LTX 2.3 Spicy Review: Twenty Seconds From One Frame, Top Rung Uneven

We don't publish the clips from our content-ladder runs; the gallery below is the everyday register, taken from the published, safe-for-work examples on the LTX 2.3 Spicy model page.

Twenty seconds, no prompt at all. A mirror-and-steam clip generated from the opening frame alone, at the endpoint's full 20-second length — the small-movement result you get by sending nothing but an image.

Rain on the glass, one continuous beat. The wet-shirt, neon-reflection clip runs the full palette of the opening still: streaked glass, one wiper sweep, one small turn toward camera.

Firelight on green velvet. A held, low-motion pose by a fireplace — hand, hem and head each move once, in sequence, rather than all at once.

A drive-in bonnet at dusk. Projector light washing across the frame in shifting blue, a blanket pulled up, headlights sweeping past — a portrait-frame clip built from a single still.

Steam clearing on a mirror. Two streaks wiped through fog resolve into a reflection; the towel stays exactly where the opening frame put it throughout the clip.

The still carries the content; the prompt, when you use one, should carry only the movement. Skip it entirely for a small plausible motion, or name one gesture and one camera note when a specific beat matters.

1. One gesture, no cuts (5 s, 720p).

She turns further toward the driver's seat, wet shirt clinging as she moves, pushes a soaked strand of hair back and mouths something with a tired smile. Rain streams down the windscreen, neon reflections crawl across the glass, wipers sweep once.

Run this prompt on LTX 2.3 Spicy →

2. A slow, sequenced pose change (5 s, 720p).

The fire surges and orange light flickers across her legs and the green velvet. She uncrosses her ankles and re-crosses them the other way, then slides her free hand slowly up the robe and draws the hem a little higher before letting it settle. She tips her head back against the chaise, eyes on the camera.

Run this prompt on LTX 2.3 Spicy →

3. No prompt at all. Send only image_url. The model infers a small, plausible movement from the still, which is often all a loop needs — the pattern behind our own no-prompt example above.

Weak: describing everything already visible in the frame. The model already sees the wardrobe and the room; repeating them in text wastes the prompt. Strong: prompt 2. One sequence of small, distinct actions, nothing the frame doesn't already establish.

We ran our five-rung content ladder on this family's one published task, image-to-video, on 25 September 2026, with an earlier restraint/impact check on 15 September. L1 Suggestive, L2 Mature themes and L3 Strong rendered clean on every run. L4 Intense rendered in 2 of 3 runs, with the third coming back with the requested action replaced. The L5 Extreme restraint/impact probe split evenly across the two dates: 2 clean, 2 softened. The L5 Extreme gore probe softened in all 3 runs — this endpoint never delivered a requested harmful action as asked; it substituted something milder every time.

Freedom Score: 89.2 of 100, pulled down almost entirely by the L5 Extreme gore result and the mixed L5 restraint/impact result.

The page label is Uncensored: we add no content filter on our side, and the model's own limits note in our documentation adds a specific warning worth repeating in ladder terms — a clothed figure prompted only for a sword wound came back at a stronger content rung than requested in two of three runs, and none of the three showed the wound as written. If your prompt is aiming at violence rather than intimacy, expect the model to substitute one for the other rather than deliver either cleanly.

Not tested yet: whether combining the loras field on the LoRA sibling changes how often the top two rungs soften; scenes with two women or two men; anime or illustrated styles; whether a closing instruction in the prompt reduces gore substitution.

Adults only, in every prompt, still and output. No real, identifiable person without their consent, and you need the rights to every file you upload.

• The same shape, with per-LoRA style control? LTX 2.3 Spicy LoRA adds a loras field for up to three of its built-in style and motion LoRAs per request, on the same 3-to-20-second, 480p-to-1080p contract.
• A top rung that held more consistently in our test? Wan 2.7 Spicy cleared every rung we ran, including both L5 Extreme probe types, though it caps at 15 seconds and costs more per second. Our Wan 2.7 Spicy review.
• An even cheaper still-to-video option? Wan 2.2 Spicy matches the $0.019-a-second 480p price but only accepts exactly 5 or 8 seconds.
• More resolution tiers on a single still? MiniMax H3 Spicy covers 480p through 1080p from $0.038 a second, 3 to 15 seconds per clip.