Lightricks Ships LTX-2.5 LoRA That Upscales Blurry Video to 4K
Lightricks dropped a new in-context LoRA for LTX-2.5 that rebuilds fine detail in soft or upscaled video and tiles its way to 4K on a single GPU.
Takeaways
- Lightricks released Refine Details, an IC-LoRA for LTX-2.5 that rebuilds fine detail in soft or upscaled video.
- Trained on 4K downscale pairs; learns only to restore the missing high-frequency band without altering framing or color.
- Uses per-step tiled latent fusion with 50% overlap, eliminating seams and per-tile flicker on 4K and 8K canvases.
- 1080p-to-4K takes one pass, about 22 minutes for 121 frames on an RTX PRO 6000, using 8 distilled steps.
- Shipped with sibling Restore LoRA for archival footage and a Native Resolution ComfyUI workflow.
- Weights available under the LTX-2.x Community License; integrates with ComfyUI and the LTX-2 repo.
Lightricks has released Refine Details weights, an open-weights, in-context low-rank adaptation (LoRA) for its LTX-2.5 base model. The add-on conditions on a soft, compressed, or previously upscaled video and synthesizes plausible fine texture while preserving composition, exposure, and color. Its Hugging Face page recorded roughly 2,900 downloads shortly after launch.
The same launch added two components for visual-effects and post-production workflows:
| Component | Purpose |
|---|---|
| Refine Details | Reconstructs edges, texture, and grain while keeping the source appearance. |
| Restore IC-LoRA | Repairs archival footage and can alter its aged or damaged appearance. |
| Native Resolution | Runs AI edits over 4K and 8K plates on one GPU using overlapping tiles. |
Rebuilding the missing frequency band
Pushing an HD shot to 8K with Lanczos interpolation enlarges the frame while leaving missing high-frequency information unresolved. That information includes fine edges, hair, fabric texture, grain, and other rapid pixel-level changes. Compression and earlier upscaling can remove those signals, leaving Refine Details to generate a temporally consistent estimate from the remaining image.
Each training example starts with 4K footage. Lightricks created a 1024x576 target through high-quality downscaling, then degraded the same tile at varying severities to produce a soft reference. The LoRA learned to map those degraded references back to native-looking targets. Its narrow objective is designed to preserve framing, exposure, and color while reconstructing the missing detail band.