LTX-2.3-22b-IC-LoRA-Golden-Hour

LTX-2.3 22B IC-LoRA Golden Hour Relighting

Golden Hour Relighting IC-LoRA for LTX-2.3-22B. Feed a midday / regular daylight clip as the reference; the adapter re-renders the same shot with golden-hour lighting while preserving composition, framing, camera movement, and subject motion.

Based on LTX-2.3.

<p align="center"> <img src="assets/landscape_20103694_golden_hour_beach_sbs.jpg" width="48%" /> <img src="assets/landscape_17584730_golden_hour_desert_sbs.jpg" width="48%" /> </p> <p align="center"> <img src="assets/landscape_27469187_backlit_golden_hour_portrait_sbs.jpg" width="48%" /> <img src="assets/landscape_13265926_golden_hour_lake_sbs.jpg" width="48%" /> </p>

Training stills — left = midday reference, right = real golden-hour target.

Example Outputs

Gallery clips are static midday still (left) + golden-hour video (right) so the reference stays sharp while the target shows motion and lighting change.

<Gallery />

Model Files

Trained on the -dev base; for inference prefer the -distilled base.

Model Details

• Base Model: LTX-2.3-22B Video
• Training Type: IC-LoRA (video-to-video)
• Control Type: Midday / daylight reference video — conditions on a day clip and relights it to golden hour
• Reference Downscale Factor: 1
• Trigger: ADD GOLDEN HOUR (include in every prompt)
• Pipeline details: No special pre/post-processing. The daytime reference is resized + center-cropped to the target resolution at inference.

Intended Use & Out-of-Scope

Intended use: Convert short real-world midday / daylight videos into a photorealistic golden-hour version of the same shot (landscapes, coasts, streets, people in motion), keeping motion and layout intact.

Out of scope: New scenes or camera moves; night→day; already-sunset references; heavy stylization. Late checkpoints (750–1000) and very high strengths often oversaturate. Clips much longer than training length can drift toward the end.

Control Signal Requirements

• Control signal type: Midday / regular daylight video
• Expected input: A single reference video
• Preprocessing: Re-encode to clean H.264 MP4 at 24 fps before inference if needed (the reference loader does not resample temporally)
• Alignment: Output frame count should satisfy frames % 8 == 1; dimensions divisible by 32
• Mask support: Not supported

How It Works

The midday reference is VAE-encoded and supplied as in-context conditioning with a text prompt. The model keeps geometry and motion from the reference and replaces midday lighting with golden hour. The prompt steers the look (sun direction, shadows, sky, rim light).

Always include the trigger ADD GOLDEN HOUR:

Reference shows <the midday scene>. Edited shows the same scene with golden hour lighting.
ADD GOLDEN HOUR <concrete look: low warm sun, long soft shadows, amber rim light, warm sky>.
Subject identity, clothing, framing, and background geometry are identical to the
reference; only lighting and time of day differ between reference and edited.

Look phrases

• low warm sun with long soft shadows, amber rim light, and a warm orange-gold sky
• late-day golden backlight, honey-colored highlights, and gentle atmospheric haze
• magic-hour warmth with copper skin tones, elongated shadows, and a glowing horizon
• soft low-angle sunlight, golden speculars, and a peach-to-amber sky gradient
• warm directional sun flare, rich amber mids, and long blue-tinted shadow fill
• sunset side-light with golden edges on subjects and a vivid warm sky

More full prompts: example_prompts.json.

Usage

ComfyUI

1. Copy the LoRA into models/loras.
2. Load LTX-2.3-22B and add ltx-2.3-22b-ic-lora-golden-hour.safetensors.
3. Start at strength 0.85–1.0 (drop to 0.6–0.8 if too warm).
4. Use an IC-LoRA video-to-video workflow from ComfyUI-LTXVideo (e.g. single-stage distilled V2V). Connect your midday clip as the reference and include ADD GOLDEN HOUR in the prompt.

CLI (ltx-pipelines)

python -m ltx_pipelines.ic_lora \
  --distilled-checkpoint-path /path/to/ltx-2.3-22b-distilled.safetensors \
  --spatial-upsampler-path /path/to/ltx-2.3-spatial-upscaler-x2-1.0.safetensors \
  --gemma-root /path/to/gemma-3-12b-it-qat-q4_0-unquantized \
  --lora /path/to/ltx-2.3-22b-ic-lora-golden-hour.safetensors 0.85 \
  --reference-video /path/to/midday_clip.mp4 \
  --prompt "Reference shows a coastal scene under bright midday daylight. Edited shows the same scene with golden hour lighting. ADD GOLDEN HOUR low warm sun with long soft shadows, amber rim light, and a warm orange-gold sky. Subject identity, clothing, framing, and background geometry are identical to the reference; only lighting and time of day differ between reference and edited." \
  --output-path ./golden_hour_out.mp4 \
  --skip-stage-2

Recommended Settings

For identity-critical shots, keep generation on stage-1 only at native resolution. A second upscale stage that drops reference conditioning can drift identity or weaken the lighting effect.

-dev validation regime: guidance_scale 3.0, num_inference_steps 30 (see training_config.yaml).

Tips & Troubleshooting

• Too orange / hot: Lower strength, or try step 250; avoid 750/1000 as default
• Too weak: Raise strength toward 1.0, strengthen the look phrase, or carefully try step 750
• Wrong time of day: Reference must look like daylight midday, not already sunset
• Motion timing off: Resample reference to 24 fps before inference
• Long-clip drift: Stay near ~2–4s for best fidelity

Dataset

72 train + 8 holdout motion-aligned pairs:

Landscape and portrait, ~97 frames @ 24 fps, dual-panel captions with ADD GOLDEN HOUR. Stills: assets/. Prompts: example_prompts.json.

Training

• Technique: IC-LoRA (rank 32, alpha 32) on DiT attention + FFN projections; flexible strategy; reference p=1.0, first-frame p=0.15
• Hyperparameters: bf16, AdamW, LR 2e-4, cosine, batch 1, grad accum 4, gradient checkpointing
• Steps: Trained to 1000; release = step 500 (oversaturation visible by ~750)
• Buckets: 960×544×49, 544×960×49
• Infrastructure: LTX-2 Community Trainer, 1× NVIDIA H100 80GB

Config: training_config.yaml.

License

See the LTX-2 community license. Adapter derived from LTX-2.3.

Training footage from Pexels stock. This repo ships weights + illustrative examples only.

Acknowledgments

• Base model by Lightricks
• Training infrastructure: LTX-2 Community Trainer
• Midday reference stills: Qwen-Image-Edit

Citation

@misc{hoffman2026ltx23-ic-lora-golden-hour,
  title        = {LTX-2.3 22B IC-LoRA Golden Hour Relighting},
  author       = {Hoffman, Omer},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Hoffm4nz/LTX-2.3-22b-IC-LoRA-Golden-Hour}}
}