Skip to content

One of three mp4 renders drew a wrong material from ~29 s; stills of the same frames are bit-identical #95

Open

Description

@jbilakhi

Version: raea 0.1.0 (macOS arm64 release tarball), 18 GB Apple Silicon Mac.

What happened

Three full mp4 renders of the same 45 s 3D composition (1920x1080, 30 fps, 1,350 frames: Bevy scene, <Instances> rain tiles, StandardMaterial boats and kerb, VolumetricFog, DepthOfField, a baked raea physics bake result). Between two of the renders only audio gains changed, nothing visual.

One of the three mp4 renders drew a large surface in the wrong (red) material from about 29 s to the end. Against PNG stills of the same frames (raea render edit.tsx --at <frames> -o dir/), that render scored PSNR 15-20 dB on 9 of 32 sampled frames (frames 870-1218), which is visibly different content. The other two mp4 renders match the PNG stills at 30-43 dB on all 32 frames (the ~30 dB frames are dense rain under H.264).

What is deterministic

  • The still path (--at) was bit-identical across two runs of 32 frames (md5).
  • raea physics bake re-bakes were sha256-identical.

So the scene and physics are deterministic; the fast mp4 path produced a different picture once.

Expected

The same composition renders the same picture every time, mp4 or stills.

I don't have a minimal repro yet (it happened once in three runs). Happy to run a diagnostic build or flag if that helps.

Activity

  1. shinyobjectz commented on Nov 12, 2026

    @shinyobjectz
    Owner

    Thanks for the careful measurements; the bit-identical stills and bakes against one differing mp4 are a strong pointer. An mp4 draws its 3D live (each frame sends what changed since the last one), while --at stills build each frame from scratch, so a once-in-three difference points at the live path keeping a stale value for a part, not at the scene. Two bugs in that path were fixed this week: <Instances> whose placements changed were frozen in a video (#85), and a mesh's emissive colour was held at its first frame's (#106). A "large surface in the wrong (red) material from 29 s on" sounds like the same family, a material change that one run's fast path missed.

    Two things would let me pin it: the composition (or a trimmed copy that keeps the surface that went red and whatever changes its material around 29 s), and the output of raea traces show last after a render. If you can share the project privately, hello@raea.io reaches me. I'll keep this open.

  2. added
    waiting on reporterCan't be reproduced or fixed without more information from whoever filed it
    on Nov 12, 2026
  3. added a commit that references this issue on Nov 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    waiting on reporterCan't be reproduced or fixed without more information from whoever filed it

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions