SF EXR

SF EXR

Load, edit, mask and save multi-layer EXR files and image sequences in ComfyUI.

• Layer dropdowns fill from the file header as soon as you set a path; nothing has to run first.
• Point at one frame of a numbered sequence (any format) and the whole sequence loads. ####, %04d and @ patterns work too.
• Pick frames with a list: empty for all, 1001-1100, 0, 15, 20, first-1050, plus a step. Negative (pre-roll) frames like shot.-010.exr are read as frame −10.
• Browse on SF EXR Load opens the ComfyUI machine's folders with sequences collapsed to one line (beauty.####.exr 1001–1100 100). Files are read where they are; nothing is copied. Dropping a single image on the node copies it into input/sf_exr/ for quick tests.
• Masks from cryptomatte names, depth bands, world-position regions and normals, or click them in SF Mask Picker. For sequences the picker has a timeline over the frames SF EXR Load will run, so you can check the mask on any frame (← / → over the preview step a frame, Shift steps 10).
• Re-rendered files are picked up when you come back to the ComfyUI tab, or with right-click → Reload files.
• OCIO through OpenImageIO's built-in ACES configs; no separate OCIO install.
• Saves keep the source layout, metadata (timecode, pixel aspect), overscan and frame numbers. Cryptomatte is always written as lossless 32-bit float.

Install

ComfyUI-Manager: search for SF EXR and install. Or with comfy-cli: comfy node install comfyui-sf-exr.

Manually: clone into ComfyUI/custom_nodes, then install the one dependency into ComfyUI's Python:

git clone https://github.com/shanef3d/ComfyUI-SF-EXR
python -m pip install -r ComfyUI-SF-EXR/requirements.txt

OpenImageIO ships prebuilt wheels for Windows, Linux and macOS. Torch and numpy come with ComfyUI. An example workflow is in example_workflows/ and appears in ComfyUI's Templates browser.

Shared ComfyUI servers

When ComfyUI only listens on this machine (the default), SF EXR can read and browse any folder. When it is reachable from other machines (--listen), SF EXR only reads, lists and writes inside ComfyUI's input, output and temp folders plus the folders in the SF_EXR_ROOTS environment variable, set on the server (separate folders with ; on Windows, : elsewhere):

set SF_EXR_ROOTS=Z:\shows;\\nas\renders

Nodes

| Node | Does | |---|---| | SF EXR Load | File or sequence → SF_EXR, the selected layer as IMAGE/MASK (as stored), and a context for SF EXR Info or SF Image to EXR | | SF EXR Info | Context → frame count, first/last frame, width, height, layer, layers, colorspace and a readable summary | | SF EXR Resize | Every layer resized the right way, with Resize Image/Mask's options: color filtered in float (HDR kept), data passes not blended across edges, motion vectors rescaled, cryptomatte re-ranked so mattes stay exact; stretching sets the pixel aspect ratio | | SF EXR Get Layer | EXR → one layer as IMAGE/MASK (scene-linear, unclamped) | | SF EXR Set Layer | Replace or add a layer from IMAGE/MASK | | SF EXR Rename Layers | Strip prefixes, find/replace, or old = new lines (layers and channels); cryptomatte metadata follows | | SF EXR Keep/Remove Layers | Keep or drop layers by name or wildcard | | SF Merge | Nuke-style merge of A onto B (over, plus, multiply, screen, in, out, …) with opacity, mask and "set bbox to"; works on EXR layers or images, output matches B | | SF EXR Save | EXR → multi-layer or multi-part file; save metadata writes the ComfyUI prompt, workflow and a summary (models, seeds, prompt text) into the header | | SF EXR Convert | Re-write EXRs with new compression, bit depth or layout, file to file | | SF Image Save | PNG / TIFF / JPG / DPX / single-layer EXR at a chosen bit depth | | SF OCIO Transform | Colorspace conversion (presets for common ACES moves) | | SF OCIO Display | Display/view transform, or its inverse | | SF EXR to Image | Scene-linear layer → 0–1 image for a diffusion model; crop to mask sends only the masked region at the model's size | | SF Image to EXR | Model output → scene-linear, pasted back (into the crop it came from) and composited through the mask | | SF EXR to Preprocessor | Render pass → control image: depth (near = white, one range for the whole shot), camera-space normals, or edges from cryptomatte boundaries / depth jumps | | SF Mask from Cryptomatte | Names or wildcards from the manifest → MASK | | SF Mask from Depth | Near/far band with falloff | | SF Mask from Position | Sphere, box, cylinder or plane in world space | | SF Mask from Normal | Surfaces facing a direction | | SF Mask Picker | Click cryptomatte objects, depth bands or 3D shapes on a preview; tune each with Z / Range / Smooth or Shape / Range / Smooth; preview any layer or the final mask |

Typical edit

SF EXR Load → SF EXR to Image ── image + crop mask → (inpaint / edit model) → SF Image to EXR → SF EXR Save
                    ↑ crop mask
              SF Mask Picker

Set SF EXR to Image's crop to crop to mask so the model works on just the masked region at about 1 MP (with padding pixels of context), instead of shrinking a 4K plate. Its crop mask output matches the cropped image for inpainting; alpha is always the layer's own alpha channel. SF Image to EXR resizes the edit back to the crop, converts it to scene-linear and composites it through the mask, so pixels outside the mask keep their original HDR values. Other passes (depth, cryptomatte) pass through unchanged; they still describe the original render where the edit changed shapes.

Compression rules

Save and Convert never let a lossy codec damage data. DWAA/DWAB are only lossy on half-float R/G/B channels and B44 on half channels, so float data layers stay single-part. When a codec would damage a data or cryptomatte layer (PXR24 on float, B44 on half data), the file is written multi-part and those layers get ZIP. compression: source keeps the original codec; with a lossy source (DWAA) each save loses a little more.

Layer names across renderers

Cryptomatte layers are found from the file's own cryptomatte/<id>/name metadata, so any renderer that follows the Cryptomatte spec works. Depth, position and normal passes are found by name (Z/depth, P/position, N/normal, including names like VRayZDepth); pick the layer explicitly in the dropdown, or use SF EXR Rename Layers, when a renderer uses something else.

Development

pip install -r requirements.txt numpy pytest ruff torch
ruff check .
pytest tests

Tests build small synthetic EXRs (beauty, depth, position, cryptomatte) on the fly; no render files are committed. Releases: bump version in pyproject.toml and add a CHANGELOG.md entry; pushing that to main publishes to the Comfy Registry.

License

MIT