← Blog
✎Blogbeitrag

Lighting-First Prompts Lost the Dog in Two Shots and the Scene in Two More

An empty brushed-steel chair in cold teal light with one warm amber lamp angled at its seat, dark floor, empty space above the chair.

Ad prompts that opened with about 170 words of light and lens lost the subject in four of six LTX shots: two had no dog, and two moved the scene outdoors or changed the breed.

Our first photoreal ad prompts opened with light and lens. In six shots, the dog was missing in two. In two more, the scene moved outdoors or the breed changed. Four of six shots failed on the subject, and the prompts were the only thing we had changed.

What the prompt looked like

The lighting-first prompts ran about 170 words, most of them describing light and the lens, with the subject arriving late in the paragraph. The pipeline now caps prompts at 200 words, as one paragraph, subject first. The cap comes from the model's own guidance, and our configuration records the failure as the reason for the rule.

We wrote the subject-first rule after the failure, not before. So the rule is a response to one batch, not a tested improvement.

Counting the failures

Our working log records the failure as 2 of 6 shots dropping the dog. The rules file records the same batch as 2 of 6 with no dog and 2 more where the scene moved outdoors or the breed changed. The two counts differ in scope. The wider count is the one we use here, because it is the one a viewer would notice.

The record does not say how the counts were made. They read as judgments by eye. No scorer or blind reviewer is recorded, so an edge case could go either way.

The fix we wrote

The rule now says: name the subject and the location in the first sentence, restate the subject once, and put the lighting after. An example in the rule reads: "the golden retriever is the main subject, centred and sharp". That sentence is a model instruction, not a finding.

What we did not do is rerun the six shots with the new wording. The rules file does not record the seeds for the original six shots, so an exact rerun is not possible from the record.

Why we did not use the other route

The other route was a reference image. Text-to-video gave a different dog in each shot. We tried an Ingredients LoRA with a sheet of the dog cut out on black, which is meant to keep one subject across shots. It was dropped when a split-screen artefact came back, and a frame-0 image route replaced it. So the subject problem has two fixes in our history, and they were not compared with each other.

What would count as a fix

The same six shots, the same seeds, and the subject-first prompts. A frame counts as holding the subject if the dog is present, of the right breed, and in the place the prompt names. A single good shot does not count.

What we are not claiming: that 200 words is the right cap, or that the subject-first sentence is the cause. Both changes arrived together, and we have not separated them.

How to prove this wrong

Keep the six original prompts, rerun each with a recorded seed, and count dogs, breeds and places in every frame. If the lighting-first prompts do not reproduce the four failures, the batch count is wrong. If the subject-first prompts lose the dog in any shot, the rule is wrong for that shot.

The open question is which of the two changes does the work, the word cap or the subject sentence. A test with each change alone would answer it, and we have not run that test.

Provenance

  • /home/mario/Brain/AssetsHQ/2026-11-19-winter-dog-ltx-commercial.md, lines 17 (the 69 to 81 second shot time, the 200-word prompt rule, and the 2 of 6 lighting-first count) and 18 (the text-to-video dog drift and the identity LoRA route) and 23 (the split-screen result that ended the LoRA route).
  • /home/mario/Projects/AssetsHQ/docs/rules.md, line 124 (the 2 of 6 no-dog and 2 of 6 drift count, and the subject-first rule with its example sentence).
  • /home/mario/Projects/AssetsHQ/pipeline.toml, line 69 (WORDS_MAX = 200, with the comment recording the 2-of-6 failure).

Kommentare

Noch keine Kommentare.

Anmelden um zu kommentieren.