FXTD®
In-house tools About us Contact us

← In-house tools

In-house tools / Model

RUDRA

Radiometric Dynamic-Range Conditioning for HDR-Aware Diffusion Models.

Convert SDR images to HDR, or use trained decoders to generate scene-linear OpenEXR directly from diffusion models.

Apache 2.0Licence
8 backbonesDecoder variants
Scene-linearOpenEXR output
ComfyUIRadiance integration
Paper

What an 8-Bit Frame Can and Cannot Say About the Scene Behind It

FXTD Studios · 2026
Bounds for inverse tone mapping, and a gate that reaches one of them

Research on the limits of reconstructing HDR from clipped images.

BoundsInverse tone mappingBibTeX
Read the paper ↗
Weights

Download models

Apache 2.0 · Image-to-Image
SDR2HDR networks + HDR decoders

Download SDR-to-HDR models, a temporal refiner and HDR decoders for eight diffusion backbones.

$ huggingface-cli download fxtdstudios/RUDRA \
    --include "rudra_*_decoder_*.safetensors" \
    --local-dir "ComfyUI/models/radiance"
Open on Hugging Face ↗
Code

Source code

Reference implementation and evaluation

Enable rudra_decoder in Radiance’s HDR VAE Decode node to output scene-linear EXR.

Apache 2.0SafeTensorsComfyUI
View on GitHub ↗Radiance docs

What’s in the release

Inverse tone mapping

Reconstruct HDR from SDR footage with image models and a temporal refiner for sequences.

shadow_v1 · s2 · s3 · image_v5 · image_v6 · temporal_v1

HDR decoders

Generate scene-linear HDR directly, without a tone-mapped intermediate.

Flux.1 · Flux.2 · Flux.2 Klein · SDXL · Qwen-Image · Z-Image · Wan · LTX-Video — full + turbo

From SDR to HDR

An SDR image can lose highlight detail through clipping. Expanding its brightness alone cannot restore that detail.

RUDRA uses HDR training references to infer plausible highlights. Reconstruction estimates missing detail rather than recovering a unique original scene.

SCENE LUMINANCE → CLIP — 1.0 RUDRA SDR MASTER DETAIL DISCARDED HERE
Everything the SDR master folded onto 1.0 is what RUDRA has to infer

Before and after

Drag each divider — left is the source, right is RUDRA
01 — Table Mountain, sun on the horizonSource · RUDRA
Source plate: the sun is a flat pale patch welded into the sky
Same plate after RUDRA: the sun disc and flare are recovered
SourceRUDRA sun clipped flatsun disc + flare recovered

The reconstructed result adds a defined sun disc and flare while retaining foreground detail.

RecoveredSun disc + flare
ForegroundHolds detail
OutputScene-linear EXR
02 — Hillside road at duskSource · RUDRA
Source plate: the dusk sky has collapsed to a flat grey field
Same plate after RUDRA: cloud structure and warmth return to the sky
SourceRUDRA sky welded to greycloud detail + warm gradient

Cloud structure and a warm horizon appear in the reconstruction while the road and foliage retain depth.

RecoveredCloud banks
HorizonWarmth restored
ShadowsDensity held
03 — Cumulus over a quarrySource · RUDRA
Source plate: bright cumulus tops merged into one white mass
Same plate after RUDRA: modelling returns inside the bright cloud
SourceRUDRA cloud tops merged whiteform holds in the brights

The reconstruction adds shape within bright clouds while retaining blue in the sky.

RecoveredCloud modelling
SkyBlue holds
SequencesTemporal refiner
Bounded

Reconstruction limits

The paper explains the limits of HDR reconstruction from an 8-bit image.

Native

Native HDR output

Generate scene-linear OpenEXR directly from supported diffusion models.

Open

Open models and code

Model weights, reference code and Radiance integration are available under Apache 2.0.

Get the models ↗ Use it inside Radiance