shares from Linux have not been tried since the fix.
Author
Owner

2026-11-18, from Charles: the sharer of the still-scrambled stream was on Linux, Flaxative around nightly 353, so they had the fix of PR #763 (in from 350). So the fix did not cover their setup. What differs from the measured setup is not yet known. The fix's second look runs only on the Wayland portal capture path (X11 skips it), and the measured case was DMA-BUF from an NVIDIA GPU; another GPU, an X11 session, or the SHM path under a different compositor could behave differently.

Needed from the sharer:

  1. Desktop and session type (echo $XDG_CURRENT_DESKTOP $XDG_SESSION_TYPE), GPU and driver (glxinfo -B | head or lspci | grep -i vga).
  2. A share of one or two minutes started with diagnostics on, then the log file:
    FLUXER_FRAME_CHECK=1 FLUXER_LOG=info,fluxer_capture=debug ./Flaxative-*.AppImage
    
    Log: ~/.local/state/flaxative/logs/flaxative.<date>.log. The lines of interest: the PipeWire format (a Modifier: line means DMA-BUF), the per-pull trace (trusted / torn / dropped), the torn-frame checker's findings as captured and as handed to the encoder, and voice: screen share encoder implementation=.
  3. The same share once more with FLUXER_CAPTURE_VERIFY_LATEST=0 added, to see whether the fix changes anything for them.
2026-11-18, from Charles: the sharer of the still-scrambled stream was on **Linux, Flaxative around nightly 353**, so they had the fix of PR #763 (in from 350). So the fix did not cover their setup. What differs from the measured setup is not yet known. The fix's second look runs only on the Wayland portal capture path (X11 skips it), and the measured case was DMA-BUF from an NVIDIA GPU; another GPU, an X11 session, or the SHM path under a different compositor could behave differently. Needed from the sharer: 1. Desktop and session type (`echo $XDG_CURRENT_DESKTOP $XDG_SESSION_TYPE`), GPU and driver (`glxinfo -B | head` or `lspci | grep -i vga`). 2. A share of one or two minutes started with diagnostics on, then the log file: ``` FLUXER_FRAME_CHECK=1 FLUXER_LOG=info,fluxer_capture=debug ./Flaxative-*.AppImage ``` Log: `~/.local/state/flaxative/logs/flaxative.<date>.log`. The lines of interest: the PipeWire format (a `Modifier:` line means DMA-BUF), the per-pull trace (trusted / torn / dropped), the torn-frame checker's findings as captured and as handed to the encoder, and `voice: screen share encoder implementation=`. 3. The same share once more with `FLUXER_CAPTURE_VERIFY_LATEST=0` added, to see whether the fix changes anything for them.
Author
Owner

2026-11-18: Charles is watching a share from a Windows user in Flaxative: the top of the frame is not scrambled so far (client and version of the sharer not yet known). So no sign of the artefact on the Windows capture path in this session; the open case remains the Linux sharer on ~nightly 353.

2026-11-18: Charles is watching a share from a Windows user in Flaxative: the top of the frame is not scrambled so far (client and version of the sharer not yet known). So no sign of the artefact on the Windows capture path in this session; the open case remains the Linux sharer on ~nightly 353.
Author
Owner

2026-11-18: the sharer of #876 (Linux Wayland, RTX 4080, 3440×1440 monitor, nightly 2026.1118.373 per the owner) still sees the top strip, so #763 is present and does not cover their case.

Tried at 3440×1440 on textchat head (PR #879's runs): run.sh --chain, MemFd and DMA-BUF (tiled modifier, stride 13760, no row padding), NVENC and OpenH264, at 1080p, 1440p and Source, 20–55 s each, FLUXER_FRAME_CHECK=1, no synthetic load. Compared with 1920×1080 and 2560×1440.

Result: not reproduced.

  • libwebrtc hands over torn copies at 3440×1440: 17–27 per 20 s on DMA-BUF, against 5–6 at 1080p and 1440p, always the picture two frames later on top.
  • #763's compare and second look dropped every one of them. Converted, encoder input and received frames had 0 mixed and 0 black-top in every run, MemFd had 0 at all stages, and the non-multiple-of-16 heights (802, 1070) encoded cleanly.
  • Not covered: the real KWin session, a 4080 driver, and load on the sharer's machine.
  • Side finding: in these debug-build runs the capture thread saturates at 3440×1440 (21–35 converted fps of 60).

Needed from the sharer: a share with FLUXER_FRAME_CHECK=1 FLUXER_LOG=info,fluxer_capture=debug, specifically:

  • the PipeWire format line (Modifier:);
  • the per-pull trace (trusted / torn / moved on);
  • frame-check warnings with side = \"captured\" and at the encoder input.

