Vfx Openexr

Vfx Openexr

Context: $ARGUMENTS

Quick start

• Dump header attributes: -> Step 1 (exr.py header)
• Quick info summary: -> Step 2 (exr.py info)
• Convert scanline -> tiled (for MIP-map access): -> Step 3 (exr.py tiled)
• Latlong -> cubemap or cube -> latlong: -> Step 4 (exr.py envmap)
• Embed a small preview thumbnail: -> Step 5 (exr.py preview)
• Multi-part (EXR 2) pack/unpack: -> Step 6 (exr.py multipart)
• Stereo multi-view pack/unpack: -> Step 7 (exr.py multiview)
• Standard attributes read/write: -> Step 8 (exr.py stdattr)
• Validate: -> Step 9 (exr.py check)
• Convert to ACES 2065-1: -> Step 10 (exr.py to-aces)
• Deep-ID manifest / metrics: -> Step 11 (exr.py manifest / metrics)

When to use

• User asks about EXR files in a VFX / compositing / lookdev context.
• Need to change EXR layout (scanline vs tiled, single-part vs multi-part).
• Deep compositing workflow (deepscanline / deeptile images).
• Preparing HDR source for a renderer's texture cache (exrmaketiled).
• ACES-ODT / ACES 2065-1 deliverables.

Step 1 — Header dump

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py header image.exr

Expands to exrheader image.exr. Shows every header attribute: dataWindow, displayWindow, chromaticities, channels, compression, pixelAspectRatio, screenWindowCenter, screenWindowWidth.

Step 2 — Concise info

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py info image.exr

Expands to exrinfo image.exr. Shorter than header; single-line-per-part summary.

Step 3 — Make tiled (for MIP-mapped texture cache)

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py tiled --in env.exr --out env_tiled.exr

Expands to exrmaketiled env.exr env_tiled.exr. Default tile size 64x64. Renderers like RenderMan/Arnold/V-Ray read tiled .exr far faster than scanline. Pair with maketx from OIIO for .tx output.

Options: --tilesize 128, --rip (RIP-map instead of MIP-map), --plevel N (levels).

Step 4 — Environment maps

Convert latlong equirectangular to a cube (6 images stacked vertically):

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py envmap --in latlong.exr --out cube.exr --type cube

Expands to exrenvmap -latlong latlong.exr -cube cube.exr.

Convert cube -> latlong:

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py envmap --in cube.exr --out latlong.exr --type latlong

Requires input already in cube layout (6 stacked squares).

Step 5 — Embed a preview

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py preview --in big.exr --out with_preview.exr --size 200

Expands to exrmakepreview -w 200 big.exr with_preview.exr. The small preview is embedded in the header so file browsers can thumbnail without decoding the full float pixels.

Step 6 — Multi-part (EXR 2+)

Combine N single-part EXRs into one multi-part file:

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py multipart pack \
  --out combined.exr beauty.exr normal.exr position.exr

Expands to exrmultipart -combine -i beauty.exr normal.exr position.exr -o combined.exr.

Split a multi-part back into per-part files:

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py multipart unpack --in combined.exr --dest ./parts/

Expands to exrmultipart -separate -i combined.exr -o ./parts/. Per-part prefix defaults to the part name.

Step 7 — Multi-view (stereo)

Combine left+right eye renders into one multi-view EXR:

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py multiview pack \
  --out stereo.exr --left L.exr --right R.exr

Expands to exrmultiview -combine -i L.exr R.exr -v left right -o stereo.exr.

Split:

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py multiview unpack --in stereo.exr --dest ./eyes/

Step 8 — Standard attributes

Read every standard attribute:

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py stdattr get image.exr

Expands to exrstdattr image.exr.

Write an attribute (example: set the owner):

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py stdattr set --in image.exr --out tagged.exr \
  --attr owner --value "Lookdev Dept"

Expands to exrstdattr -owner "Lookdev Dept" image.exr tagged.exr. Standard attributes include: owner, comments, capDate, utcOffset, longitude, latitude, altitude, focus, expTime, aperture, isoSpeed, envmap, chromaticities, whiteLuminance, adoptedNeutral, renderingTransform, lookModTransform, xDensity, wrapModes, keyCode, timeCode, framesPerSecond. Full reference in references/standard-attributes.md.

Step 9 — Validate

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py check image.exr

Expands to exrcheck image.exr. Structural validation: header integrity, chunk table, pixel-data presence. Non-zero exit on any corruption.

Step 10 — Convert to ACES

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py to-aces image.exr aces.exr

Expands to exr2aces image.exr aces.exr. Sets chromaticities to AP0 (ACES 2065-1) and adds SMPTE ST 2065-4 ACES-compliant attributes.

Step 11 — Deep ID manifest + metrics

Read the deep-ID manifest (cryptomatte-adjacent feature):

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py manifest deep.exr

Expands to exrmanifest deep.exr.

Compression size/speed stats:

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py metrics image.exr

Expands to exrmetrics image.exr.

