LTX-2 19B and LoRA Camera Moves: Sound, Dolly Shots and Limits Tested

LTX-2 19B makes a 5 to 20 second clip with a stereo sound track already in it, and the LoRA build lets you load up to three LoRA files to direct the camera or restyle the whole clip. In our tests on 5 October 2026 the official Dolly In, Dolly Out and Dolly Left camera LoRAs worked at scale 1, a community anime-style LoRA changed the look at scale 0.5, and any LoRA file over 2 GiB failed to load. Nothing was refused or filtered, and image-to-video held back more than text-to-video on our adult-content ladder.

LTX-2 19B comes from Lightricks and is an open audio-video model that generates picture and sound together. We ran the base model through the LTX-2 19B family and the LoRA build through the LTX-2 19B LoRA family. The prompts below are everyday scenes with clothed adults; our content probes are summarised by level, not quoted.

At a glance

Sound on every clip

The headline feature is that every clip arrives with an audio track. You steer it inside the prompt, usually with a final sentence that starts with "Audio:". We confirmed on every output that the file carries AAC stereo at 24 kHz. The publisher lists a limitation: audio without speech may be of lower quality, and prompt following depends heavily on prompting style.

Here is a text-to-video clip built around sound:

LTX-2 19B, text-to-video, 16:9, 720p, 6 seconds, seed 201. Frames at the start, middle and end. The slow push-in happens as asked.

A woman in her thirties in a gold sequined gown sings a slow jazz ballad into a vintage microphone in a smoky lounge, eyes half closed, a piano player blurred behind her. Warm spotlight, slow push-in. Audio: her soft, husky voice singing, piano chords, glasses clinking in the background.

We judged picture and structure frame by frame. We did not score audio quality, so treat any promise about a voice or a mix as something to check on your own clip.

A second example puts the sound in the action:

Text-to-video, 16:9, 720p, 5 seconds, seed 202.

A bartender in his forties with a trimmed grey beard, black shirt with rolled sleeves, shakes a cocktail shaker hard behind a dark wooden bar, then pours an amber drink into a coupe glass and slides it toward the camera with a grin. Amber backlight through bottles. Audio: ice rattling in the shaker, liquid pouring, low bar chatter and soft music.

The shake, the pour and the grin all happened in order. The bottle labels are decorative rather than readable.

Size, length and speed

Text-to-video takes a 16:9 or 9:16 ratio and builds the frame from that and resolution. Image-to-video follows the photo: a 16:9 photo gave 832 x 448, 1216 x 704 or 1856 x 1088, a 2:3 photo 512 x 768 at 480p.

Text-to-video, 9:16, 720p (704 x 1280), 5 seconds, seed 203, with the sound of waves and a laugh requested.

That speed is the other half of the story: LTX-2 19B returned a 720p clip in seconds, where a heavier model like HunyuanVideo 1.5 takes minutes. Draft at 480p, check the sound and the framing, then rerun the keeper at 720p or 1080p. The family pages show the current per-second price for each resolution.

Image-to-video: describe what happens next

The first frame of an image-to-video clip is your photo. The danger is a prompt that describes a different place. In our test a 2:3 portrait of a singer on a stage was paired with a prompt about a bed, and the clip cut to a bedroom shot after the first frame. The model followed the words and left the photo.

The fix is to describe what happens next in the picture:

Image-to-video, 720p, 5 seconds, seed 204, static camera.

The woman grips the barbell, lifts it in one clean deadlift, holds it at the top for a breath and lowers it back to the floor, chalk dust drifting in the light. Static camera. Audio: her sharp exhale, the plates clanking as they touch down, the hum of the gym.

The setting, her outfit and the light stayed. The movement did not match the word we used: she pressed the bar overhead instead of a floor deadlift, so name the motion in plain physical steps (grips, lifts to the chest, presses overhead) when it matters.

Image-to-video, 720p, 6 seconds, seed 205: "strums the acoustic guitar and hums a gentle melody, tapping his bare foot... Slow push-in. Audio: warm acoustic guitar chords, his quiet humming, crickets starting up."

Image-to-video needs a prompt: unlike some models, both image_url and prompt are required.

Camera LoRAs: the Dolly moves

A prompt like "the camera dollies in" is a request. A camera LoRA makes it a habit. Lightricks publishes official camera-control LoRAs for the LTX-2 19B model: Dolly In, Dolly Out, Dolly Left and Dolly Right, plus Jib Up, Jib Down and Static. In the LTX-2 19B LoRA family you pass a direct download link to the .safetensors file in the loras array:

{
  "prompt": "The camera dollies in toward a woman in her thirties in a crimson evening gown standing at the top of a marble staircase at a gala, until her face fills the frame as she turns and smiles. Chandeliers, warm light. Audio: string quartet, murmuring crowd.",
  "aspect_ratio": "16:9",
  "resolution": "720p",
  "duration_seconds": 5,
  "seed": 301,
  "loras": [
    {
      "path": "https://huggingface.co/Lightricks/LTX-2-19b-LoRA-Camera-Control-Dolly-In/resolve/main/ltx-2-19b-lora-camera-control-dolly-in.safetensors",
      "scale": 1
    }
  ]
}

Dolly In, Dolly Out and Dolly Left are 327 MB each and loaded without trouble at scale 1. Describe where the camera should end up, as the prompt above does with "until her face fills the frame".

