Skip to main content
Uncensored — This model declares no content filter. SpicyAPI forwards the request without adding a platform content scan; legal and consent duties remain with the customer.

LTX-2 19B LoRA Review: Your Own Weights, Sound On

Pick LTX-2 19B LoRA when you bring your own camera or style file; check that it is under 2 GiB and trained for LTX-2, or skip it and use the plain build.

Create in StudioOpen the Playground

Spicy Index

57.1/ 100

Preliminary

Freedom Score

75.5/ 100

Content levels

Tested 6 of 6 probe types · 51 runs

  • Mild · The lightest prompts in the test

  • Moderate · Where filtered models start to refuse

  • Strong · Mid-scale; must be tested before a model is ranked

  • Very strong · Where most models start to hold back

  • Maximum · The top of the scale, scored in two halves: class A and class B

Price
$0.12 per 5-second clip
Median time
47 s
Success rate
Collecting data

LTX-2 19B LoRA is the same Lightricks audio-video model as the plain LTX-2 19B family, with one addition: a loras field that takes up to three LoRA files per request, each a direct download link and a strength from 0 to 4 (default 1). It runs on text-to-video and image-to-video, makes clips of 5 to 20 whole seconds at 480p, 720p or 1080p, and puts stereo sound on every clip. On 5 October 2026 we tested the ladder with the field empty and then tested the LoRA field on its own. With no file loaded, the build matched the plain model on the same seeds: L1 Suggestive (lingerie, a slipping strap), L2 Partial, L3 Strong and the consensual-restraint probe of L5 Extreme rendered on text-to-video, while L4 Intense (an intimate scene between two adults) and cinematic gore came back softened. In our 5 October blind test it scored 46.5 out of 100.

What LTX-2 19B LoRA made in our tests

Straight from our test runs. Copy a prompt or open it in the playground to try it yourself.

  • Text to video

    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.

    Try this model

  • Text to video

    Starting close on a man in his thirties at the helm of a sailboat, white shirt with rolled sleeves and sunglasses, the camera dollies out to reveal the whole boat heeling over on a deep blue sea at golden hour. Audio: wind in the sails, water slapping the hull, gulls.

    Try this model

  • Image to video

    The camera dollies in slowly toward the singer in the green velvet dress as she starts to sing into the vintage microphone, ending on a close-up of her face in the spotlight. Audio: her voice singing a slow ballad, soft piano.

    Try this model

  • Image to video

    The couple ride the yellow scooter along the coastal road, laughing, as the camera dollies out to reveal the cliffs, the village and the glittering sea. Audio: scooter engine, wind, her laughter.

    Try this model

  • Text to video

    The camera dollies in from the back of an empty concert hall toward a man in his forties in a charcoal suit playing a grand piano alone on the lit stage, ending close on his hands and face. Warm stage light, dark red seats. Audio: a slow romantic piano piece echoing in the hall.

    Try this model

  • Text to video

    Try this model

  • Text to video

    Try this model

  • Text to video

    Try this model

  • Text to video

    Try this model

Make your own with LTX-2 19B LoRA

Create in StudioOpen the Playground

The verdict

The LoRA tests are the reason to read on. The official Dolly In and Dolly Out camera LoRAs loaded at strength 1 and moved the camera as asked. A community anime-style LoRA turned a photoreal prompt into hand-drawn animation at strength 0.5, and the same prompt without the file stayed photoreal. A LoRA trained for a different base model, a Wan 2.1 file, loaded without any error and changed nothing visible. And three kinds of file failed outright: anything over 2 GiB, a gated download and a repository page link.

Pick it when the look or the camera move is the point and you already know which file you want. If you do not use a LoRA, LTX-2 19B gives you the same content results for less money.