Gotchas

• 11 compression types, pick right: NONE, RLE, ZIPS (per-scanline zip), ZIP (16-scanline block zip), PIZ (wavelet, best lossless for grainy film), PXR24 (lossy 24-bit float to 24-bit), B44 / B44A (fixed-size lossy, fast random access), DWAA / DWAB (DCT-based lossy, best size/quality), HTJ2K (EXR 3.2+ JPEG-2000-like). Read references/compression.md for the full tradeoff table.
• EXR 2.x adds multi-part + deep data. EXR 3.x reimplements the library (OpenEXRCore C API, thread-safe) and split Imath out as a separate lib. Pin your IO lib version if you need bit-exact compat.
• Deep images have two flavors: deepscanline (unordered samples) and deeptile (tiled). They require different APIs and most renderers write the scanline flavor.
• PIZ is lossless; DWAA/DWAB are not. Users sometimes assume DWA is lossless because it preserves floating-point precision — it doesn't. DWAB compresses harder than DWAA.
• chromaticities attribute is where "what primaries does this EXR use?" is stored. Missing chromaticities = assume Rec.709. ACEScg primaries look like red=0.713,0.293; green=0.165,0.830; blue=0.128,0.044; white=0.3127,0.329. ACES 2065-1 (AP0) is different.
• screenWindowCenter / screenWindowWidth are for CG origin, not for cropping. Don't set these to change the display window.
• displayWindow vs dataWindow: displayWindow is the intended canvas; dataWindow is the rendered area (may extend beyond displayWindow for overscan, or be smaller for partial renders). Comp apps crop to dataWindow on read.
• Tiled EXR random-access != random-access guaranteed. Tile size matters for how many tiles a renderer reads per sample. 64x64 is a good default; 128x128 for large-area sampling.
• Environment maps (exrenvmap) expect specific layouts. Latlong: full 2:1 equirect with rows = latitude. Cube: 6 faces stacked vertically as +X,-X,+Y,-Y,+Z,-Z (1:6 aspect), sized 6*N rows x N cols.
• exrcheck does NOT verify perceptual quality. It confirms the file is well-formed; it can't tell you the pixels are correct.
• exr2aces only rewrites the header chromaticities. It does NOT re-transform pixels. If pixels are already ACEScg but you mark as ACES2065-1, you've mislabelled color — use OIIO / OCIO to actually transform pixel values.
• Canonical tools list verified from openexr.com/en/latest/tools.html: exrheader, exrinfo, exrmaketiled, exrenvmap, exrmakepreview, exrmultipart, exrmultiview, exrstdattr, exrcheck, exr2aces, exrmanifest, exrmetrics. No exrconv, exrconvert, or exr2png — those are NOT real tools.

Examples

Example 1: "Tile a big latlong IBL for Arnold texture cache"

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py tiled --in studio.exr --out studio_tiled.exr --tilesize 128

Example 2: "Combine my AOV passes into one multi-part EXR"

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py multipart pack --out shot.exr \
  beauty.exr diffuse.exr specular.exr normal.exr position.exr

Example 3: "Check this deep render didn't get corrupted"

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py check deep.exr

Example 4: "Convert this latlong HDRI to a cubemap"

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py envmap --in sky.exr --out sky_cube.exr --type cube

Example 5: "Flag this file as ACES 2065-1 for studio ingest"

Only after confirming pixels are actually in AP0:

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py to-aces render.exr render_aces.exr

Example 6: "List all the standard attributes on this plate"

uv run ${CLAUDE_SKILL_DIR}/scripts/exr.py stdattr get plate.exr

Troubleshooting

Error: exrheader: command not found

Cause: OpenEXR CLIs not installed. Solution: brew install openexr, apt install openexr, or build from source (github.com/AcademySoftwareFoundation/openexr).

Error: Invalid image data: cannot read chunk table

Cause: Truncated file (upload/download interrupted), or EXR 2/3 version mismatch. Solution: Re-copy the file. Confirm reader lib version: exrheader image.exr | grep version.

Renderer says "tiled EXR required"

Cause: Scanline file passed to a renderer that wants tiled for MIP pyramid. Solution: exr.py tiled. Some renderers (RenderMan) also want .tx — use maketx from OIIO (see vfx-oiio skill).

Deep file loads in Nuke but not in RenderMan

Cause: Deep file is deeptile but one tool only expects deepscanline. Solution: Check exrheader for type: deeptile vs type: deepscanline. Rewrite via OIIO if a converter is needed (no built-in converter ships with OpenEXR CLI suite).

Preview embedding fails on very large images

Cause: Some exrmakepreview builds can't allocate huge intermediate buffers. Solution: Pre-downscale the source with OIIO, then embed.

Reference docs

• Compression trade-offs table (all 11 types, lossy/lossless, speed) -> [references/compression.md]
• Standard attributes full list (fields + expected types + meaning) -> [references/standard-attributes.md]