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 Refine Details is an open adapter that rebuilds fine texture in compressed or upscaled video, and the model card gives one timing: about 22 minutes for 121 frames on an RTX PRO 6000. That makes it a finishing tool for a few key shots. For bulk enlargement, Sume's hosted video upscale is a different operation with a published per-second rate.
Facts are from the Refine Details card, read on 2026-10-02.
What does Refine Details do?
It takes a clip and optionally a reference image, for example a face or a logo, and re-renders it with the texture and sharpness of a native high-resolution capture. It is an in-context LoRA on LTX-2.5 with the Gemma-4 12B text encoder. The card says it is best at taking 1080p toward 4K, and that you can reach up to 8K through multi-pass refinement.
What does 22 minutes buy you?
The card recommends 8 steps and processes 1024x576 tiles with tiled fusion. On an RTX PRO 6000 it reports roughly 22 minutes for 121 frames. At 24 frames per second that clip is a little over 5 seconds, so a 30-second shot would take over two hours at the same rate, which is arithmetic from the card's one figure and not a benchmark of ours.
The card also lists what it is not for: colour restoration, pristine cinema raw footage and single-pass 8K. It slightly desaturates archive footage and smooths codec noise.
| Item | Card value |
|---|---|
| Tile size | 1024x576 with tiled fusion |
| Steps | 8 |
| Timing | About 22 minutes for 121 frames, RTX PRO 6000 |
| Best range | 1080p toward 4K |
| Not for | Colour restoration, pristine raw, single-pass 8K |
What does the hosted alternative cost?
Sume Video Upscale 1.0 accepts a public https video_url, an upscale_factor of 1.1 to 4, and a duration_seconds up to 30 (5 if omitted). The catalog lists $0.009 per input second, so a 6-second clip is $0.054 and a 30-second clip is $0.27. Jobs are asynchronous. The model behind it is chosen server side, and it takes no reference image.
That last point is the real difference. Refine Details can be told what a face or logo should look like. Sume's route has no such input, so it cannot restore a specific identity detail.
Which one for which shot?
Use Refine Details on the one hero shot where texture matters and you own a large GPU, and check the tiles for seams. Use a hosted upscale for batches, social cuts and anything where waiting 22 minutes per five seconds is not acceptable. If you are unsure, run both on a single 5-second clip and compare at 100 percent. The GPU break-even post shows how to do the same rent-or-hosted sum for generation.
As ever, the licence is the LTX community licence from the card, and Sume lists no LTX model in its video catalog.
Sources
Related posts
More in Models
- 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.
- Luma API camera concepts: 34 named moves vs Sume's prompt-only camera
Luma's API lists 34 named camera concepts you can layer, such as orbit_left and dolly_zoom. Sume has no camera field, so here is how to ask for the same moves.
- Luma Photon API: 4 image refs and a character ref vs Sume references
Luma's photon-1 image API takes up to 4 image references, a style reference and a 4-image character reference, each with a weight. Here is the Sume equivalent.
Written by Sume