Local AI Video Models on a DGX Spark: the Scene Renders, the Verb Does Not

One model makes video on this box: LTX-Video 2B. Same method as everything else here — every probe carries a written prediction, made before generating.

Four text-to-video probes, each targeting a temporal failure rather than a compositional one. All four failed, and they failed identically.

• the cat is sharp, holds eye contact for all 161 frames, and never pushes the mug
• the corgi never completes approach → impact → recoil
• the duck riding a robot vacuum is a yellow ball on a stick
• the sloth and squirrel are two generic rodents, no high-five

The pattern: the model renders the scene it was given and does not perform the verb. Appearance is often excellent — real forest depth of field, convincing glass reflections in a lobby, coherent bodies. The requested event simply doesn't occur.

Image-to-video is a different story

1 pass, 2 partial, 1 fail — and the reason is structural rather than a matter of quality.

Text-to-video has to invent a subject every run, so "is it the same cat at frame 90" has nothing to check against. Image-to-video starts from a still that has already been generated and judged, so drift is measurable against a control — and the 2B model never has to invent the subject at all.

The practical pipeline is FLUX for the still, LTX to move it.

Two findings that survived strict grading

Conditioned content is preserved; newly revealed regions are confabulated. The most informative clip of the day was a slow push toward an enamel shop sign. The letters do not melt — the conditioned sign stays correct the whole time it's visible. Then the camera pans, and the model invents a second sign in the newly exposed area, reading gibberish.

Rendered text is usually the least stable thing in a frame. Here it was stable exactly where it was conditioned and fabricated everywhere else. On this model, a static camera is a correctness decision, not an aesthetic one.

Materials and light are convincing; geometry is not. In the kitchen clip the wood grain, brushed steel and shadow falloff are all excellent — and the dishwasher juts out past the cabinet fronts, with cabinet doors of mismatched widths. I graded that clip "nearly right" on a first pass and was corrected. It's a plausible texture of a kitchen, not a kitchen.

That's the single most important habit for evaluating this model: grade on geometry and object identity, never on vibe. A single frame at a glance is precisely the altitude at which this model looks fine.

The length ladder never broke

One prompt, walked up the frame count until failure.

All four rungs cleared. The question was "max seconds before it OOMs or degrades — note which happens." The answer is neither. The ceiling reached is the vendor's documented maximum of 257 frames, not the hardware's: peak memory is 36.6 GB against the 121 GB a GB10 has, and cost per second of video is flat at 15.6–17.5s regardless of clip length. Longer clips are free in relative terms.

The part where I was wrong, again

An earlier round of video testing produced two confident conclusions about model limitations. Both were my bugs. By now this is the most reliable finding in the whole project.

"The model liquefies small repeated subjects." Round one turned six rubber ducks into a single amorphous yellow smear, and I wrote that up as a property of the model. It was a missing VAE decode parameter. At vendor settings the ducks render as distinct, well-formed rubber ducks with visible eyes and beaks.

"Extreme scale ratios do not survive motion." Round one had a tiny astronaut vanish from a sleeping cat's back, and I concluded the scale relationship couldn't hold through animation. It holds end to end. At frame 160 the astronaut is still there, still tiny, still pushing the mower.

Round one is retained in full as _void-round1/ and marked void rather than deleted, because the comparison is the evidence.

Here's why this keeps happening with generative models specifically: a model failing is indistinguishable from a harness misconfiguring it, because both produce plausible-looking wrong output rather than an error. With the coding models a harness bug at least showed up as a stuck score. Here it shows up as a picture that looks fine.

One probe did get genuinely worse at vendor settings: the raccoon shrug. Round one produced a legible shrug and even absorbed the still's spurious third arm; at 161 frames the raccoon lunges at the camera in heavy motion blur. Longer is not uniformly better, and reporting only the improvements would be its own kind of lie.

What this does not tell you

Every video statement here is about one small model. LTX-Video 2B has a 7.2 GB transformer — its text encoder is roughly three times larger than the video model itself. The 13B 0.9.7-dev checkpoint is downloaded and has never been run. No Wan, no CogVideoX, no HunyuanVideo. "Video models can't perform verbs" is not supported by this work; "this 2B model doesn't" is.

And six single-subject text-to-video prompts — a pug in a wizard hat, a self-playing piano, a traffic light on wet tarmac — were designed around what the model can hold together, and never run. The honest reading is that four adversarial prompts failed, not that the model can't do text-to-video at all.

Where the notebook stands

That's the end of this arc. Two boxes, nine coding models, three generative ones, a contract that took an hour to write and outperformed everything built around it, and a running tally of forty-four harness bugs against zero confirmed model limitations.

The question I bought the hardware for is still open. I know what a fix costs now — two-thirds of a cent, sixteen agents wide, on one of the two boxes. What I don't know is whether any of it survives contact with a real codebase, which by definition exceeds one context, and where the last from-scratch attempt destroyed 92% of its own output.

That's the next notebook.