Ten LTX-2.5 clips generated 2026-08-17 on one B200, against the H3 clips already on disk. Same 768×1344 at 24fps, same in-model audio, same concepts and verbatim dialogue. Clips are full quality — nothing re-encoded — so skin texture is judgeable. Unmute to check the audio; both models generate it in-model.
What to look for. The result splits by conditioning path, and that matters more than the price.
Image-conditioned (green tag): LTX is genuinely competitive — identity, set, wardrobe and props hold for the whole clip. H3 still runs cooler and more matte, which is the look the house prompts were tuned for, and it holds framing where LTX wanders.
Text-only (red tag): LTX drifts. In walk_p1 the
person changes and the lighting jumps mid-shot. H3's ref2va holds a
presenter across scene changes and LTX-distilled has no equivalent —
that gap, not the cost, is what decides adoption for recurring-cast work.
Best case for LTX. Identity, kitchen, wardrobe and props hold for the whole clip, and the comedy arc lands: amused → wide-eyed → the mock-trailer bellow → collapsing into laughter.
15s image-conditioned. Identity holds across all 361 frames. LTX pulls in closer and is more animated; H3 holds the reference framing more strictly.
Short control clip. Note LTX drifts off-prompt near the end — stops addressing camera and drops the earbud the prompt named.
THE FAILURE CASE. Watch the LTX side: the person visibly changes partway through, and the lighting jumps from cool daylight to hard golden sunset mid-shot. H3's ref2va holds one presenter and one lighting state for 15s.
Text-only bedroom. H3 uses ref2va to keep the same presenter in a new room; LTX has no equivalent, so this is a fresh person each time.
Text-only bathroom. Same structural gap as above.
Incidental win for LTX: it came out correctly PORTRAIT. H3's only version of this concept is 1344×768 LANDSCAPE — the ref2va + aspect_ratio:auto trap, a full-price unusable clip.
No H3 counterpart — never generated.
No H3 counterpart — never generated.
No H3 counterpart — never generated.
The only character-consistency adapter in the LTX family. Left is LTX-2.5 text-only on the same concept; right is Ingredients. It works — but it is a 5-second landscape model on the older 2.3 base.
The headline pair. LTX-2.5 (left) changes the person and jumps the lighting partway through 15s. Ingredients (right) holds one person and one lighting state. NOTE it is only 5s — a third the duration is a third the chance to drift, so this shows the mechanism works, not that it survives 15s.
Bedroom scene. Same presenter carried from the reference sheet into a room that is not in the sheet.
Bathroom scene. Note the softness — 0.34 MP against LTX-2.5's 2.09 MP at 1080p.
Forced to OUR house format: 9:16, 15s. The card calls this out of distribution and it shows. It runs without erroring — it degrades rather than fails — but this is why Ingredients is not adoptable as-is.
The audio is your own Gemini-3.1-flash TTS from the voice feed, passed through unmodified — the clip carries the actual Lissin voice, not a re-synthesis. Video prompts were written per clip to match each voice's director prompt and timbre. Left is with NO image conditioning, right conditions frame 0 on a photoreal gpt-image-2 still. Unmute these.
sports-01 · Play-by-Play Hype · voice Fenrir · excitable
Video built to peak where the audio peaks. Tests whether A2V reads energy out of the waveform.
true-crime-01 · Noir Detective · voice Algenib · gravelly
Deliberately near-static staging so the audio's long pauses have somewhere to land.
wellness-01 · Sleep Meditation · voice Vindemiatrix · gentle
Total camera stillness is the point — motion here would mean A2V is not reading the audio's pace.