Working with EXR & HDR

LTX-2.5 works with high dynamic range (HDR)

• Bring your own EXR — Native EXR Workflow. Prefer scene-linear ACEScg OpenEXR (still or frame folder). Also accepted: scene-linear sRGB / Rec.709, or already-encoded ACEScct. Declare the space with --hdr, run the normal LTX pipelines, and get HDR masters back end to end.

• Generate HDR from SDR — HDR IC-LoRA (beta). If your source is standard-dynamic-range (SDR) video and you want an HDR version synthesized from it, the HDR IC-LoRA generates the extra dynamic range (HDR IC-LoRA is compatible with LTX-2.3 and that LTX-2.5 compatibility is under development).

Native EXR Workflow

Pass OpenEXR into the normal LTX pipelines with --hdr and get HDR masters out. Supported formats:

What You Get

HDR Video Files Are Not Accepted as Direct Input

Pipelines accept EXR stills or a folder of EXR frames — not HLG / PQ / HDR10 / mp4 or mov as conditioning. If you point the CLI at an HDR video file, it’s treated as ordinary SDR video decode and you won’t get HDR behavior.

If your source is an HDR video file, convert it first to a directory of OpenEXR frames in one of the supported formats. Do not leave display-encoded HLG/PQ inside the EXR:

my_clip_exr/
  frame_00000.exr
  frame_00001.exr
  ...

Then point the CLI at that folder (or a single still .exr via --image) and pass matching --hdr (ACESCG, SRGB_LINEAR, or ACESCCT).

Inputs

Outputs

When --hdr is set and you write, e.g., --output-path /tmp/out.mp4:

If --hdr is omitted (no EXR condition), output is standard SDR H.264 only.

How to Use

Image-to-video from an ACEScg HDR plate:

$
python -m ltx_pipelines.ti2vid_two_stages \
>
  --checkpoint-path /path/to/checkpoint.safetensors \
>
  --prompt "your prompt" \
>
  --image /path/to/plate.exr 0 1.0 \
>
  --hdr ACESCG \
>
  --output-path /tmp/out.mp4

Produces /tmp/out.mp4 (HLG) and /tmp/out_exr/*.exr (ACEScg-tagged).

Video conditioning / IC-LoRA from an EXR sequence:

$
# First convert HDR video -> a folder of EXRs (prefer ACEScg linear), then:
$
python -m ltx_pipelines.ic_lora \
>
  --distilled-checkpoint-path /path/to/distilled.safetensors \
>
  --video-conditioning /path/to/exr_frames/ 1.0 \
>
  --hdr ACESCG \
>
  --prompt "your prompt" \
>
  --output-path /tmp/out.mp4

Retake on an EXR sequence:

$
python -m ltx_pipelines.retake \
>
  --distilled-checkpoint-path /path/to/distilled.safetensors \
>
  --video-path /path/to/exr_frames/ \
>
  --frame-rate 24 \
>
  --hdr ACESCG \
>
  --start-time 1.0 \
>
  --end-time 2.0 \
>
  --prompt "your prompt" \
>
  --output-path /tmp/out.mp4

Generating HDR from SDR (HDR IC-LoRA, Beta)

The HDR IC-LoRA converts standard-dynamic-range (SDR) video into HDR output, producing per-frame EXR files suitable for professional color grading. Unlike the Native EXR path above (which carries your HDR through the pipeline), this adapter synthesizes dynamic range from an SDR source — use it when you don’t have an HDR original.

It runs via a ComfyUI workflow or a standalone Python script.

Model: Lightricks/LTX-2.3-22b-IC-LoRA-HDR (ltx-2.3-22b-ic-lora-hdr-0.9.safetensors)

ComfyUI Workflow

Files for ComfyUI:

Steps:

1. Install ComfyUI-LTXVideo custom nodes (see ComfyUI installation).
2. Download the HDR IC-LoRA and place it in ComfyUI/models/loras/.
3. Download the model checkpoint and distilled LoRA if you don’t already have them.
4. Get a free LTX API key from console.ltx.io; the workflow uses the Gemma API for text encoding.
5. Load the HDR workflow in ComfyUI and enter your API key in the LTX API KEY node.

$
export OPENCV_IO_ENABLE_OPENEXR=1

Example workflow steps:

1. Load your source SDR video.
2. The HDR IC-LoRA is applied via the LTX IC-LoRA Loader Model Only node.
3. Text conditioning uses the prompt "HDR footage" with a negative prompt for quality control.
4. Generation runs with CFG 1, the euler_ancestral sampler, and 8 inference steps.
5. The LTXVHDRDecodePostprocess node decodes the result into both an SDR preview and HDR linear output.
6. Save the HDR output as EXR frames.

Key nodes:

• LTX IC-LoRA Loader Model Only — loads the HDR IC-LoRA checkpoint and applies it to the model.
• LTXVHDRDecodePostprocess — decodes and post-processes the HDR output, producing tonemapped (SDR preview) and hdr_linear (full HDR linear data for EXR export). Its exposure parameter adjusts the SDR preview brightness in stops (±10 EV) and does not affect the HDR linear output.

Python Script

Files for the Python script:

Single video:

$
python run_hdr_ic_lora.py \
>
    --input ./my_video.mp4 \
>
    --output-dir ./hdr_output \
>
    --distilled-checkpoint /path/to/ltx-2.3-22b-distilled.safetensors \
>
    --upscaler /path/to/ltx-2.3-spatial-upscaler-x2-1.1.safetensors \
>
    --lora /path/to/ltx-2.3-22b-ic-lora-hdr-0.9.safetensors \
>
    --text-embeddings ltx-2.3-22b-ic-lora-hdr-scene-emb.safetensors \
>
    --exr-half \
>
    --high-quality

Batch processing: point --input at a directory of .mp4 files and the pipeline loads once and processes all of them sequentially. A run_hdr_batch.sh script is also provided, which spawns a fresh process per video for better memory isolation.

Output — each video produces a tonemapped SDR preview (my_video.mp4) and a folder of per-frame HDR EXR files (my_video_exr/). The EXR frames contain linear scene-referred values in Rec.709/sRGB primaries. Use --skip-mp4 for EXR only, or --no-save-exr for the preview only.

Options:

Viewing EXR Files

Standard image viewers can’t display HDR EXR files correctly. The frames contain linear scene-referred values and need no tonemapping for standard viewing in an OCIO-aware viewer.

In any OCIO-aware viewer (DJV, Nuke, Mari), use:

DJV is a free, open-source viewer designed for professional image sequences including EXR.

Importing EXR Files into DaVinci Resolve

HDR EXR files need specific project settings in DaVinci Resolve to display correctly.

1. In General Preferences, enable “use 10-bit precision in viewers if available”.
2. Open Project Settings → Color Management and set:

Leave all other settings at their defaults. Dolby Vision, HDR10+, and HDR Vivid should remain disabled.