{"id":5818,"date":"2026-03-25T18:53:38","date_gmt":"2026-03-25T10:53:38","guid":{"rendered":"https:\/\/crepal.ai\/blog\/?p=5818"},"modified":"2026-03-25T18:53:40","modified_gmt":"2026-03-25T10:53:40","slug":"ltx-2-3-spatial-temporal-upscaler","status":"publish","type":"post","link":"https:\/\/crepal.ai\/blog\/aivideo\/ltx-2-3-spatial-temporal-upscaler\/","title":{"rendered":"LTX 2.3 Spatial and Temporal Upscaler: How to Use It"},"content":{"rendered":"\n

Hi guys, I’m Dora. Last week I sat staring at a 512\u00d7768 LTX 2.3<\/strong> output thinking it was almost perfect \u2014 the motion was clean, the composition worked \u2014 but it looked genuinely soft at anything larger than a phone screen. I knew the upscalers existed. I’d been putting off learning them because ComfyUI intimidates me whenever I have to wire up a two-stage pipeline from scratch.<\/p>\n\n\n\n

So I blocked off an afternoon, read through the official docs, broke things about four times, and finally got a setup that works reliably. This guide is that afternoon compressed into something you can actually follow.<\/p>\n\n\n\n

What the LTX 2.3 Upscaler Does (Spatial vs Temporal \u2014 What Each Fixes)<\/h2>\n\n\n\n

LTX 2.3<\/a><\/strong> ships with two separate upscaler checkpoints that solve two completely different problems. The spatial upscalers allow creators to generate at a manageable resolution and scale up afterward, while the temporal upscaler doubles the frame rate of existing clips. Used together, they make high-resolution, high-frame-rate output genuinely accessible on consumer hardware \u2014 which wasn’t really true before this release.<\/p>\n\n\n\n

Spatial upscaler<\/strong> (ltx-2.3-spatial-upscaler-x2-1.0.safetensors<\/code>) \u2014 takes your generated latent and doubles the spatial resolution. If you generated at 512\u00d7768, it upscales to 1024\u00d71536 while working entirely in the latent space, before VAE decode. This means it’s not just stretching pixels \u2014 it’s adding detail that the generation pass established but didn’t render at full fidelity.<\/p>\n\n\n\n

\n
\"\"<\/figure>\n<\/figure>\n\n\n\n

Temporal upscaler<\/strong> (ltx-2.3-temporal-upscaler.safetensors<\/code>) \u2014 doubles frame rate in the latent space. Generate at 24fps, get 48fps output. The temporal upscaler increases frame rate directly in the latent space, allowing creators to scale videos to higher FPS without regenerating the entire sequence. This is a much better result than RIFE-style interpolation, which guesses intermediate frames from pixels \u2014 the LTX temporal upscaler understands the motion structure of the generation itself.<\/p>\n\n\n\n

The key insight: you don’t have to run both every time. Spatial alone is the most common use case. Temporal matters most for motion-heavy content where 24fps looks choppy.<\/p>\n\n\n\n

When to Use the Upscaler vs Generating at Full Res Natively<\/h2>\n\n\n\n

This tripped me up at first. Why not just generate at 1080p from the start?<\/p>\n\n\n\n

The honest answer: you can, and sometimes you should. But generating at half resolution in Stage 1 and upscaling in Stage 2 gives you a meaningful workflow advantage \u2014 you get to preview motion, composition, and prompt adherence at low cost before committing to a full-resolution pass. Most creators prefer to generate smaller, faster previews first, and then upscale them \u2014 today’s upscalers take minutes to upscale a 10-second clip into 4K resolution.<\/p>\n\n\n\n

The practical rule I follow:<\/p>\n\n\n\n

Scenario<\/td>Approach<\/td><\/tr>
Quick iteration, testing prompt<\/td>Generate at 512\u00d7768, no upscaler<\/td><\/tr>
Reviewing motion quality<\/td>Add spatial upscaler after motion looks right<\/td><\/tr>
Final delivery, social\/web<\/td>Full two-stage with spatial upscaler<\/td><\/tr>
Cinematic output, slow motion<\/td>Add temporal upscaler on top of spatial<\/td><\/tr>
Very short clip (<5s), time-sensitive<\/td>Generate full-res natively<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

The spatial upscaler adds significant VRAM overhead (more on that below), so it’s not worth running on every draft. Wait until the generation is worth upscaling.<\/p>\n\n\n\n

\n
\"\"<\/figure>\n<\/figure>\n\n\n\n

How to Add the Upscaler to Your ComfyUI Workflow (Step-by-Step)<\/h2>\n\n\n\n

Before you start: make sure you’re on a recent ComfyUI nightly build, since LTX 2.3 nodes require it. If nodes are missing when loading a workflow, your ComfyUI version may be outdated \u2014 the Desktop version auto-updates when a new stable release is available.<\/p>\n\n\n\n

Step 1: Download the upscaler checkpoints<\/strong><\/p>\n\n\n\n

Both upscaler files are hosted on Hugging Face \u2014 Lightricks\/LTX-2.3<\/a>. Download:<\/p>\n\n\n\n