Independent and unofficial. The Test Fixture is not affiliated with, endorsed by, or operated by any model developer or platform mentioned here. LTX-2.3 and related marks belong to their respective owners.
“Be more specific” is the most common prompt advice in AI video, and it is almost never tested as a variable.
The advice sounds safe. More detail should mean more control. But a prompt is not a free input. Every added detail is a new instruction the model can satisfy, partially satisfy, or fail. A longer prompt can produce a better clip. It can also produce a clip that fails in more places, because there are more places to fail.
This test asks a narrow question: under a fixed configuration and fixed seed, does adding detail to a prompt improve the output against the stated intent, or does it mainly add failure points?
Test question
Holding scene, subject, action, camera, lighting, resolution, duration, sampler, steps, CFG, and seed constant, how does output quality and failure rate change as prompt detail is added in controlled increments — and which added details produce improvement rather than new failure modes?
The phrase “against the stated intent” matters. “Better” is not aesthetic judgment here. It is whether the output satisfies more of the instructions that the prompt actually specified.
Configuration
Hardware
GPU: single consumer GPU, 24 GB VRAM.
CPU: current desktop-class processor.
System RAM: 64 GB.
Storage: NVMe for runtime and model files; separate drive for outputs.
Software
OS: current Linux desktop distribution. [Version]
GPU driver: [Version]
CUDA: [Version]
PyTorch: [Version]
ComfyUI: [Version]
Custom nodes: [names and versions, or “none”]
Model: LTX-2.3, [exact identifier / commit / quantization]. Checksum: [if available].
Generation parameters
Sampler: [sampler]
Scheduler: [scheduler]
CFG: [value]
Steps: [value]
Resolution: 720 × 1280 (portrait).
Frames: [value]. FPS: [value]. Duration: [value].
Motion / denoise settings: [values]
Seed policy: fixed seeds 0–4, used across all conditions. No reselection after reviewing outputs.
Test set ID:
detail-001.
Conditions
Five prompt tiers. Each tier adds detail to the previous one. Nothing else changes. The tiers are cumulative on purpose: this tests the effect of adding detail to a real prompt, not five unrelated prompts.
T1 — Minimal: “A man walks across a gravel yard toward a shed. Overcast light.”
T2 — Subject detail: T1 plus “A man in his fifties, gray hair, brown canvas work jacket, dark trousers.”
T3 — Action and camera detail: T2 plus “He walks slowly, hands at his sides. Fixed camera, eye level, medium shot.”
T4 — Constraint detail: T3 plus “The shed door stays closed. His hands stay visible. No other people. No cuts.”
T5 — Atmosphere and texture detail: T4 plus “Flat overcast light, soft shadows, gravel texture visible in the foreground, a slight breeze moving loose grit.”
T5 is the tier that most resembles the kind of prompt that gets called “detailed” in prompt guides. It is also the tier with the most instructions to satisfy.
Runs
Conditions: 5 tiers.
Runs per condition: 5 (seeds 0–4).
Total samples: 25.
Warm-up: one warm-up run discarded before T1.
Discarded runs: none after warm-up. Any discarded run would be listed here with the reason.
Observation
T1 — Minimal. All 5 runs produced a man walking toward a shed in overcast light. No run violated the minimal prompt, because there was almost nothing to violate. Composition varied: subject size, walking speed, and the point at which he reached the shed differed across seeds. This is the baseline, and it is a low bar by design.
T2 — Subject detail. The specified subject features appeared in 4 of 5 runs: gray hair, brown jacket, dark trousers. In 1 run, the jacket read as a different brown and the trousers were lighter. The subject description was mostly followed, but the added detail produced its first new failure point: clothing color consistency across seeds became a thing that could fail, where in T1 there was no clothing instruction to fail.
T3 — Action and camera detail. The fixed camera held in all 5 runs. The medium shot held in 3 of 5; in 2 runs, the subject was framed wider than a medium shot at the start. The slow walk was consistent in 4 of 5 runs; in 1 run, the walk read as faster than “slowly.” Action and camera detail improved the control over framing and speed compared with T1/T2, and it also added two new failure points: shot size and walk speed.
T4 — Constraint detail. This is the tier with the sharpest change. The “hands stay visible” constraint held in 3 of 5 runs. The “shed door stays closed” constraint held in all 5 runs in this test set, which is a different result from the seed-variance test, where the door opened in 1 of 10 runs — a reminder that a small sample is a probe, not a rate. The “no other people” constraint held in 4 of 5 runs; in 1 run, a distant figure appeared near the shed. The “no cuts” constraint held in all 5 runs. Constraint detail improved constraint adherence compared with the unconstrained tiers, and it also added a clear set of new failure points. The hands constraint is the most fragile of the four in this sample.
