Read Books Books

Earlier benchmark evidence: October 25 baseline white paper · Original interactive clip-review gallery. The original gallery covers the RTX 5070 / 32 GB RAM configuration and reports time to a finished file.

Before and after · 1080p and 2K · Memory · Watch the fast takes · Data and methods

Executive summary

I am building toward a local studio for short vertical dramas: roughly a minute to a minute and a half of story, assembled from stills, short video performances, dialogue and editing. The first question after upgrading my computer was practical: how long would I wait to see the next idea?

The strongest before-and-after evidence comes from 18 matched pairs: three scenes, three video recipes, two resolutions and five-second clips. At 720p, median server render time fell by 44.9% to 59.3%, depending on the recipe. Those are whole-computer observations, with the same image, prompt, seed and generation settings. GPU, RAM, storage conditions and cache behavior were not isolated as separate experiments. Matched measurements

The upgraded setup also completed all 96 configurations in the new matrix, including 48 at 1080p and 2K. The most immediately useful result for my studio planning is LTX 8+3: a five-second 1080p clip took a median 53.2 seconds to render and 61.0 seconds to a finished file. A ten-second clip took 2m 12s to render, so “about a minute” needs a clip length attached to it. New measurements

This first paper is about turnaround time and available working choices. A separate watch page shows the fastest completed examples. The clips are deliberately selected by measured speed, rather than presented as polished episode footage.

What changed in the computer

The earlier paper documented an HP OMEN desktop running local H3 on its existing RTX 5070. This follow-up compares that earlier configuration with the upgraded workstation. The Ryzen 7 8700F processor remained the same.

ComponentEarlier configurationUpgraded configuration
GPUNVIDIA RTX 5070NVIDIA RTX PRO 5000 72GB Blackwell
Dedicated GPU memory12 GB VRAM72 GB VRAM
System memory32 GB RAM128 GB RAM, four 32 GB modules
CPUAMD Ryzen 7 8700FAMD Ryzen 7 8700F
Operating systemWindows 11 HomeWindows 11 Home
RAM configuration recorded for this testClock not established in the selected old snapshot3600 MT/s configured; modules rated for 5600 MT/s

The four-module setup prioritizes 128 GB capacity. This study does not compare it with a two-module, 64 GB configuration at a higher clock. Models were also moved to the internal SSD during the upgrade work. The timings therefore describe the working setup as used; they do not isolate a GPU-only gain, a RAM-only gain or an SSD-only gain. Hardware and scope

VRAM and system RAM remain different memory pools. They should not be added together and described as the memory capacity of one GPU.

The cleanest comparison: the same shots at 720p

The three repeatable packets were a cart boarding action, a cart departure and a close-up of a pass. Each before-and-after pair used the same reference image bytes, motion prompt, seed, generated dimensions, frame count, frame rate and sampling recipe. We also checked the saved workflow graphs: differences were input/output filenames and an H3 frame-count helper replaced by its identical recorded value of 124 frames.

The bars below show the median of those three scenes for each recipe. Each scene has one measured run per configuration; these are not repeated-run confidence estimates.

Matched 720p five-second clips: H3 Turbo drops from 170.1 to 71.347 seconds; LTX native 8 from 75.834 to 30.841; LTX 8+3 from 46.172 to 25.460. Lower is faster.
Figure 1. Same three scenes, five-second clips, 720p. All bars start at zero. Server time includes loading as needed, inference, decoding and saving. [1]
720p recipe · 5-second clipEarlier medianUpgraded medianLess render timeEarlier ÷ upgraded
H3 Turbo 4170.1 s71.3 s58.1%2.38×
LTX native 875.8 s30.8 s59.3%2.46×
LTX 8+346.2 s25.5 s44.9%1.81×

That changes the rhythm of trying a shot. A roughly three-minute H3 wait became a little over a minute in this sample. Both LTX recipes came back in roughly half a minute. These ratios divide the old three-scene median by the new three-scene median; they are not an average of individual speedup ratios or a general rating for either GPU.

The matched 480p results also improved. Median times went from 63.9 to 23.9 seconds for H3 Turbo, 43.1 to 25.4 seconds for LTX native 8, and 27.1 to 15.2 seconds for LTX 8+3. The range matters: new native LTX at 480p varied from 13.0 to 32.7 seconds across the three scenes. Natural cache/loading states were retained, and the data do not isolate the reason for that spread. The full paired CSV includes every pair, not just the fastest one.

See the matching 480p graph Matched 480p five-second clips: old and upgraded median render times are 63.9 and 23.9 seconds for H3 Turbo, 43.1 and 25.4 for native LTX, and 27.1 and 15.2 for LTX 8+3.

Direct 1080p is now a practical option to test

The original production idea was to audition shots at 480p, cut the scene, then finish selected material at a higher resolution. That remains worth testing. The new LTX measurements add another option: some starting footage may be quick enough to generate directly at 1080p.

