Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents: 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →

DiffHDR Turns Ordinary 8-Bit Video Into HDR and Outputs EXR Frames for Nuke

In 8-bit SDR video, a window often becomes a white blob and a dark corner a black hole, and color grading can't bring those details back. DiffHDR finds these areas and paints in the missing brightness through a LoRA for the Wan2.1-VACE-14B model, drawing on neighboring frames. The result is HDR (high dynamic range) in EXR frames for Nuke and DaVinci Resolve. The code and weights are open, so you can test the method on your own footage right away.

AI-processed from TG: @cgevent (Метаверсище и ИИще); edited by Hamidun News
DiffHDR Turns Ordinary 8-Bit Video Into HDR and Outputs EXR Frames for Nuke
Source: TG: @cgevent (Метаверсище и ИИще). Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceTG: @cgevent (Метаверсище и ИИще) ↗
Editor-in-Chief
AuthorsEyeline Labs, Netflix and Texas A&M; the paper has been on arXiv since April 7, 2026
Releasedthe code, two LoRAs (for video and for HDRI) and two ComfyUI integrations
Key metric #3Input is MP4, processed internally in Log-Gamma; output is linear OpenEXR frames

Ordinary MP4 video stores brightness in 8 bits. This is SDR, or standard dynamic range. If a window was overexposed during shooting, it is simply white in the file, and if a corner of the room fell into shadow, it is black.

Color correction is powerless here: the file no longer contains any data you could "pull back." HDR is video with a wide brightness range, where the same window shows clouds and the corner shows furniture. DiffHDR tries to produce that kind of HDR from an ordinary SDR clip: it finds the lost areas and fills them in with a neural network.

The result is saved as EXR, a frame format with headroom for brightness that is used in post-production. If you work with archival footage or clips from simple cameras, or you build ComfyUI graphs, you can already download the project and try it.

How it works

The source describes the pipeline as a chain of steps. The input is ordinary 8-bit MP4 video. The algorithm first looks for blown highlights and crushed shadows, meaning areas where information has been lost. The image is then converted to Log-Gamma, a logarithmic representation of brightness, and the Wan2.1-VACE-14B video model takes over. The DiffHDR LoRA is attached to it: a small fine-tuned add-on that teaches a large model one narrow task.

The model does not look at a frame in isolation. It uses neighboring frames, scene context and optional hints. A simple example: in one frame a lamp has become a white circle, while in the next frame part of the filament is still visible. The model can "match" those frames and reconstruct how the lamp should look. After that, the values are converted from Log-Gamma back to linear space and saved as a sequence of OpenEXR frames.

This kind of output matters: EXR files can be opened directly in Nuke, DaVinci Resolve and other compositing and color-grading tools. The result does not stay in a demo player but drops into a colorist's normal workflow.

What you can download now

Here is the timeline, based on the post. The first version of the DiffHDR paper appeared on arXiv on April 7, 2026, and the third revision is dated August 31, 2026. As of November 2026, the following are available:

  • source code in the Eyeline-Labs/DiffHDR repository on GitHub;
  • two LoRAs: one for video and one for HDRI (lighting panoramas);
  • official weights on Hugging Face;
  • a ComfyUI port by gireshahid33;
  • an alternative ComfyUI integration by claussteinmassl.

Eyeline Labs, Netflix and Texas A&M worked on the project. The two independent ComfyUI nodes suggest that the community has picked it up. For anyone building graphs in ComfyUI, this is the shortest way to try the method without wrestling with research code.

Where the news comes from and the caveats

The author of the @cgevent channel learned that DiffHDR was still alive from a neighboring story. According to the post, the author of one of the ComfyUI ports, gireshahid33, works as a compositor at Industrial Light & Magic. The day before, he published the code for a Nuke node for Corridor Key 2.0. That is a separate compositing tool, and it merely prompted the observation that DiffHDR is under active development. The source has no details on Corridor Key 2.0, so there is nothing to assess here.

The source gives no quality metrics, VRAM requirements or processing speeds. Wan2.1-VACE-14B is a model with 14 billion parameters, so problems on weak hardware are likely. You will need to look for exact figures in the repository and the paper. Also, recovering blown highlights is generation, not extraction of real data: the model paints in plausible brightness. For documentary or scientific footage, that is a fundamental difference.

What to do in practice

For studios and small post-production teams, the value is clear: old footage and clips from 8-bit cameras can be brought up to HDR delivery without reshoots. For tool developers, this is an example of how a LoRA on top of a large video model becomes a narrow professional utility with output in an industry format.

If you have test footage with blown highlights, start with the ComfyUI ports. Pick a few difficult scenes: windows, street lamps, a sunset sky. Compare the result with the usual way of stretching SDR into HDR, and check whether artifacts have appeared in the painted-in areas.

Story

Key Questions Answered

How it works?

The source describes the pipeline as a chain of steps. The input is ordinary 8-bit MP4 video. The algorithm first looks for blown highlights and crushed shadows, meaning areas where information has been lost. The image is then converted to Log-Gamma, a logarithmic representation of brightness, and the Wan2.1-VACE-14B video model takes over. The DiffHDR LoRA is attached to it: a small fine-tuned add-on that teaches a large model one narrow task. The model does not look at a frame in isolation.

What you can download now?

Here is the timeline, based on the post. The first version of the DiffHDR paper appeared on arXiv on April 7, 2026, and the third revision is dated August 31, 2026. As of November 2026, the following are available: source code in the Eyeline-Labs/DiffHDR repository on GitHub; two LoRAs: one for video and one for HDRI (lighting panoramas); official weights on Hugging Face; a ComfyUI port by gireshahid33; an alternative ComfyUI integration by claussteinmassl.

Where the news comes from and the caveats?

The author of the @cgevent channel learned that DiffHDR was still alive from a neighboring story. According to the post, the author of one of the ComfyUI ports, gireshahid33, works as a compositor at Industrial Light & Magic. The day before, he published the code for a Nuke node for Corridor Key 2.0. That is a separate compositing tool, and it merely prompted the observation that DiffHDR is under active development. The source has no details on Corridor Key 2.0, so there is nothing to assess here.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…