Read Books Books

Research snapshot: October 25, 2026, 11:36 PM EDT. Adapted into this archival white paper on October 29, 2026. These are results from the original computer configuration, before the later GPU and RAM upgrade.

Open the original interactive clip-review gallery · Five-second results · Ten-second results · Download all 43 entries

What this experiment was trying to answer

Before investing in a larger workstation, I wanted to know what my existing author computer could do with local video models. Could it make usable short performances from a reference image? How long did I have to wait? What happened when I asked for more resolution or a longer clip?

This paper preserves the review snapshot behind the original Author Video Lab gallery. It follows the October 24 H3 study and precedes the upgraded-computer comparison. The gallery lets readers inspect the pictures alongside their timings, rather than treating speed as proof of quality.

The computer and the scope

The original setup had an NVIDIA RTX 5070 with 12 GB of VRAM and 32 GB of system RAM. The gallery’s delivery clips are silent and run at 24 fps. Original generated audio remains in the research archive.

The saved public snapshot contains 43 completed clips: 21 H3, 17 LTX and 5 Wan. Twenty-eight entries use the Nina close-up; fifteen cover cart boarding, cart departure, a pass, a sedan action and a combined cart sequence. The published selection is a review collection, not every run in the research archive or a complete factorial matrix. It cannot supply an overall failure rate.

The recipes include H3 Base 20, H3 Turbo 4, native eight-step LTX and two-stage LTX 8+3, plus twenty-step Wan. One H3 ten-step pilot is labeled separately. A two-stage recipe combines a smaller first pass with latent upscaling and refinement; its step count is not directly comparable to the same number of native-resolution steps.

Every time in this paper is elapsed time to a finished file, including loading required by the current cache and output processing. Setup and downloads are excluded. These are individual observations, with no repeated-run confidence estimates or standardized cold/warm-cache comparison. The October 29 comparison paper also reports server execution time, which is a different timing boundary.

Five-second clips: what the original computer could do

To avoid blending different scenes into one model ranking, this table uses only the Nina close-up. Times are rounded to the nearest second; the download retains the original precision.

Nina close-up · 5-second clip480p720p1080p
H3 Base 203m 34s13m 04s53m 51s
H3 Turbo 41m 02s3m 03s11m 53s
LTX native 80m 52s1m 08s3m 15s
LTX 8+30m 31s0m 52s3m 25s
Wan 202m 27s7m 14s24m 10s

The fastest Nina five-second observation was LTX 8+3 at 480p, at 31.4 seconds to the finished file. At 720p, it took 52.5 seconds, native LTX took 67.6 seconds, and H3 Turbo took 183.1 seconds. Those waits suggest very different rhythms for trying a composition or camera action.

Five-second Nina clips at 720p on the original computer: LTX 8+3 takes 52 seconds; native LTX 1 minute 8 seconds; H3 Turbo 3 minutes 3 seconds; Wan 7 minutes 14 seconds; H3 Base 20 13 minutes 4 seconds.
Figure 1. One Nina observation per recipe, 720p, five-second output. Time to a finished file. All bars begin at zero; shorter is faster.

There is an important exception to any simple “8+3 is always faster” story: at 1080p for five seconds, native LTX finished in 3m 15s, while 8+3 took 3m 25s. Cache and processing conditions were not isolated, so this pair establishes what happened in those runs, not why.

The separately labeled H3 ten-step 1080p pilot took 28m 09s in this gallery snapshot. It must not be substituted for the twenty-step result of 53m 51s. The fact that both produce five seconds of video does not make them the same generation recipe.

Ten-second clips made the limits clearer

Nina close-up · 10-second clip480p720p1080p
H3 Base 209m 40s40m 52sNo completed entry
H3 Turbo 42m 29s8m 52sGPU out of memory
LTX native 81m 13s2m 55s40m 22s
LTX 8+30m 48s1m 45s6m 12s
Wan 205m 53s19m 53sNo completed entry

For the Nina ten-second 1080p case, native LTX took 40m 22s to a finished file, while LTX 8+3 took 6m 12s. The longer duration changed the time budget sharply. These figures include file preparation; the corresponding server-only times cited in the later paper are about 40m 10s and 6m 03s.

Ten-second Nina clips at 1080p on the old computer: native LTX takes 40 minutes 22 seconds to a finished file, compared with 6 minutes 12 seconds for LTX 8+3.
Figure 2. Same scene label and output length, two different LTX recipes. Single observations; file-ready times, not server-only times.

The H3 Turbo 4 attempt at 1080p for ten seconds ran out of GPU memory and produced no finished clip. The snapshot records that further H3 1080p/ten-second attempts were omitted by request. The absent Base 20 and Wan entries at that setting are labeled “No completed entry”; absence is neither a zero-second render nor evidence that those recipes failed in the same way.

More than one kind of shot

The gallery also includes movement and detail shots. At 720p for five seconds, LTX 8+3 took 52.9 seconds for boarding, 53.6 seconds for departure, 48.3 seconds for the pass, and 48.8 seconds for the sedan. Corresponding H3 Turbo observations ranged from 2m 53s to 2m 58s.

The combined ten-second cart sequence took 1m 51s with LTX 8+3 at 720p, compared with 8m 46s for H3 Turbo at that resolution. These are distinct shot packets; combining their times with Nina would conceal that change in content. The downloadable data identify scene, recipe, duration, resolution, seed and exact clip ID for every entry.

Saved review thumbnail from the Nina close-up LTX 8+3 benchmark.
A saved review thumbnail from the Nina LTX 8+3 720p five-second take. Open the original gallery to judge the moving result; a still does not establish motion quality.

Open the October 25 clip-review gallery and filter by scene, model, resolution or duration. It is a fixed snapshot, not a live queue. Use the filters to keep the scene and output length consistent before comparing recipes. The download controls preserve access to individual takes.

The snapshot explicitly marks visual quality as pending. Completed media and shorter waits do not prove that a shot is creatively approved, that dialogue matches, or that a full episode can be produced at that rate. This was a hardware-and-workflow exploration. The next production question is how much of the generated material survives into a finished edit.

Where this fits in the research

This original configuration could generate local footage, including some 1080p results. It also exposed substantial waits and at least one documented memory failure. That is the baseline against which later work should be read—not a claim that the old machine could do nothing.

The October 29 hardware paper compares 18 matched old/new pairs and reports the upgraded 1080p and 2K matrix. Its matched subset is not identical to this curated 43-clip gallery. The cost and studio-capacity paper then asks what ownership, rental and APIs might mean for a goal of five to ten finalized short episodes per day.

Sources and preserved data

  • Original interactive gallery, with its October 25, 2026, 11:36 PM EDT snapshot label.
  • Preserved 43-entry JSON snapshot and readable CSV, including exact file-ready times and links to the original clip URLs.
  • Source and adaptation notes. The narrative and figures use the saved local gallery files associated with the same Sites project. The Sites connector confirmed the active public gallery and version 3; an unauthenticated live data fetch returned 403, so live-versus-saved JSON byte equality was not verified.

Research dated October 25, 2026; archival write-up prepared October 29 with AI assistance and Elizabeth Ann West directing publication. Original clips, measurements and gallery remain preserved.