Here is what “one minute” means in this dataset. The five-second LTX 8+3 clip finished in about a minute, including file preparation. The ten-second clip took about two and a half minutes to the finished file.

At 1080p, LTX 8+3 takes 53.2 seconds to render a five-second clip and 61.0 seconds until the file is ready. A ten-second clip takes 132.1 seconds to render and 145.5 seconds until ready.
Figure 2. Upgraded computer, LTX 8+3, three scenes per clip length. File-ready time adds upload, polling, copy, crop/trim and encoding. It excludes creative review and editing. [2]

This is not a promise that a minute and a half of drama arrives after eighteen minutes of rendering. A finished episode also needs still generation, voices, chosen performance ranges, retries, editing and sound. But it gives us a measured starting point for the next production experiment.

At ten seconds, LTX and H3 occupy very different time budgets

For this new matrix, every combination used all three packets. The table shows server render medians for ten-second output clips. These high-resolution figures are measurements of the upgraded setup; they are not matched old/new comparisons.

Recipe1080p render2K renderAdditional 2K render time
LTX 8+32m 12s2m 36s18.0%
LTX native 83m 44s4m 27s19.4%
H3 Turbo 413m 37s18m 07s33.0%
H3 Base 2064m 43s81m 32s26.0%
LTX ten-second clip medians: 8+3 takes 2 minutes 12 seconds at 1080p and 2 minutes 36 seconds at 2K; native 8 takes 3 minutes 44 seconds and 4 minutes 27 seconds.
Figure 3. LTX render time, in minutes, for ten-second clips. Three scenes per bar. [2]
H3 ten-second clip medians: Turbo 4 takes 13 minutes 37 seconds at 1080p and 18 minutes 7 seconds at 2K; Base 20 takes 64 minutes 43 seconds and 81 minutes 32 seconds.
Figure 4. H3 uses a longer horizontal scale than Figure 3 so its values remain readable. Both graphs use minutes and start at zero. [2]

“8+3” is a separate LTX pipeline: a smaller first pass is enlarged and refined in a second stage. Native 8 works on the target grid. The result is not a simple contest between eight and eleven equally expensive steps. H3 Turbo also uses an adapter, rather than merely lowering the base model’s step count. The tested H3 adapter is named for 768p; its 1080p and 2K results are exploratory uses of that recipe.

Here, 2K means 2048 × 1152, about 13.8% more delivered pixels than 1920 × 1080. It does not mean 2560 × 1440 or 4K. Higher-resolution generations are new performances, even with the same seed; they are not upscaled versions of the smaller clips.

For five-second clips, the upgraded 1080p medians were 53 seconds for LTX 8+3, 1m 22s for native LTX, 4m 19s for H3 Turbo and 19m 38s for H3 Base 20. Four-step H3 is therefore a different time proposition from the roughly one-minute LTX result.

The old computer was not incapable of 1080p

The historical record needs a little care. The old setup did complete some 1080p work. Its earlier ten-second Nina shot took 40m 10s of server time with native LTX and 6m 03s with LTX 8+3. The new matrix uses boarding, departure and pass shots, so those old times are useful context but do not support a controlled high-resolution speedup ratio.

The old ten-second 1080p H3 Turbo attempt recorded a GPU out-of-memory error. It did not produce a finished video. In the upgraded matrix, all three ten-second 1080p H3 Turbo cases completed. That expands the demonstrated set of working configurations; it does not make the scenes identical or guarantee every future job will fit. Historical outcomes

The old 1080p H3 Base 10 result must also stay separate from Base 20. Substituting one for the other would make a hardware comparison look cleaner than the experiment actually was.

A few frames, with the moving examples on their own page

These are frames at two seconds from actual five-second benchmark outputs. They are selected from the fastest completed take at the labeled resolution. They illustrate the footage; they are not a controlled sharpness comparison because the selected scenes and recipes differ.

Generated cart departure frame from the fastest measured 720p five-second take, LTX 8+3.
720p cart departure · LTX 8+3 · 25.4 seconds to render. Click through to watch. [5]
Generated pass close-up from the fastest measured 1080p five-second take, LTX 8+3.
1080p pass insert · LTX 8+3 · 53.1 seconds to render, 60.9 seconds to its finished file. [5]

The watch page has the fastest completed five-second example at each of the four resolutions, plus a joint-fastest H3 example. These are individual minima, not the medians used in the graphs. The complete results remain downloadable so the fastest examples do not stand in for the whole experiment.

What the memory readings do and do not tell us

For ten-second LTX jobs, the highest sampled use across the three scenes was:

Recipe and resolutionGPU memory usedSystem RAM used
LTX 8+3 · 1080p48.65 GiB39.16 GiB
LTX native 8 · 1080p47.69 GiB38.93 GiB
LTX 8+3 · 2K50.15 GiB42.28 GiB
LTX native 8 · 2K49.23 GiB42.11 GiB

These are machine-wide samples taken approximately every five seconds. They include other GPU/system use and may miss instantaneous peaks. GiB is the 1024-based reporting unit. These readings are not isolated model allocations, and the GPU and system columns must not be added together. [2]

The tested jobs left substantial room in the installed 128 GB of system RAM. They do not establish that 128 GB was necessary, that 64 GB would perform identically, or that two simultaneous render jobs would be a good idea. This matrix used one GPU render at a time.