Try text-to-video → · Try image-to-video →

Key facts

  • Pass rate: with loras empty on 5 October 2026, text-to-video delivered 4 of 7 ladder probes and image-to-video 2 of 7, identical to the plain build on the same seeds; none were refused.
  • Price: $0.018 per second at 480p, $0.024 at 720p, $0.036 at 1080p, the same on both tasks and the same whether you load zero, one or two LoRAs. A five-second 720p clip is $0.12.
  • LoRA limits: up to three files, each a direct download link under 2 GiB (2,147,483,648 bytes), strength 0 to 4, a fourth file is rejected at submit.
  • Sources: every scored run is in the LTX-2 19B LoRA test report, the scoring rules are on our review methodology page, and the full parameter list is on the LTX-2 19B LoRA model page.

What we liked and what held it back

What we liked

  • Official camera LoRAs work. Dolly In, Dolly Out and Dolly Left are 327 MB each and loaded at strength 1; a prompt that names where the move ends gets that ending.
  • A community style file changed the whole clip. At strength 0.5 the same prompt and seed gave a clear hand-drawn cel look, where the file-free control stayed photoreal.
  • Files stack. An anime-style LoRA plus the Dolly In file, both at 1, produced a clip that kept the style and the move.
  • It works on a photo too. On image-to-video the style file turned a photographic first frame into animation inside the first second.
  • The price does not move with the LoRA count. Zero, one or two files cost the same per second.
  • Sound on every clip. The same stereo track as the plain model, with no switch to turn it off.

What held it back

  • The 2 GiB limit rules out a lot of files. Many LTX-2 LoRAs on Hugging Face are 3.4 GB exports, and two of the official camera LoRAs (Jib Up and Static, 2.2 GB each) are also over the limit. They fail with "Failed to load LoRA".
  • A wrong-base file fails silently. A Wan 2.1 LoRA loaded with no error and no visible change. Run a with-and-without pair before you trust a new file.
  • Links must be direct and public. A repository page link and a gated download both failed with a LoRA error. Failed requests were not charged.
  • A strong style file can pull the picture away from your prompt. At strength 1 the anime-style file changed the set and the characters more than at 0.5.
  • The hard rungs behave like the plain model. L4 and cinematic gore softened on both tasks, and image-to-video also softened L2 and the restraint probe.
  • It costs more per second than the plain build. The LoRA slot adds about a quarter to the price.

What the clips look like

The gallery shows published camera-LoRA examples from the model page, plus stills from our own with-and-without runs, all clothed and safe for work.

Dolly In at a gala. A woman in a crimson gown at the top of a marble staircase; the camera pushes in until her face fills the frame. The move ends where the prompt said it would.

Dolly Out on a boat. The clip starts close on a man at the helm and pulls back to show the whole sailboat heeling on a deep blue sea.

Dolly In from a photo. A stage-singer photo starts wide and ends in a close-up at the microphone, with the person and wardrobe fixed by the still.

Dolly Out on a scooter. A couple on a yellow scooter, then a pull-back that reveals cliffs, a village and the sea.

Dolly In in a concert hall. From the back of an empty hall to the pianist's hands and face.

Anime style, with and without. Two stills from the same prompt and seed: photoreal without the file, hand-drawn cel animation at strength 0.5.

A Wan LoRA on this base. Two stills from one prompt and seed, with and without a Wan 2.1 file. Same room, same dress, same pose; the file changed nothing we could see.

Prompts that worked

Camera LoRAs respond to prompts that state the end of the move. Style LoRAs usually want their trigger phrase in the prompt, but in our control run the trigger text without the file did nothing, so the file does the work and the phrase only helps it.

1. A reveal with Dolly In at strength 1.

text
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.

Run this prompt on LTX-2 19B LoRA →

2. A pull-back with Dolly Out at strength 1.

text
Starting close on a man in his thirties at the helm of a sailboat, white shirt with rolled sleeves and sunglasses, the camera dollies out to reveal the whole boat heeling over on a deep blue sea at golden hour. Audio: wind in the sails, water slapping the hull, gulls.

Run this prompt on LTX-2 19B LoRA →

3. A move on a photo.

text
The camera dollies in slowly toward the singer in the green velvet dress as she starts to sing into the vintage microphone, ending on a close-up of her face in the spotlight. Audio: her voice singing a slow ballad, soft piano.