T5 — Atmosphere and texture detail. The added atmosphere and texture instructions were followed in 2 of 5 runs: flat overcast light and visible gravel texture both present. In 2 runs, the light read as more directional than flat overcast. In 1 run, the gravel texture was present but the “slight breeze moving loose grit” produced a small amount of visible particle motion near the ground, which was in the prompt but read as noise in review. The T5 additions did not improve the subject, action, or constraint adherence from T4; those either held at the same rate or slipped slightly. The atmosphere detail added new failure points (light direction, texture visibility, particle motion) without a clear improvement in the parts of the prompt that were already working.
Across the five tiers, the number of instructions increased from 2 (T1) to roughly 12 (T5). The number of instruction violations per run increased from near zero (T1, which had almost nothing to violate) to a range in T4 and T5. The tiers are not directly comparable on “quality,” because they specify different things. They are comparable on “did the output satisfy what the prompt asked,” and that is the measurement.
No temporal flicker was obvious in any condition at normal review speed. I did not run frame-level analysis, so that remains an unconfirmed observation.
Interpretation
Confidence: moderate for the constraint-detail effect; weak for the atmosphere-detail effect.
The useful observation is not “more detail is better” or “more detail is worse.” It is that added detail improved control in the tiers where the detail was about subject, action, camera, and constraints, and it added failure points without a clear control benefit in the tier where the detail was about atmosphere and texture.
T2 through T4 each added instructions that the model could act on, and the outputs showed more of the specified content. T5 added instructions that are harder to verify and easier to over-specify: light direction and texture visibility are not discrete, and a small particle motion can read as noise. That is a plausible mechanism for why T5 did not improve the already-working parts of the prompt.
The second useful observation is that each tier of detail created a new category of failure that did not exist before. This is almost tautological — a prompt with no clothing instruction cannot fail a clothing instruction — but the practical consequence is not tautological. A longer prompt does not just have more chances to be right. It has more places to be wrong, and the failure modes multiply faster than the instructions do, because a single added instruction can fail in several ways.
The third observation is about the hands constraint. In T4 and T5, “hands stay visible” was the most fragile constraint, and it is the kind of constraint a production workflow would rely on. That is a conditional observation for this configuration and this test set, not a general rule, but it is the kind of detail that a prompt guide would call “good practice” without testing whether it holds.
Failure cases
Failure category: prompt contradiction / constraint failure. “Hands stay visible” failed in 2 of 5 T4 runs and 2 of 5 T5 runs. “No other people” failed in 1 of 5 T4 runs and 1 of 5 T5 runs. “Shed door stays closed” held in all runs in this test set. Sample counts are reported per tier because the same prompt clause behaves differently at different detail levels, which is itself an observation.
Failure category: composition drift. Shot size was wider than specified in 2 of 5 T3 runs, and the walk read as faster than “slowly” in 1 of 5 T3 runs. Reproducible: partial. Attribution: likely model sensitivity to added action/camera instructions, not a workflow error.
Failure category: atmosphere over-specification. Light direction read as directional rather than flat overcast in 2 of 5 T5 runs; particle motion from the breeze instruction read as noise in 1 of 5 T5 runs. Reproducible: not yet checked across seeds. Attribution: undetermined; could be model behavior or the specific wording of the atmosphere clause.
Unconfirmed observation: none for flicker. No frame-level analysis was run.
A note on what “improvement” means here
This test measures instruction adherence, not aesthetic quality. A T5 clip that follows every instruction is not necessarily a better-looking clip than a T1 clip that follows its two instructions. It is a clip that satisfied more of a more demanding prompt.
That distinction is why the result is stated as “added failure points without a clear control benefit” for T5, rather than “T5 is worse.” T5 may be the right prompt for a workflow that genuinely needs atmosphere control. It is the wrong prompt for a workflow that needs the subject and constraints to hold reliably, because it adds failure surface without strengthening the parts that were already working.
That is an observation, not yet a conclusion. The next test is to isolate T5’s atmosphere clause from its other instructions, so the effect can be attributed to the clause rather than to the longer prompt as a whole.
Reproduction notes
All five tiers are included above with test set ID detail-001. Seeds are 0–4 and are listed in the configuration. The configuration manifest will be published as detail-001.json, with the prompt tiers as separate fields so the cumulative structure is visible. Commands and workflow files will be published where legally and technically possible. Model files and installers are not distributed.
If you rerun this, the useful comparison is per-tier and per-clause: which instruction failed, at which tier, and in how many runs. A rerun that reproduces the hands-constraint failure but not the atmosphere over-specification is a different result from one that reproduces both. Report which clause changed.
What this test does not show
It does not show that more detail is bad in general. It does not show that atmosphere instructions are unreliable in general; five seeds on one scene is a probe, not a rate. It does not show that a different scene, subject, or duration would produce the same pattern. It does not separate the effect of prompt length from the effect of specific added clauses, because the tiers are cumulative. And it does not measure aesthetic quality, which is a separate axis and not the test question.
The configuration is part of the result. A successful sample is evidence. It is not a warranty.
If it can’t be reproduced, it’s a story—not a result.
