raubatz/ltx/Lightricks/LTX-2.3-22b-IC-LoRA-Colorization - Storage Bucket

LTX-2.3 22B IC-LoRA Colorization

This is a Colorization IC-LoRA trained on top of LTX-2.3-22B, which restores natural color to grayscale, monochrome, or desaturated video while keeping subject identity, framing, and geometry untouched — only the color information changes.

It is based on the LTX-2.3 foundation model.

Model Files

ltx-2.3-22b-ic-lora-colorization-0.9.safetensors — the single released checkpoint (used for the published inference samples).

Model Details

• Base Model: LTX-2.3-22B Video
• Training Type: IC-LoRA (video-to-video)
• Control Type: Video-to-video — a grayscale/monochrome input (reference) video drives a color-restored output video.
• Reference Downscale Factor: 1 (the reference is encoded at 1× the output resolution).
• Pipeline details: No special pre/post color transform — the reference video is VAE-encoded as the control signal and the model predicts the colorized result.

Intended Use & Out-of-Scope

Intended use: Colorizing black-and-white, monochrome, or heavily desaturated footage — restoring plausible, natural color to a clip while preserving the original subject, composition, motion, and background geometry.

Out of scope: Any task other than colorization (it is not a deblur, denoise, decompression, or upscaling model); relighting or scene re-composition; and content far outside the training distribution. Generating far above the 960×544 training bucket can weaken the colorization effect.

Control Signal Requirements

• Control signal type: Grayscale / monochrome / desaturated source video.
• Expected input: A reference video (.mp4 / .mov / .mkv / .webm / .avi).
• Preprocessing: None required — the reference is VAE-encoded directly. The reference is used at 1× the output resolution (downscale factor 1).
• Alignment: The output matches the reference frame count, FPS, resolution, and aspect ratio. Best results at the 960×544×121 @ 24fps training bucket (both landscape 960×544 and portrait 544×960 were seen in training).

How It Works

The model is conditioned on both the reference video latents and a text prompt that describes the grayscale source and the desired color result. The prompt convention learned in training is:

Reference shows {grayscale scene description}. Edited shows the same scene with natural colors restored. COLORIZE {vivid natural-color description of the same scene}. Subject identity, framing, and background geometry are identical to the reference; only color information differs between reference and edited.

Representative prompt from a real run:

Reference shows a small wild rabbit sitting among rough textured boulders with a fallen log and dry grass behind it, rendered in high-contrast monochrome with soft natural daylight emphasizing the fine fur and the coarse stone surfaces. Edited shows the same scene with natural colors restored. COLORIZE a young brown cottontail rabbit with warm tan and grey-brown fur, a pale cream underside and soft pink inner ears, perched on weathered grey granite boulders flecked with green and ochre lichen. Behind it a bleached driftwood log and clumps of golden dry grass catch the warm late-afternoon sun, while muted green vegetation softens the blurred background. The light is gentle and warm, giving the rocks subtle earthy browns and the whole scene a calm woodland tone. Subject identity, framing, and background geometry are identical to the reference; only color information differs between reference and edited.

Usage

🔌 ComfyUI

1. Copy the LoRA weights into models/loras.
2. Load the LTX-2.3-22B base model and add ltx-2.3-22b-ic-lora-colorization-0.9.safetensors as the LoRA.
3. Start at strength 1.0 and adjust to taste.
4. Use an IC-LoRA (video-to-video) workflow from the LTX-2 ComfyUI repository, which already wires the reference-video control nodes. Connect your grayscale clip as the reference video and write the prompt using the COLORIZE convention above. Because the reference downscale factor is 1, a generic reference encode at output resolution is correct.
5. Start at or near the 960×544 training bucket; generating far above it can weaken colorization on high-frequency detail.

Recommended Settings

• LoRA strength / weight: 1.0 (the published samples used strength 1.0).
• Resolution & frames: Trained and validated at 960×544×121 @ 24fps (frames satisfy (frames-1) % 8 == 0). Both landscape (960×544) and portrait (544×960) were in training. Start near this bucket for the strongest, most consistent effect.
• Prompting: Follow the Reference shows … COLORIZE … only color information differs structure documented in How It Works. Describe the same scene in both halves; only change the color description. Keep identity/framing/geometry language intact so the model only alters color.
• Production inference recipe (what we used): Run via the distilled ltx_pipelines.ic_lora pipeline with the identity-safe, stage-1-only native hi-res recipe — render on a 2× canvas with --skip-stage-2 --tile-reference-encode (stage 2 is skipped so the reference stays anchored for the whole denoise), LoRA strength 1.0, seed 42, 121 frames @ 24fps. The distilled checkpoint uses fixed sigmas, so there is no CFG / guidance scale and no negative prompt. A dev-trained LoRA loads cleanly on the distilled checkpoint.

Examples

References

• Code: GitHub Repository
• Inference Pipeline: ltx_pipelines.ic_lora (LTX-2 distilled IC-LoRA pipeline)

Tips & Troubleshooting

• Weak or partial colorization at very high resolution: the model generates stage-1 at the full output resolution, which is well above the 960×544 training bucket. If color looks washed out or incomplete, lower the generation/output resolution toward the training bucket.
• Color bleed or oversaturation: drop the LoRA strength slightly (e.g. 0.8–0.9) and make the COLORIZE description more specific about the intended hues.
• Identity drift: keep the stage-1-only recipe (do not use the two-stage path for identity-critical clips) so the reference stays attached for the entire denoise.

Dataset

The model was trained using a proprietary dataset.

Training

• Technique: IC-LoRA (rank 128, alpha 128, dropout 0.05) on the DiT transformer; target modules attn1.to_q, attn1.to_k, attn1.to_v, attn1.to_out.0, ff.net.0.proj, ff.net.2.
• Hyperparameters: bf16 mixed precision, AdamW optimizer, learning rate 1.5e-4, cosine scheduler, gradient checkpointing on, batch size 1.
• Steps: 1500 total training steps configured. Released checkpoint: ltx-2.3-22b-ic-lora-colorization-0.9.safetensors.
• Infrastructure: LTX-2 Community Trainer (single-node, 8 GPU).

License

See the LTX-2-community-license for full terms.

Acknowledgments

• Base model by Lightricks
• Training infrastructure: LTX-2 Community Trainer