The cost and daily-production target

My operating goal is five to ten finalized 90-second episodes per day. The GPU and RAM upgrade cost me a reported $11,750. Those facts make usable output, retry rates and directing time as important as the fastest individual render. The companion cost paper compares ownership, rental and API scenarios, and estimates the rendering portion of that daily target. No five-to-ten-episode production day or local-agent editing pipeline has yet been demonstrated.

What I would test next in the studio

At this stage, I am looking for useful pictures and a workable production pace. I do not need every benchmark take to be a finished scene. The next question is whether we can turn a small script into an edited vertical short with a predictable amount of trial and revision.

LTX 8+3 is the fastest high-resolution recipe in these measurements and a practical place to start that test. I also want to compare two finishing routes: generating directly at 1080p, and auditioning at 480p before upscaling selected takes. Upscaling preserves a chosen take’s action when timing is maintained; regeneration at a higher resolution can change the performance. Neither the upscale workflow nor its end-to-end cost has been measured here.

Sampled frames from the ten-second tests showed continuity and camera changes, and native LTX did not demonstrate a universal creative advantage over 8+3. That is a reason to judge the specific shot, rather than assume that more steps or more pixels must produce the preferred result. A full-motion quality study and an accepted-shot rate remain future work.

The prepared mini-production uses seven starting compositions for twelve story beats. Its script, stills and audio sequence can help us test the production process before a full episode batch. This white paper does not claim that a new episode, pilot storyboard or cast-voice pipeline has already been completed.

Data and methods

The earlier matched runs were recorded October 25, 2026, New York time; the upgraded matrix ran October 28–29. Four profiles × three packets × four resolutions × two clip lengths produced 96 matrix cells. Every cell has a completed output. An app restart interrupted the original first high-resolution attempt; its records remain preserved and its successful replacement has a separate attempt identifier. The completion count is not a claim that every attempt ran without interruption.

Delivery is exactly five or ten seconds at 24 fps. The graphs use server execution time, including cache-dependent model loading, inference, decoding and initial save. Finished-file time additionally includes reference upload, completion polling, copying and crop/trim/encoding. Setup, downloads, technical verification, creative review, editing and Dropbox copying are excluded. Cold/warm cache states and execution order were not randomized or standardized.

The current workflows receive one clean starting image and a written motion prompt. No old finished video was supplied as the input to these image-to-video tests. H3 used the retained INT8 checkpoint family, with Base 20 or the four-step Turbo adapter. LTX 2.5 used the retained NVFP4 distilled transformer with native 8 or the two-stage 8+3 recipe. Other model families, reference-video modes and visible-dialogue synchronization are outside this matrix.

Model-compatible generation sizes were cropped to delivery sizes. The 1080p raw generation was 1920 × 1088, delivered at 1920 × 1080; 2K was 2048 × 1152. The published clips are silent delivery copies; original generated audio is retained in the research archive and its generation cost remains in timing. These are landscape hardware tests serving a vertical-studio research goal, not a validated native 9:16 production benchmark.

The 18 matched pairs were checked against saved references, prompts, workflow graphs and receipts. The 48 high-resolution results were additionally audited for successful renderer histories, video header dimensions/frame counts, recorded timing/resource agreement and media checksums. Technical checks do not establish creative acceptance. The ten-second visual observations use six sampled frames per take, with one additional full-size inspection of a sampled frame; they are not continuous-motion or blind reviews.

Sources and downloadable evidence

[1] Matched before/after evidence: 18 paired measurements, CSV. Includes both server and finished-file times, scene/profile/seed, source identifiers, and reference/prompt hashes. All three packets are included. The plotted ratios are ratios of three-scene medians.

[2] Upgraded matrix: all 96 completed results, CSV · group medians, ranges and sampled memory, CSV · chart data and matching method, JSON. Public downloads omit personal filesystem paths.

[3] Hardware and implementation: sanitized method record, based on the recorded hardware/software snapshots, original protocol and retained model-file manifests. Earlier published RTX 5070 study.

[4] Historical 1080p context: selected earlier outcomes. This separately identifies the Nina scene, completed LTX timings and failed H3 Turbo attempt. No zero-second result or high-resolution matched speedup is assigned to a missing or failed output.

[5] Moving examples: fastest-take gallery · selection rule and file checksums. Media are existing benchmark outputs; no new generation was required to make this paper.

Prepared with AI assistance from retained local test records, with Elizabeth Ann West directing the research and publication. Evidence cutoff: October 29, 2026. These are early workflow measurements on one upgraded workstation, not a universal model ranking or an estimate of finished episodes per day.