Run this prompt on LTX-2 19B LoRA →

Paste the Dolly In or Dolly Out link into loras for these. Weak: a style LoRA at strength 1 with a prompt that describes a different scene; the clip drifted from the prompt. Strong: strength 0.5 and a prompt that matches the file's subject.

How far it goes

We ran seven probes per task at 480p and five seconds on 5 October 2026, one run per probe, with loras empty and the same seeds as the plain build. Results matched: text-to-video delivered L1, L2, L3 and the restraint probe, and softened L4, gore and the probe with a man in frame; image-to-video delivered L1 and L3 and softened the other five. The page label is Uncensored: we add no content filter, and none of the files you load are reviewed. In our 5 October Freedom run it scored 75.5 out of 100.

A LoRA changes how far a probe goes. With the community anime-style file at strength 1, our L3 prompt came back as a featureless figure rather than the plain model's result, so a style file can soften a probe the base model delivers. We did not test adult-content LoRAs on this endpoint, and nothing here promises how a file you choose will behave.

Not tested yet: adult character LoRAs, files near the 2 GiB limit, three files at once, and strengths above 1 for camera files.

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, LoRA files included.

Loading a LoRA file

Each entry in loras is a path, which must be a direct download link, and a scale from 0 to 4 with a default of 1. Up to three entries are accepted per request, and an entry that fails to load fails the whole request without a charge.

We ran every case on 5 October 2026 and report what we saw.

  • Official camera LoRAs at strength 1: works. Dolly In and Dolly Out each landed the move we asked for on both text-to-video and image-to-video.
  • Community anime-style LoRA at strength 0.5: works. A hand-drawn cel look with the same prompt and seed; at strength 1 the stylisation was stronger and the set and characters drifted further from the prompt. Stacked with Dolly In, both held.
  • Wan 2.1 LoRA at strength 1: no effect. It loads, and the clip is not bit-identical to the control, but composition, wardrobe and motion were the same. Use files trained for LTX-2.
  • Files over 2 GiB: fail. The error is "Failed to load LoRA" with a size message. Look for a smaller export or a lower-rank version.
  • Gated files and page links: fail. A link that needs a login fails with an authorisation error, and a repository page link asks for the direct download link.

A practical routine: run one clip with loras empty, one with your file at 0.5, then raise the strength in small steps if the effect is too weak.

Alternatives

  • No LoRA, lower price? LTX-2 19B delivers the same ladder results from $0.0144 a second.
  • Built-in style presets instead of your own files? LTX 2.3 Spicy LoRA ships its presets with the endpoint; on 5 October its live Freedom score was 83.8.
  • A different base for your LoRA files? Wan 2.2 LoRA takes Wan-trained files; on 5 October its live Freedom score was 88.8.

How far LTX-2 19B LoRA goes

Each level, on each task, based on confirmed results. A level earns points only when the content actually appears, and a single run earns 60% at most.

Freedom points by level and task
LevelImage to videoText to video
MildThe lightest prompts in the test
4/4 delivered20.0 / 20 pts
4/4 delivered20.0 / 20 pts
ModerateWhere filtered models start to refuse
2/4 delivered15.0 / 20 pts
4/4 delivered20.0 / 20 pts
StrongMid-scale; must be tested before a model is ranked
4/5 delivered18.0 / 20 pts
1/6 delivered11.7 / 20 pts
Very strongWhere most models start to hold back
0/4 delivered10.0 / 20 pts
0/4 delivered10.0 / 20 pts
MaximumThe top of the scale, scored in two halves: class A and class B
Class A: 1/4 deliveredClass B: 0/4 delivered11.3 / 20 pts
Class A: 3/4 deliveredClass B: 1/4 delivered15.0 / 20 pts
Task score74.2 / 74.2Lowest: pulls the average downTested Oct 5, 202676.7 / 76.7Tested Oct 5, 2026

Tested Oct 5, 2026 to Oct 5, 2026

This site stays safe for work: results are written up as text, any image shown is pixel-censored, and test prompts are never published. We never test anyone who looks underage, real people’s likenesses or non-consent.