If those stay at 0 while viewers see the strip, the cause is after capture.

2026-11-18: the sharer of #876 (Linux Wayland, RTX 4080, 3440×1440 monitor, nightly 2026.1118.373 per the owner) still sees the top strip, so #763 is present and does not cover their case. **Tried at 3440×1440 on textchat head (PR #879's runs):** `run.sh --chain`, MemFd and DMA-BUF (tiled modifier, stride 13760, no row padding), NVENC and OpenH264, at 1080p, 1440p and Source, 20–55 s each, `FLUXER_FRAME_CHECK=1`, no synthetic load. Compared with 1920×1080 and 2560×1440. **Result: not reproduced.** - libwebrtc hands over torn copies at 3440×1440: 17–27 per 20 s on DMA-BUF, against 5–6 at 1080p and 1440p, always the picture two frames later on top. - #763's compare and second look dropped every one of them. Converted, encoder input and received frames had 0 mixed and 0 black-top in every run, MemFd had 0 at all stages, and the non-multiple-of-16 heights (802, 1070) encoded cleanly. - Not covered: the real KWin session, a 4080 driver, and load on the sharer's machine. - Side finding: in these debug-build runs the capture thread saturates at 3440×1440 (21–35 converted fps of 60). **Needed from the sharer:** a share with `FLUXER_FRAME_CHECK=1 FLUXER_LOG=info,fluxer_capture=debug`, specifically: - the PipeWire format line (`Modifier:`); - the per-pull trace (trusted / torn / moved on); - frame-check warnings with `side = \"captured\"` and at the encoder input. If those stay at 0 while viewers see the strip, the cause is after capture.
Author
Owner

Charles captured the artefact in a screenshot on 2026-11-19 (not attached: it shows another member's screen).

Measured on the screenshot (1460x1254, scaled from the stream view): the stream shows a browser window. Rows 45-67 are solid black across the full width (about 22 rows, 1.8% of the height, starting 3.6% from the top), directly under the browser's toolbar and cutting the page's own header line in half. Above it the toolbar is drawn normally; below it the picture is normal. So the band is not at row 0 and it is black, not stale picture.

At 1440 source rows that is roughly rows 52-77, at 1080 roughly 39-58: about one to two 16-pixel block rows. That makes a lost or wrongly coded slice / block row in the encoder or decoder a candidate alongside the capture path. Still needed: who was sending (OS, GPU, screen or window share, chosen resolution), which client the screenshot was taken on, and whether the band flickers or stays.

Charles captured the artefact in a screenshot on 2026-11-19 (not attached: it shows another member's screen). Measured on the screenshot (1460x1254, scaled from the stream view): the stream shows a browser window. Rows 45-67 are solid black across the full width (about 22 rows, 1.8% of the height, starting 3.6% from the top), directly under the browser's toolbar and cutting the page's own header line in half. Above it the toolbar is drawn normally; below it the picture is normal. So the band is not at row 0 and it is black, not stale picture. At 1440 source rows that is roughly rows 52-77, at 1080 roughly 39-58: about one to two 16-pixel block rows. That makes a lost or wrongly coded slice / block row in the encoder or decoder a candidate alongside the capture path. Still needed: who was sending (OS, GPU, screen or window share, chosen resolution), which client the screenshot was taken on, and whether the band flickers or stays.
Author
Owner

Charles (2026-11-19): it does not happen when the sender is a Windows client, so this is a source-side problem on Linux senders, not the viewers. An agent is now investigating the Linux sending path (capture, fit/convert, encoder) with these facts.

Charles (2026-11-19): it does not happen when the sender is a Windows client, so this is a source-side problem on Linux senders, not the viewers. An agent is now investigating the Linux sending path (capture, fit/convert, encoder) with these facts.
Author
Owner

The fix (#945) is in nightly 2026.1119.406: on Linux with the NVIDIA driver, screen capture now uses memory buffers instead of dmabuf. Not yet confirmed on an affected sender. To confirm, the Linux/NVIDIA person who shares runs FLUXER_FRAME_CHECK=1 FLUXER_LOG=info,fluxer_capture=debug flaxative, shares for a few minutes and checks the log: it should show buffers="MemFd" nvidia_driver=true and no black-band warning. Stays open until then.

The fix (#945) is in nightly 2026.1119.406: on Linux with the NVIDIA driver, screen capture now uses memory buffers instead of dmabuf. Not yet confirmed on an affected sender. To confirm, the Linux/NVIDIA person who shares runs `FLUXER_FRAME_CHECK=1 FLUXER_LOG=info,fluxer_capture=debug flaxative`, shares for a few minutes and checks the log: it should show `buffers="MemFd" nvidia_driver=true` and no black-band warning. Stays open until then.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
charles/fluxer-native#589
No description provided.