AI video restoration, built on LTX-2.3
Two purpose-built models for targeted restoration. Remove compression artifacts from low-bitrate sources and recover sharpness from defocused footage. Both keep identity and geometry locked; only degradation changes.
.svg)
Key Capabilities
Decompression: Remove compression artifacts
Remove macroblocking, chroma bleed, ringing, and banding from low-bitrate video while preserving subject, framing, and geometry. Not for blur, denoise, upscale, or colorization.
Deblur: Recover defocused footage
Invert spatial defocus to sharp focus across three blur families and four strength tiers. Not motion blur, compression repair, denoise, or super-res. Lower to 0.8 if over-sharpening.
Identity-preserving, open-source, batch-ready
Identity-safe recipe keeps reference attached. Deblur: native 1920×1088. Decompression: strongest near 960×544. Open weights via ComfyUI/ltx_pipelines, 80GB+ VRAM (32GB FP8).
Clean up compressed archives
Restore compressed masters, UGC ingest, and legacy exports for streaming and broadcast without losing quality. Decompression removes macroblocking, chroma bleed, ringing, and banding, rebuilding clean edges while preserving the original subject and composition.
Rescue defocused shots
Recover out-of-focus footage that would otherwise be unusable. Deblur regenerates plausible detail in defocused shots while handling optical defocus only, not motion blur. Sets accurate expectations for technical buyers who need predictable results.
Target restoration over black-box solutions
Stop guessing what an "enhance" button does. Pick the restoration model that matches your specific degradation: Decompression for compression, Deblur for defocus. Run them as separate targeted passes and know exactly what each one changes.
Restore batch library, self-hosted
Deploy open weights via a self-hosted ic_lora pipeline, integrate into your render or ingest infrastructure, and run the same settings on every clip, without per-clip licensing. Results still depend on each clip's condition.
Built for
Media supply-chain and archival engineers re-delivering compressed or degraded catalogues. Post specialists and pipeline TDs building targeted, diagnosable, self-hosted restoration 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 restoration works
Input
Technical characteristics:
- Reference video — The degraded clip in .mp4, .mov, .mkv, .webm, or .avi. VAE-encoded, no preprocessing, 1× resolution. Deblur applies to whole frame only.
- Text prompt — ENHANCE QUALITY (Decompression) or DEBLUR (Deblur) convention. Describe same scene in both halves, keep identity, framing, geometry language intact.
- Base model and LoRA — LTX-2.3-22B + ltx-2.3-22b-ic-lora-decompression-0.9.safetensors or ltx-2.3-22b-ic-lora-deblur-0.9.safetensors at LoRA strength 1.0. Deblur: use checkpoint step 1000.
- Pipeline — ltx_pipelines.ic_lora distilled pipeline (identity-safe, stage-1-only native hi-res). No CFG or negative prompt.
Output
Technical characteristics:
- Restored video — Matches your reference in frame count, FPS, resolution, and aspect ratio, with only the targeted degradation removed and content preserved.
- Performance notes — Decompression is strongest near the 960×544 training bucket. Deblur generates well at native 1920×1088.
- Hardware requirements — Self-host via ComfyUI or ltx_pipelines.ic_lora on 80GB+ VRAM (or 32GB with FP8 quantization).
FAQs
Are these one restoration model?
No. Decompression and Deblur are two targeted IC-LoRAs that can be run as separate passes depending on your degradation type.
Will Deblur remove motion blur?
No. Deblur v2 was trained on spatial and optical defocus only, so there is no motion or temporal blur in its training data. Motion blur is out of scope.
What does Decompression actually fix?
Macroblocking, chroma bleed, ringing, and banding from low-bitrate and heavily compressed sources. It's not a deblur, denoise, upscaler, or colorizer.
Do these models change the content of my footage?
No. Both condition on your reference and preserve subject identity, framing, and geometry. Only the targeted degradation changes. Extreme degradation where content is essentially destroyed may be hallucinated rather than faithfully reconstructed.
What hardware and license do I need for Reframe?
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.
