AI video colorization, built on LTX-2.3
Restore natural color to grayscale, monochrome, or desaturated video while preserving the original footage. Subject identity, framing, and scene geometry stay locked to the source. Only the color information changes.
.svg)
Key Capabilities
Identity-preserving colorization
VAE-encoded reference is the control signal. Model predicts color only. Subject identity, composition, motion, and geometry stay locked. COLORIZE convention keeps output faithful to original
Temporal consistency across the whole clip
Video diffusion model on full reference ensures color coherence across every frame. No flicker or palette drift. Identity-safe, stage-1-only recipe keeps reference attached. Best at 960×544×
Open weights, self-hosted, batch-ready
Single checkpoint free on HuggingFace under LTX-2 Community License. Run via ComfyUI or ltx_pipelines on 80GB+ VRAM (32GB FP8). Repeatable results, no manual per-clip grading.
Archive and documentary colorization at scale
Colorize black-and-white and monochrome archive for streaming and broadcast re-release. Identity and framing stay intact, so historical footage preserves its integrity while gaining plausible natural color. Temporal coherence across long clips eliminates frame-by-frame inconsistencies.
Art-directed color via prompt control
Direct specific hues for skin tones, foliage, materials, and surfaces through the COLORIZE prompt description. Lower LoRA strength to 0.8–0.9 and refine the description to control bleed or oversaturation. Repeatable creative control, not a black box.
Replace frame-by-frame colorization pipelines
Replace per-frame colorization pipelines with one reference-conditioned pass. Get temporally consistent color without frame-by-frame stitching or post-blending. The upgrade from DeOldify-style workflows that fight flicker.
Batch colorization in production pipelines
Process large video libraries with one consistent color treatment. Open-source self-host means no per-clip cost or seat license, and it integrates directly into existing render infrastructure. Every clip in the batch gets consistent palette.
Built for
Restoration and archival engineers bringing black-and-white catalogues to color for broadcast re-release. Colorists and finishing teams building art-directable, temporally consistent colorization into post workflows.
Product teams, AI startups, and developers building AI-powered video features. Add production-grade video generation as a product capability, not a research project. One API, production-ready results, and no custom orchestration.
Brands, agencies, and creative teams producing high volumes of content. Turn existing assets into video at scale. Faster iteration, lower production cost, and more output from what you already have.
Teams that require full control over deployment and data. Run video generation in your own environment. On-premises, no cloud dependency, and full infrastructure ownership.
Platforms powering creative tools with multiple AI models. Upgrade your video output with a best-in-class engine. Improve generation quality, retain users, and differentiate with a model built for production, not prototypes.
How LTX-2.3 colorization works
Input
Technical characteristics:
- Reference video — Grayscale, monochrome, or heavily desaturated video (.mp4, .mov, .mkv, .webm, .avi). VAE-encoded, no preprocessing, 1× resolution.
- Text prompt — COLORIZE convention. Describe same scene in both halves, color description only differs. Keep identity, framing, geometry language.
- Base model and LoRA — LTX-2.3-22B + ltx-2.3-22b-ic-lora-colorization-0.9.safetensors at LoRA strength 1.0.
- Pipeline — ltx_pipelines.ic_lora distilled pipeline (identity-safe, stage-1-only native hi-res). No CFG or negative prompt required.
Output
Technical characteristics:
- Colorized video — Matches your reference in frame count, FPS, resolution, and aspect ratio with only color changed.
- Performance notes — Strongest and most consistent results at or near the 960×544 training bucket. Generating far above it can weaken colorization on high-frequency detail.
- Hardware requirements — Self-host via ComfyUI or ltx_pipelines.ic_lora on 80GB+ VRAM (or 32GB with FP8 quantization).
FAQs
Does it work on color footage or only black-and-white?
Built for grayscale, monochrome, or heavily desaturated sources. Feeding already-colored footage is out of scope.
Will it change the content of my footage?
No. The model is conditioned on the reference clip, so subject identity, framing, motion, and geometry are preserved. Only color information changes.
How is this different from a frame-by-frame colorizer like DeOldify?
This is a video diffusion model conditioned on the whole reference, so color stays temporally consistent across frames rather than being inferred per-frame and stitched. Per-frame approaches produce flicker and palette drift.
What resolution should I run?
Trained and validated at 960×544×121 @ 24fps in landscape and portrait. Start near that bucket for the strongest effect. Very high resolution can wash out the colorization.
What hardware and license do I need?
Open weights on HuggingFace under the LTX-2 Community License, requiring 80GB+ VRAM (or 32GB with FP8). The license is free under $10M ARR; enterprise licensing applies above that threshold.