See every test record →See the uncensored leaderboard →

Run your own prompt and see how far it goes

Create in StudioOpen the Playground

How LTX-2 19B LoRA earns its Spicy Index

Each dimension below links to its full workings: the table row every point came from and why points were lost.

Public leaderboards 45% of quality
Not yet tested

No exact match on a tracked public leaderboard yet. How it works

Blind output test 35% of quality
46.5
Arena quality 20% of quality
Not yet tested

Counts once the model has 100+ blind votes and a 95% interval under 200 Elo. How it works

Capability 10.5% weight
64.8
Freedom Score 30% weight
75.5

Only our blind test covers this model so far, and its agreement with public human-vote leaderboards is limited on this board. The preliminary Index is ranked by its current score; independent evidence is still pending.

Open the full test report →LoRA support75.0 / 100

Capability, item by item

Points from the public scoring table. Upscaled tiers and limits we set ourselves don’t count.

Native max resolution1080p
25.0 / 25
Longest single clip20 s
11.0 / 20
Native audioYes
15.0 / 15
Reference inputsNo reference images · no video or audio
0.0 / 15
Frame controlFirst frame only
5.0 / 10
Task coverageImage to video, Text to video
5.0 / 10
Customization (LoRA)LoRA supported
3.8 / 5

What LTX-2 19B LoRA can do on SpicyAPI

Straight from the live model specification on SpicyAPI. Some creators offer more in their own apps; this is what you can call through our API today.

Capabilities by endpoint
FeatureImage to videoText to video
Resolutions480p · 720p · 1080p480p · 720p · 1080p
Clip length5–20 s5–20 s
Aspect ratios—16:9 · 9:16
Native audioYesYes
Seed controlYesYes
Negative promptNoNo
See every parameter and code sample →

How long it really takes

Measured on real jobs over the last 30 days, from request to finished result. Times include our own test runs.

Median time
47 s
90th percentile
89 s
Wait per second of video
5.4 s per output second
Success rate
Collecting data

A time appears once at least 10 completed jobs are in. Success rates count customer jobs only and stay hidden for every model until we have a full 30 days of traffic.

Models to consider instead of LTX-2 19B LoRA

What do you think?

Was this review useful?

Log in to vote.

Comments

Comments are checked automatically and by our team. We remove spam, offensive content, personal attacks and anything that names a real person without consent. Report a comment and we review it within one business day.

Tried this model? Share what worked. You need a free account to comment, so we can keep spam out.

LTX-2 19B LoRA FAQ

Try LTX-2 19B LoRA yourself

Generate in Studio in a few clicks, or set every parameter in the Playground and see the exact price before you run it.

$0.12 per 5-second clip

More reviews

See the leaderboards →
  • A woman in an emerald silk dress standing among large tropical leaves in a glass conservatory at dusk

    Wan 3.0

    Pick Wan 3.0 for cheap uncensored 30-second takes with sound from text, a still or references; skip it when a scene must stop where your script stops.

    Spicy Index
    87.7
    Freedom Score
    96.0

    Tested Sep 24, 2026

    Read the review
  • A woman in a flowing gold silk dress wading through the surf at sunset, the fabric streaming out behind her

    Seedance 2.5

    Pick Seedance 2.5 when a 30-second take must stay on script or follow a big reference cast; for uncensored clips on a budget, MiniMax H3 costs far less.

    Spicy Index
    81.4
    Freedom Score
    83.6

    Tested Sep 24, 2026

    Read the review
  • A dark room with a wall of tall windows over a glittering city at night, a single small figure at the glass

    Wan 3.0 Prime

    Pick Wan 3.0 Prime for faster text or reference jobs at 36% more; for image-to-video, base Wan 3.0 costs less and clears the same intimate-scene level.

    Spicy Index
    75.8
    Freedom Score
    80.7

    Tested Sep 24, 2026

    Read the review