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.
A few months ago I watched a generated video clip that looked excellent. Then I tried to reproduce it.
The prompt was public. The settings were not. The seed was missing. The model version was described as “latest.” The hardware was “a 4090.” The workflow was a screenshot with three nodes cropped out. The comments section had the usual mix of admiration and people asking why their own output looked nothing like it.
That experience is why this blog exists.
AI video is not short on impressive clips. It is short on test reports.
I spent fifteen years in industrial reliability and verification. In that world, a measurement without conditions is gossip. If someone tells you a bearing lasted 10,000 hours, you ask at what load, at what temperature, with what lubrication, and how many samples. If the answer is “it looked fine,” you do not have a result. You have a story.
Generative video has the same problem, just with better lighting.
A demo clip can be useful. It can show what a model is capable of under some conditions. But it cannot tell you whether the result is repeatable, whether it depends on a lucky seed, whether the workflow is stable across versions, or whether the failure you are about to hit is a model limitation or a configuration mistake. Those are different problems, and they need different fixes. A clip does not distinguish between them.
The configuration is part of the result. That is not a slogan. It is the minimum condition for treating a generated video as evidence rather than entertainment.
This site is called The Test Fixture because a test fixture is a controlled setup. It is the thing that holds the experiment steady so you can change one variable and know what changed. That is the standard I want to apply here.
The first model I am testing in public is LTX-2.3. I am starting there because it is an interesting system and because the gap between a good demo and a repeatable workflow is visible enough to study. But the method is model-neutral. The brand is model-neutral. The site is independent and unofficial. I do not represent a model developer, I do not sell official tools, and I do not have access to internal builds.
Here is what you will find.
Controlled runs. Tests where I change one variable at a time: seed, prompt structure, resolution, duration, motion instruction, sampler settings, runtime version. Each report will include the test question, the configuration, the observation, the interpretation, and a clearly labeled speculation section when I am guessing.
Failure analysis. Flicker, identity drift, object deformation, camera instability, prompt contradictions, failed renders, and the most useful category of all: the impressive clip that cannot be reproduced. A failure case is not an embarrassment. It is part of the result.
Test bench notes. Hardware, VRAM pressure, inference time, quantization, dependency conflicts, and the boring details that decide whether a workflow is practical. Boring details are usually where the truth lives.
Field notes. Evaluation philosophy, test design, responsible reporting, and lessons from outside the bench. Some of those lessons come from industrial verification. Some come from photographing industrial remnants around Denver, where the light changes and the structure does not. Some come from hiking at altitude, where the weather is a variable whether you planned for it or not.
What you will not find: official affiliation, insider claims, “secret settings,” guaranteed results, unlimited commercial use promises, or brand criticism without evidence. I will not turn one successful run into a universal conclusion. A successful sample is evidence. It is not a warranty.
I will also show failed runs. This is not a marketing decision. It is a testing decision. If I only publish the clips that worked, I am not publishing test results. I am publishing a highlight reel with footnotes.
Every model-specific article will carry an independent and unofficial notice. Every article will show a byline and publication date. When I get something wrong, I will update the article and note what changed. I will not quietly delete a failed test. The correction is part of the record.
That is an observation, not yet a conclusion. But it is enough to start.
If it can’t be reproduced, it’s a story—not a result.