A Diagnostic Tree for Generations That Come Back Wrong

Three generations in, still wrong, and the reflex is to rewrite the whole prompt and roll again. That's usually the expensive move. A generation that comes back wrong is giving you information — which part of the input mattered and which part the model ignored — and reading that signal correctly is the difference between a two-credit fix and a fourth full reroll.

The instinct to blame yourself is backwards, too. Even a strong model gets it wrong more often than the marketing implies. On VideoPhy, a benchmark built specifically to check whether generated video obeys the physics its caption implies, the best-performing model tested satisfied caption adherence and basic physical commonsense together on only 39.6% of instances — not "close enough," fully-correct-on-both-counts on fewer than two out of five tries. If you're rerolling and still hitting a wall, you're not doing it wrong. You're running into the category's actual failure rate, and the fix is a better diagnosis, not more attempts.

What follows is a diagnostic tree: start from what you observe in the bad output, not from what you originally asked for, and let the observation tell you which input to change.

Start from the symptom, not the intent

The reflex when a generation is wrong is to reread your own prompt and ask "what did I mean?" That's the wrong starting question, because your intent isn't what the model responded to — your prompt, your reference image, your chosen model, and (for video) the physical and temporal behavior it happened to render all were. Four different inputs, four different fixes, and they don't share a remedy.

The tree branches on one question first: is the problem what's in the frame, or how it behaves?

Branch 1: The subject or scene is wrong

If the output shows the wrong thing entirely — a wrong object, a wrong setting, a wrong count of something, a detail your prompt named that just isn't there — that's a prompt adherence failure, not a rendering failure. The model didn't misfire on execution; it didn't parse your intent from the words you gave it in the first place.

The fix lives in the prompt, not the generation settings:

• Rewrite the sentence naming the missing or wrong element more concretely — trade an abstract noun for a specific, visualizable one.
• If the prompt was already detailed and specific, the mismatch may be the model itself. Some models are simply stronger at literal instruction-following than others, and switching is often faster than five more prompt edits.
• For anything layout- or composition-heavy — multiple named elements, specific spatial relationships — consider whether you're asking one model to do two jobs (subject fidelity and scene composition) that a different model splits better.

Branch 2: The scene is right, one region is broken

This is the "why do I have to reroll an entire eight-second clip for one bad half-second" situation, and it doesn't need a full reroll. If everything else in the frame is correct and only a specific, boundable region is wrong — a hand, a sign, a background element — the efficient fix is local, not global.

For video specifically, LTX 2.3 Retake exists for exactly this: it takes the source video's URL plus a prompt describing the desired change, and optional start_time, duration, and retake_mode parameters to regenerate only that window and blend it back into the untouched footage on either side. You're not paying for the ninety percent of the clip that already worked.

Branch 3: The scene is right, the physics or motion is wrong

This is the VideoPhy failure mode specifically — a floating object, liquid that doesn't pour like liquid, weight that doesn't load correctly onto a surface, motion that stutters or reverses illogically. It's the hardest branch to fix with a prompt edit alone, because you're not missing an instruction — you've hit a category the model wasn't good at simulating to begin with.

Two moves actually help here, roughly in order:

1. Simplify the physical interaction the prompt asks for. Every additional physical claim in one shot — pouring and splashing and catching light correctly — compounds the odds that at least one of them breaks.
2. Switch models before you rewrite the prompt again. Physical plausibility varies more by model than by prompt wording. A different model on the identical prompt is a legitimate diagnostic step, not a cop-out — compare options on /models rather than assuming the one you started with is the right tool for a physics-heavy shot.

Branch 4: You genuinely can't tell what's wrong

Sometimes the output is subtly off and you can't articulate why. It "feels" wrong but doesn't cleanly match any of the three branches above. Don't guess — ask for a structured read before spending another generation on a guess.

A real Versely walkthrough: point the agent at the generation and ask something like "Does this actually match what I asked for?" — that routes to the same capability behind getting AI feedback on a generation. It runs a vision check against your original prompt and returns whether it matches, exactly what's wrong, a severity rating (ok / minor / broken), and a suggested fix. A "minor" verdict with a named issue tells you which branch above you're actually in; a "broken" verdict on something that looked fine to you at a glance is worth a second look before it ships. This is a lightweight check, not a new generation — the right first move when you're stuck, not the fourth reroll.

Putting the branches together

A realistic sequence: generate, get a result that's off, run the feedback check if the problem isn't obvious, then act on exactly one branch — rewrite the prompt (Branch 1), retake the specific segment (Branch 2), simplify the ask or swap models (Branch 3). What you don't do is all three at once. Changing the prompt and the model and asking for a segment retake in the same pass tells you nothing about which change actually fixed it — which means you've learned nothing useful for the next generation in the same style.

Before you publish, check once more

Even a generation you're satisfied with is worth a last pass before it goes out, especially if you fixed it via a targeted branch rather than a clean first take — a segment retake can occasionally leave a faint seam at the boundary it blended into. Run it through the same pre-publish review step you'd use on any clip, whether it came out right in one pass or got stitched together across three fixes.

When to stop the tree and start over

The tree has a bottom. If you've worked through two branches and the output is still wrong in a new way each time, that's a signal the prompt is asking for something at the edge of what any available model does well — not that you haven't found the right branch yet. At that point, changing the ask itself (a different angle, a shot that avoids the hard element entirely) usually beats another round of diagnosis. Knowing which branch to try is what saves the first three generations. Knowing when to stop trying branches is what saves the tenth.