Login
You're viewing the techdon.dev public feed.
  • Jan 6, 2026, 11:30 AM

    2. For AF there's a working WIP branch for the swISP, so we can expect progress here as well (plus multiple actuators having working drivers now).
    3. We already have a working software stack for photos and videos, via Pipewire and Gstreamer. Just start Snapshot or Kamoso. (surely still missing quite a few features)
    4. Drivers for ISPs will stay out of reach on many devices - that's why we have the swISP, which is about to turn into a gpuISP by default (likely in the next libcamera release).

    ...

    💬 3🔄 0⭐ 1

Replies

  • Jan 6, 2026, 11:39 AM

    5. Hardware video encoding is already available and working on devices like the OP6. In Snapshot you can turn it on in the settings - it's only disabled by default because of stability issues/bugs (e.g. on the Pixel3a), another area where we can expect improvements in the near future.
    ...

    💬 1🔄 0⭐ 0
  • Jan 6, 2026, 11:43 AM

    6. With the gpuISP tuning steps like color correction are essentially free regarding performance. I already have a first test set of tuning matrices for the IMX363 applied on my Pixel3a (running libcamera with latest gpuISP patches) - it currently makes the result worse, but you can expect some comparison images soon where this changes :)
    Fairphone even gifted a FP4 and FP5 to the libcamera maintainers so they can start doing this kind of tuning.

    💬 0🔄 0⭐ 0
  • Jan 6, 2026, 11:47 AM

    TL;DR: the time to start playing with tuning is now - and in fact that's what we're doing already :)

    💬 1🔄 0⭐ 0
  • Jan 6, 2026, 12:09 PM

    P.S.: here's a quick video shot with Snapshot on an OP6 with pmOS edge + libcamera gpuISP, with hw-encoding enabled. There's still a lot to be desired quality-wise - most importantly AF, sensor tuning - but the basics for fullHD video capturing are increasingly in place (edit: the v4l2 encoder certainly also needs improvements).

    Image attached toot
    💬 2🔄 1⭐ 0
  • Jan 6, 2026, 12:27 PM

    @rmader the @libcamera matrix room is an exciting follow right now. I love it when you share updates, thank you. Cameras is one of the things I'm most excited for this year in mobile linux.

    💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Jan 6, 2026, 1:22 PM
    @rmader 1. Agreed. 2. Well... more work is needed, like actually merging the stuff in, right? And then we'll have to do _tuning_ of the AF algorithm. And probably someone needs to do PDAF. I tried in clicks machine, but again, more is missing than tuning. 3. "Missing few features" as in "will need complete redesign"? :-). This will get longer reply. 4. Agreed. Another missing piece. I'm eagerly waiting for gpuisp to hit the tree, but it is not there yet. 5. Ok, I was not aware of this one, it is good to see progress. (But again, more than tuning is needed). 6. Good to see that. Does gpuisp already do lens shading correction? That one is a bit important. Plus it should really output YUY2, because that's what encoder needs. Plus there's more work there, for example we'd want controls for manual focus, manual exposure; plus it needs to resize and not crop. But that's all work that should happen after it hits the tree.

    Thanks for video. I guess we can agree that it needs working AF first, and then we'll see what else is needed? :-).
    💬 0🔄 0⭐ 0
  • Jan 7, 2026, 11:42 AM

    @pavel Re 4: Looking really good now 🤞
    5. Yeah, I'm betting on a positive feedback loop here - the more stuff works, the more motivated people are to help fixing the remaining issues.
    6. AFAICS lens shading, denoising and YUV formats are high on the list for the next steps after the main gpuISP implementation has landed. And I fully agree that YUY2/YUV420 and NV12 would be great to have. Not having the later currently causes an extra conversion/copy when recording in Snapshot.

    💬 0🔄 0⭐ 0
  • Jan 6, 2026, 1:32 PM
    @rmader 3. Lets take about this one, because I don't believe it is as simple as "just start Snapshot".

    Gpuisp will need to include sensor tuning, correct? It is needed for video capture. And for still image capture, we'll need to capture full-size raw image, bypassing gpuisp, correct? That will get us bayer image.

    But now we'll need to transform that bayer image into high-quality jpg. What is the proposal to do that?

    Anyway I'm playing with libcamera-based photo taking here -- https://gitlab.com/tui/libcamera/-/commit/7c3c0d04c026d3a4a030fb7882095c028cd4ee8e -- but there's a lot of work to be done :-(.
    💬 0🔄 0⭐ 0
  • Jan 7, 2026, 11:58 AM

    @pavel Hm, why would you not use the gpuISP for taking still pictures? Sure, you can further improve quality with that in the long run - and the upcoming version of libcamera will allow you to choose between gpuISP and raw - but for starters I'd argue it's fine to just use the gpuISP.

    💬 1🔄 0⭐ 0
  • Jan 7, 2026, 12:23 PM
    @rmader Well... the thread began with "more than tuning is needed", and this is decision which will lead to redesign later. We'll likely want to use same tuning for gpuISP and then for hq-still-photo-component, and hq-still-photo-component does not exist at the moment :-).

    And "why not use gpuISP at the moment". Well, besides quality, gpuISP only has one output. We need that output for preview. And that means we have photos in BAYER.

    Being able to feed libcamera with BAYER data, and saying "Librem 5 back camera took this BAYER, please convert it to RGB in high quality" would be good.
    💬 0🔄 0⭐ 0
  • Jan 7, 2026, 1:00 PM

    @pavel Re "Well, besides quality, gpuISP only has one output. We need that output for preview" - the way how e.g. Snapshot works right now is that it uses the same buffers for preview/viewfinder as well as encoding video or still images.
    A proper multi-stream architecture will certainly be nice long-term - but I don't see why we wouldn't just implement that in the gpuISP when we get to that point (so it can have multiple outputs).

    💬 1🔄 0⭐ 0
  • Jan 7, 2026, 1:11 PM

    @pavel P.S.: encoding RGB/YUV to jpeg (or any image/video format) is something best left to the clients, IMO. Those already have plenty of options to do so, such as using #gstreamer, #ffmpeg or other libraries.

    💬 0🔄 0⭐ 0