gregt/T2_IC_SDR2HDR_LTX_LoRA

LTX-2.3 SDR → HDR IC-LoRA

A LoRA adapter for LTX-Video 2.3 (22B) that converts SDR video into HDR (LogC3 encoded), trained as an IC-LoRA (in-context LoRA) using the video_to_video strategy from ltxv-trainer.

The model takes an SDR clip as a conditioning reference and generates a matching HDR (LogC3) version, suitable for grading downstream in DaVinci Resolve, Baselight, or Nuke.

Inspired by Lightricks' LumiVid paper (arXiv:2604.11788) — same core idea (LogC3 target, IC-LoRA-style reference conditioning, ~10K steps).

Checkpoint

lora_weights_step_07000.safetensors — step 7,000 of a planned 10,000-step run.

Usage (ComfyUI)

Use LTXICLoRALoaderModelOnly from Lightricks/ComfyUI-LTXVideo:

1. Load LTX-2.3 base model
2. Load this LoRA via LTXICLoRALoaderModelOnly
3. Connect your SDR clip as the reference video
4. Run the IC-LoRA pipeline — Stage 1 (low-res, LoRA active) → Stage 2 (upsample, no LoRA)

No trigger word — the LoRA is always active when loaded.

Recommended CFG: up to 1.5 works well.

Training Details

Base model	LTX-Video 2.3 22B (ltx-2.3-22b-dev)
Text encoder	Gemma 3 12B
Strategy	IC-LoRA (video_to_video in ltxv-trainer)
LoRA rank / alpha	32 / 32
Target modules	attn1/2.to_{k,q,v,out.0}, ff.net.0.proj, ff.net.2
Resolution	1280 × 736
Frames per clip	49
Batch size	1
Learning rate	2e-4, cosine schedule
Steps	7,000 (target 10,000)
Precision	bf16 + gradient checkpointing
Log curve	LogC3 (validated optimal via KL divergence in LumiVid)

Dataset

• 200 clips rendered from PolyHaven HDRIs with virtual camera moves
• Paired SDR ↔ LogC3 HDR clips
• SDR side intentionally degraded (compression, blur, contrast, white-balance shift) to mimic real-world camera capture

Intended Use

• Converting SDR (Rec.709) video to LogC3 HDR for further grading
• VFX / DI pipelines that already speak LogC3 (Nuke, Baselight, DaVinci Resolve)
• Outputs are designed to be exported to ProRes 4444 or OpenEXR sequences

Limitations

• Trained primarily on PolyHaven HDRI environments — limited human/motion diversity in v1
• Output is LogC3, not display-referred HDR — you still need a grading pass to map to PQ/HLG/Rec.2100
• The model can do a surprisingly good job of hallucinating believable detail back into clipped highlights (skies, practicals, blown-out windows), but this is generative reconstruction — it is not faithful recovery of the original photons. Use with eyes open in any pipeline where photographic accuracy matters.
• Inference is two-stage (low-res → upsample); expect LTX-2.3's typical compute footprint

Citation

If this is useful, please reference the LumiVid paper that inspired the approach:

@article{lumivid2026,
  title={LumiVid: Distilling LDR Video Diffusion Models for HDR Video Generation},
  author={Lightricks},
  journal={arXiv:2604.11788},
  year={2026}
}

License

Apache 2.0. Base model license (LTX-Video) applies to inference use.