LTX-2.5 Pixel Spatial Upscaler: draft at 280p, then upscale 2x
The LTX-2.5 Pixel Spatial Upscaler re-renders a low-res draft 2x with invented detail. How the draft workflow works and how Sume's draft-then-final differs.
The LTX-2.5 Pixel Spatial Upscaler is an open adapter that takes a low-resolution video and re-renders it at twice the size, synthesizing fine detail instead of interpolating pixels. Lightricks recommends drafting at around 280p to settle composition, then upscaling. On Sume you can draft cheaply too, but with a different model and a different final step.
Specs come from the model card, read on 2026-10-02.
What is the Pixel Spatial Upscaler?
It is an in-context LoRA on LTX-2.5 22B. During denoising it consumes a text prompt and your low-resolution reference video alongside the noisy latents, and learns to map downscaled references to high-resolution renderings. The card says composition and motion are preserved while plausible high-frequency detail is generated.
- Input: a low-resolution video at a 2x downscale factor.
- Output: 2x higher resolution with synthesized spatial detail.
- Precision BF16, adapter rank 32, LoRA strength 1.0 recommended.
- The reference must match the target framing and duration.
Why draft at 280p first?
The card's intended workflow is to generate drafts at very low resolution, about 280p, to establish composition cheaply, then apply the 2x upscaler for the creative high-resolution final. The saving is real because most generations are thrown away. A bad composition costs a small render, not a full-size one.
The catch is in the name: the detail is invented. If a logo, a face or a piece of text matters, check the upscaled result against the draft, because the adapter creates texture the draft never had.
How does that compare with a draft on Sume?
Sume's catalog is hosted models, listed by GET /v1/videos/models, and each reports its own supported_resolutions. The docs list wan-3.0 at 2 to 30 seconds. Sume's pricing table in the repo carries Wan 3.0 provider list prices of $0.05, $0.10 and $0.20 per second at 480p, 720p and 1080p, and Sume bills list times 1.25, so $0.0625, $0.125 and $0.25 per second. A 5-second 480p draft is about $0.31 and a 5-second 1080p final is $1.25.
The difference is that on Sume the final is a fresh generation at the higher resolution. It will not match the draft shot for shot, and there is no field that conditions it on the draft. The LTX adapter is built around that match. If you need the final to be the same shot as the draft, the open route is the closer fit; if you only need a good clip, a fresh hosted render is simpler.
| Step | LTX-2.5 plan | Sume plan |
|---|---|---|
| Draft | About 280p on your GPU | 480p generation, about $0.0625 per second |
| Final | 2x upscale conditioned on the draft | New 1080p generation, $0.25 per second |
| Same shot as draft | Yes, by design | Not guaranteed |
| Setup | LTX-2 pipeline or ComfyUI | One API key |
Can Sume upscale a finished clip instead?
Yes, with Video Upscale 1.0, a separate route whose factor runs from 1.1 to 4 and whose reserve is $0.009 per input second. It enlarges the clip you already have, which is the safer way to keep the shot. See how to upscale video to 4K for the arithmetic. Neither route is free of artefacts, so compare at full size before you publish.
Sources
Related posts
More in Models
- LTX-2.5 quantization: fp8-cast, fp8-scaled-mm or NVFP4?
LTX-2.5's repo offers fp8-cast with bf16 checkpoints and fp8-scaled-mm on Hopper; the Hugging Face card lists NVFP4 and int8. Which flag fits which GPU.
- LTX-2.5 Refine Details: 22 minutes per clip on an RTX PRO 6000
The LTX-2.5 Refine Details LoRA rebuilds texture up to 4K but takes about 22 minutes for 121 frames on an RTX PRO 6000. A hosted upscale compared.
- LTX-2.5 Slow-Motion Control: speed 0.025 to 1.0, 40x slower
The LTX-2.5 Slow-Motion Control LoRA sets motion speed from 0.025 (40x slow) to 1.0 while playback stays 24 fps. Sume's video API has no speed field.
- LTX-2.5 Water Simulation LoRA: add rain or a flood to a clip
LTX-2.5 Water Simulation adds water to a dry reference clip at native 1920x1088. How its prompt works, and the hosted edit route on Sume instead.
Written by Sume