AI video upscaling, built on LTX-2.5.
The Pixel Spatial Upscaler IC-LoRA synthesizes fine detail into low-resolution video instead of simply interpolating pixels. Open weights, self-hosted.
.svg)
Key capabilities
Generative 2x upscaling
Synthesize new fine detail instead of interpolating pixels. A creative upscale, not a pixel-accurate resize.
Draft-low, finish-high workflow
Generate a draft at low resolution to lock composition and motion cheaply, then upscale to a high-resolution final.
Cost-efficient drafting
Draft at low resolution to nail composition and motion, then upscale only the shots that make the final cut.
Efficient finals at 2x
Turn a fast, low-resolution draft into a sharper final render at double the resolution, with synthesized detail instead of a soft interpolated upscale.
Self-hosted pipelines
Run upscaling inside your own pipeline instead of relying on a closed SaaS tool.
Stays in your existing workflow
Add the upscaler as a LoRA node in the same ComfyUI workflow you already generate in, no exporting footage to a separate app.
How it works
Input
Technical characteristics:
- Reference clip — A low-resolution reference clip, ideally a draft generated around 280p for composition and motion.
- Model — LTX-2.5-22B with the ltx-2.5-22b-ic-lora-pixel-spatial-upscaler-x2-1.0.safetensors adapter.
- Workflow — A video-to-video IC-LoRA workflow through the LTX-2 ComfyUI repository.
- Settings — LoRA strength set to 1.0, with weights pre-scaled for full strength, and a reference downscale factor of 2.
Output
Technical characteristics:
- Upscaled clip — A 2x upscaled clip with synthesized fine detail, not a faithful pixel-accurate reproduction of the source.
- Tunable fidelity — Fewer steps and lower guidance stay closer to the reference, more steps and higher guidance allow more creative detail.
- Out of scope — Not designed for denoising, compression-artifact removal, or live-action footage that requires factual accuracy. Use video restoration for those workflows.
What users are building with LTX
FAQs
Is this the same as video restoration?
No. Restoration removes compression artifacts and sharpens damaged footage while preserving fidelity to the source. This upscaler generates new detail and isn't intended for footage where accuracy to the source matters.
Is this the same as Diffusion Fidelity Rendering?
No. Diffusion Fidelity Rendering allocates compute by scene complexity during the initial generation. The Pixel Spatial Upscaler IC-LoRA is a separate, post-generation upscaling step.
Will it upscale real or live-action footage accurately?
Not recommended. It's a creative generative step, not a pixel-accurate refiner.
What resolution increase does it support?
Currently a 2x upscale.
How much control do I have over the result?
Guidance and step count let you trade off fidelity to the source against more hallucinated detail.
How do I run it?
Run the IC-LoRA video-to-video workflow from the LTX-2 ComfyUI repository, with the LTX-2.5-22B base model, under the LTX-2.x Community License.
