OCIO Write

OCIO Write is where the pack's whole argument lands: it takes your image batch and writes it to disk color-managed, in the formats a real pipeline actually uses - EXR, DPX, TIFF, PNG, JPEG for stills and sequences; ProRes, DNxHR, FFV1, h264, hevc in MOV, MXF, MP4 or Matroska for video. It's Nuke's Write, and it exists because Save Image quietly 8-bits everything, and that's where dynamic range goes to die. The pack measured one LTX-2.5 generation: 2304 distinct brightness levels from the model, 220 through the stock write path. This node keeps all of it.

The core is two fields and one idea: input_colorspace (your working space, default sRGB - Display) → output_colorspace (what you encode into - the format picks a sensible default, EXR → ACEScg). Nothing is clamped on the way, and values below 0 and above 1 survive into any float container. The profile dropdown presets the whole node for known HDR sources: LTX 2.3 HDR (linear Rec.709 → ACEScg, because Lightricks' own node already undid the LogC3 curve), the two LumiPic log profiles, Seedance 4K 10-bit (a placeholder until that encoding is published), and SDR Rec.709 delivery. There is no LTX 2.5 preset on purpose - 2.5's HDR is ACEScct, and you undo that curve with OCIO LogConvert instead.

The choices that matter

• container - still image, sequence, or video. That drives everything below it.
• video_codec - seventeen of them. The measured table is worth internalizing: ffv1 (.mkv, 16-bit 4:4:4) is the only one that returns this pack's input bit-exact - md5 of the stream equals what went in, and it's a Library of Congress preservation format. hevc_444_12 is the only genuine 12-bit encode - ffmpeg's ProRes and DNxHR encoders top out at 10 bits no matter what their names claim (ProRes reads back as 12 and carries 10). h264/hevc are review codecs, and on an HDR output space they auto-move to 10-bit (BT.2100 has no 8-bit form).
• bit_depth - narrows to the format: EXR 16f/32f, DPX 10/16, TIFF 8/16/32f, PNG 8/16, JPEG 8.
• compression (EXR only) - zip (default) is lossless. dwaa/dwab are lossy: measured, they cost about half the distinct values, multiply the error by ~80, and quantise float32 to half before compressing - a 32f + DWAA file is a half-precision file that still says "float." Review copy yes, master or data pass never.
• auto_range - pulls frame range, start number and fps from the upstream OCIO Read automatically. Handy on canvas; API-only note: an API prompt gets whatever is in its JSON, so set the four fields explicitly for farm work.
• view (optional) - how to cross the scene/display boundary. Left on (none), the conversion is colorimetric: scene-linear 4.0 leaves as 1.78 and the container clips it flat. Pick an ACES 2.0 view and the output transform rolls highlights off instead. Use a view for a review deliverable, not a master.

Wire in alpha for RGBA, fps to carry the source rate, audio to mux a track, and metadata from OCIO Read to carry the plate's camera/lens/timecode into the file. A native ComfyUI video input lets you record a Load Video clip (or a Seedance-style API node's VIDEO) straight out with all these settings.

Where people get burned

The empty-folder "success." If a render reports success and the folder is empty, ComfyUI's front end likely dropped this output node via a partial-execution target list - a skipped node has nothing to report its own absence with. Queue from the graph instead. Second, EXR needs OPENCV_IO_ENABLE_OPENEXR=1 set before ComfyUI starts. Third, the preview tells the truth: the node shows what it wrote in its output colorspace, so a wrong colorspace pick looks visibly wrong rather than quietly wrong - trust that. Install via Manager (Nightly - the numbered registry versions are stale/flagged) or clone https://github.com/SlavaSexton/ComfyUI-OCIO + pip install -r requirements.txt. Video containers need a full ffmpeg build on PATH.