fxtdstudios/rudra-models

RUDRA models — FXTD Studios

The models behind RUDRA, FXTD Studios' SDR → scene-linear HDR reconstruction for VFX and grading. Apache 2.0, commercial use permitted.

Folder	What it is	Size
models/RUDRA_ACES_2/	RUDRA ACES 2: the learned ACES 2.0 reconstruction. 8-bit SDR in, an ACEScg master out, whose ACES 2.0 SDR view keeps the source's 8-bit codes. Ships as TorchScript and ONNX (DirectML/CPU).	2.2 MB
models/rudra_genfill_v3/	RUDRA AI fill v3: a LoRA for Wan2.1-VACE-1.3B that draws the shape inside fully clipped highlights, and a small brightness estimator (brightness.json) that sets how bright they are.	271 MB

Use

In the RUDRA app (Windows; macOS builds are unsigned for now):

• RUDRA ACES 2: File › Model packages, or Look › Download the model. Every file is checked against its SHA-256 before it is used.

• AI fill: File › Model packages › AI fill v3 › Download. The app then uses this LoRA and brightness.json. It also needs:

  • the base model Wan2.1-VACE-1.3B-diffusers (set its folder in Deliver › AI fill › Setup…, or put it in <RUDRA folder>/models/genfill/);
  • Python (scripts\GENFILL.ps1 -Setup);
  • a 16 GB GPU.

  Then use Deliver › AI fill.

From the command line (RUDRA repository):

rudra-native aces-v2 models/RUDRA_ACES_2 <frames or movie> --out <dir>
python -m rudra.genfill.run --master <EXR folder> --out <new folder> \
    --base Wan2.1-VACE-1.3B-diffusers --lora models/rudra_genfill_v3/rudra_genfill.safetensors

run.py loads brightness.json from the LoRA's folder by default. Add --brightness-offset <stops> to shift the brightness (±3).

RUDRA ACES 2: what it guarantees, and its evidence

• Codes kept: the master's ACES 2.0 SDR 100-nit view gives back the source's 8-bit codes within 1 code, fully clipped pixels included.
• 2.0 (w4 + lock): in partially clipped highlights the channels the camera recorded are held to the source (the lock), and only the clipped channel is rebuilt. Needs RUDRA 1.2.0 or later; older apps refuse the package.
• Held-out test (The Crew, ARRI ALEXA LogC3, 56 cases from 18 scenes, registered before it was run, run once): better than the plain ACES 2.0 inverse at −3 EV and +3 EV and in highlight colour; at the delivered exposure within 0.1 code RMS of it, not better (0.311 against 0.242). One production and one camera: a result on that footage, not on all footage.
• Rebuilt highlights are estimates, marked in the rudra.reconstructed EXR channel. Not camera RAW.

What the AI fill guarantees

• Codes kept: no 8-bit code of the master's ACES 2.0 SDR 100-nit view changes. Pixels it does not touch are copied bit for bit.
• Highlights only: it fills only fully clipped areas (code 255), and adds at most +3 stops over the master.
• Marked: every generated pixel is marked in a rudra.generated EXR channel.
• What it draws is plausible, not recovered. It is experimental and off by default in the app.

Evaluation (AI fill v3, pre-registered, 30 Oct 2026)

The test used 58 clips. The verdict counts only 25 Poly Haven scenes that were held out and never seen before (registration_v3.json, report_v3.json).

Gate	Result
Clipped-highlight log2 error vs the exact inverse	−0.128 stops (95 % CI −0.226 … −0.042)
Detail (PU21) at −3 / −6 EV	−7.1 / −6.5 (lower is better)
Brightness estimate vs leaving it	−0.199 stops (CI −0.339 … −0.076); overshoot > 1 stop in 3.3 %
8-bit codes / shadows / flicker / window seams	kept / unchanged / 0.007 EV / 1.0

On footage with small clipped areas it is neutral.

Training data

All training data is commercial-clean:

• Poly Haven HDRIs (CC0);
• FXTD Studios' own footage and renders;
• procedurally generated scenes;
• CC0 plates.

The base model, Wan2.1-VACE-1.3B, is Apache 2.0 from Wan-AI.

Not covered here: the older Beta 2 weights (sdr2hdr_image_v8, sdr2hdr_shadow_v1) are a separate, non-commercial release.

Licence

Apache License 2.0. Copyright 2026 FXTD Studios. Wan2.1-VACE-1.3B is © Wan-AI under Apache 2.0.

www.fxtdstudios.com · FXTD Studios