SDR → HDR Universal

If you only ever add one node from this pack, it's this one. SDR → HDR Universal is the pack's SDR-to-HDR orchestrator: a deterministic expansion, the learned RUDRA reconstruction, or both blended - one node, one dropdown, no VAE, no latent round trip. It's also the node the README's quick start tells you to wire first: Read → SDR → HDR Universal → Viewer.

The mechanism

Three stages, and knowing them makes the widgets make sense.

First the inverse EOTF: inverse_oetf decodes your display-encoded SDR (sRGB, Rec.709, Gamma 2.2, Gamma 2.4, or None) to scene-linear. Input is clamped to 0–1 first.

Then the expansion. Below knee nothing happens at all - the image is untouched. Above it, luminance expands up to peak_nits through a shoulder whose hardness is shoulder_gamma. The source comments are unusually careful here: the join between the two is made continuous in gradient, not just in value, so you don't get a visible ridge where the sky crosses the knee. That's the bug they fixed, and it's why 1.0 on shoulder_gamma is a documented mistake - it's a straight line to peak and it steps the gradient.

knee_mode decides how the knee is placed. adaptive measures it per frame as a luma percentile (knee 0.75 = the 75th percentile), which is right for footage that changes exposure. With batch_mode set to Video Frames, temporal_smoothing EMA-filters that knee across frames so it doesn't crawl; leave it adaptive and on Independent Images and you can watch the whole image pump between shots.

Then the learned pass. processing_mode is your choice: Expand (deterministic only), Recover (learned only), Hybrid (default - expansion plus masked learned recovery). The learned half is the same RUDRA pixel model as the dedicated Recover node, blended by rudra_blend; set that to 0 and you've disabled it without changing mode.

What to set, what to wire

Your two nits fields are the ones people confuse. peak_nits (default 1000) is the ceiling - what the PQ/HLG encode targets, and the hard limit on the output. reference_white_nits (default 203) is where SDR diffuse white lands: a white shirt, a page, a cloud. Headroom above 203 belongs to speculars.

The tooltip also mentions the escape hatch: setting reference_white_nits equal to peak_nits restores the old pre-3.4 behaviour of slamming SDR white to the display peak. Useful if you're matching an old graph, ugly otherwise.

pixel_recovery_mode defaults to highlights because, in the author's words, it's the safest and avoids hallucinating chroma in deep shadows. all will touch shadows too. pixel_strength (Direct Pixel backend only) doubles or kills the learned residual over its own inverse tone curve.

Outputs: image, highlight_mask, shadow_mask, highlight_confidence, shadow_confidence, and - the one to actually read - report. It's a string that tells you which path ran and why. Wire image into the Viewer for review, or into the pack's Write node with output_encoding set to Linear ACES2065-1 (AP0) for EXR delivery.

Install

Manager → search Radiance → Install → restart → refresh the browser. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/fxtd-studios/radiance.git
cd radiance
python -m pip install -r requirements.txt

Heavier than the average node pack - OpenEXR, OpenImageIO, opencolorio, transformers, scipy - and on Windows portable you should run the pip line with python_embeded\python.exe. The RUDRA weights (~5 MB) drop into ComfyUI/models/radiance/ on first use; they're non-commercial while the code is GPL-3.0.

Where it bites

The fallback is the big one, and it's documented: with no checkpoint, downloads disabled (RADIANCE_ALLOW_DOWNLOADS=0), HF_HUB_OFFLINE=1 or an unwritable models folder, Recover and Hybrid produce output bit-identical to Expand and don't complain loudly. Read report. A failed download is retried once per ComfyUI session, so restart to try again.

Second: Expand deliberately keeps SDR midtones at their SDR display level - below the knee the signal stays at 1.0 = 100 nits, and only the knee-to-white span opens up. If your grade looks flat compared to a BT.2446-style lift, that's the design. Raise reference_white_nits or grade after.

Third: the CLI-adjacent honesty from the community when this pack surfaced in March 2026 - someone in the announcement thread insisted it "won't create real 16/32-bit EXR." For Universal that's wrong, but only because of what you do downstream: pick Linear or AP0 out, write EXR from the pack's Write node, and pass float all the way. Encode to PQ and save a PNG and you've made a display-referred file with an HDR-shaped curve.