Dolly In LoRA at scale 1, text-to-video, 720p, 5 seconds, seed 301. The camera ends on her face, as the prompt asked. A man at the left of the first frame was our model's addition.

Dolly Out LoRA at scale 1, text-to-video, 720p, 5 seconds, seed 302: "Starting close on a man... the camera dollies out to reveal the whole boat heeling over on a deep blue sea."

A third shot on a concert stage starts from the back of the hall and pushes in to the pianist:

Dolly In LoRA at scale 1, text-to-video, 720p, 5 seconds, seed 307.

The move is the strong part. Small things drift: the lettering on the piano came out garbled.

Which direction for which job

Image-to-video with a camera LoRA

The photo fixes the person and the wardrobe, and the LoRA fixes the move:

Dolly In LoRA at scale 1, image-to-video, 720p, 5 seconds, seed 304. The first frame is the photo; the clip ends on a close-up of her face.

Dolly Out LoRA at scale 1, image-to-video, 720p, 5 seconds, seed 305.

Two other clips show where a sideways move gets loose:

Dolly Left LoRA at scale 1, text-to-video, 9:16, 720p, seed 303. The camera slides left as asked, and the cast at the end is not the cast we started with.

Dolly Left LoRA at scale 1, image-to-video, 720p, 5 seconds, seed 306. We asked him to keep strumming; the camera slid away and he left the frame.

The takeaway: a camera LoRA makes the move reliable, not the subject. When the subject must stay, ask for a shorter move or keep the action simple.

An anime-style LoRA, stacking and scale

We also tried a community anime-style LoRA (428 MB) with a trigger phrase at the start of the prompt, at the same seed and prompt each time.

Start a style LoRA at 0.5 and raise it in small steps. Put the trigger words from the LoRA's model card at the very start of the prompt, then send the same loras array with every prompt in a series so shots match. Combine a style with a camera file at modest strengths; up to three LoRAs fit in one request, and the number of LoRAs does not change the price. We kept the style frames out of this page, since they are test material rather than finished clips.

Why a LoRA fails to load

The loras field takes a direct download link to one .safetensors file. Everything else we tried failed in a predictable way, and a failed request is not charged:

The last row is the dangerous one, because nothing tells you it did nothing. Run a with-and-without pair at the same seed. If the two clips match in composition, wardrobe and motion, the file was trained for another model. When a LoRA is too large, look for a smaller export of the same LoRA, or one trained at a lower rank.

How far it goes

Nothing filters the prompt, the LoRA files or the result. We ran the same seven-probe ladder as for the other models, at 480p and 5 seconds, all subjects adults. For image-to-video the starting photos were clothed adults. The LoRA build with no LoRA loaded matched the base on the same seeds. We summarise results by level and do not print the probe prompts.

Zero refusals, zero errors, no blurring. Text-to-video delivered 4 of 7 probes and image-to-video 2 of 7, so image-to-video held back more. With the anime-style LoRA loaded, the L3 probe came back as a featureless figure, so a style LoRA softened it further. One probe per cell is a sample, not a rate.

Where it falls short

• Image-to-video cuts away when the prompt describes another place.
• Subjects drift under strong camera moves, in clothing and in who is in frame.
• LoRA files over 2 GiB fail, and several popular ones are that size.
• A LoRA for the wrong base fails silently.
• Sound is prompt-steered and unscored by us; the publisher flags lower quality for audio without speech.
• No audio switch. Every clip carries a track.
• Image-to-video is gentler on adult content than text-to-video.

LTX-2 19B or the LoRA build

Start with the base model, and move to the LoRA build when a prompt keeps missing the camera move or the style. If you need a heavier, silent model for faces, see our HunyuanVideo 1.5 test, and for a general approach to LoRAs see running your own LoRA through an API.

FAQ

Does LTX-2 19B generate sound?

Yes, every clip has a stereo AAC track at 24 kHz and there is no way to switch it off. You steer the sound with an "Audio:" sentence in the prompt, and the publisher notes that audio without speech may be of lower quality.

How long can a clip be?

Any whole second from 5 to 20, at 480p, 720p or 1080p. A 20-second 480p clip came back in 36 seconds. Requests for 4 or 21 seconds are rejected at submit.

Why did my image-to-video clip cut to a different scene?

The prompt described a different setting from your photo. In our test a stage portrait plus a prompt about a bed cut to a bedroom. Describe what happens next in the picture instead.

Which camera LoRAs worked?

The official Lightricks Dolly In, Dolly Out and Dolly Left LoRAs, 327 MB each, loaded and moved the camera at scale 1. Jib Up and Static are 2.2 GB, above the limit.

What is the LoRA file size limit?

2 GiB, or 2,147,483,648 bytes. Larger files fail with "Failed to load LoRA", and a failed request is not charged. Link to the .safetensors file directly, not to the repository page.

Can I stack several LoRAs?

Yes, up to three per request, each with its own scale from 0 to 4. A style LoRA and the Dolly In camera LoRA, both at 1, worked together in our test and the style held.

How do I know my LoRA actually did something?

Run the same seed and prompt with and without it. A LoRA trained for a different base model loaded without an error in our test and changed nothing visible.

Is LTX-2 19B uncensored?

Yes, nothing filters the prompt, the LoRA or the result. In our ladder text-to-video delivered 4 of 7 probes and image-to-video 2 of 7, with no refusals, and the model itself softened the rest.