Best Open-Source NSFW Video Models in 2026: MiniMax H3 vs Hunyuan vs Wan vs LTX

MiniMax H3 is the strongest open-weights video model we host that you can call uncensored: the H3 LoRA build scored 68.7 in our blind test and 98.3 on our Freedom score, with zero refusals in 60 content-ladder runs. Of the open models added on 5 October 2026, HunyuanVideo 1.5 and Wan 2.2 LoRA tie on Freedom at 88.1 and both return silent clips. LTX-2 19B is the cheapest of the eight families here and the fastest we measured, and it makes sound, but the LTX-2 19B pair held back the most and scored lowest in the blind test.

If you come from ComfyUI, "open source" means downloadable weights: a community that trains LoRAs, and no hosted app's filter between prompt and result. This comparison covers the uncensored text- and image-to-video families in the SpicyAPI catalogue that run publicly released weights as published. Apart from the publishers' model cards, it uses our own data: the Freedom ladder, our blind test, the live catalogue and a same-prompt run on 6 October 2026. For every uncensored video model, open or not, read Best Uncensored AI Video Generator in 2026.

At a glance

Rows are ordered by blind-test score. Scores are from our live index on 6 October 2026 and move as we add runs. The blind test is our side-by-side quality vote out of 100; Freedom is how much of our content ladder a model delivers, out of 100. Specs and prices (lowest tier) come from the live catalogue.

\*Plain MiniMax H3 delivered all 48 of its scored runs. Its Freedom score is low because task-and-level cells we have not run yet count as zero, not because anything was refused.

Every family takes text or a photo, except Singularity, which has no text-to-video; the three MiniMax families also take reference images. Only HunyuanVideo and Wan 2.2 LoRA take a negative prompt, and only on text-to-video.

Quick picks:

• Widest range, your own LoRAs, sound included: MiniMax H3 LoRA.
• A silent single-person clip for the least money: HunyuanVideo 1.5, or Wan 2.2 LoRA image-to-video.
• The cheapest draft with sound, back in seconds: LTX-2 19B.
• One unbroken 20-second take up to 4K: LTX 2.5.

How far each one goes

Every family here is labelled Uncensored: the platform adds no content filter to the prompt, the image or the result. What a clip shows is decided by the model's own training, so the same label hides real differences. Our ladder has five levels:

Each cell is delivered runs out of runs. Every other run came back softened: the clip kept the setting and mood but did not draw what the probe asked for. Across all eight families there was not one refusal. All subjects were adults, and image-to-video runs started from clothed photos.

What that means for a project:

• L1 to L3 is safe ground on every MiniMax, Wan and Hunyuan build. Only the LTX-2 19B pair softened some L2 and L3 runs.
• L4 is where the field splits. The H3 builds and LTX 2.5 drew the two-adult scene almost every time. HunyuanVideo softened all 8 tries and Wan 2.2 LoRA 5 of 7, so plan for a suggestion of the scene on those two.
• Gore is its own question. HunyuanVideo draws it (6 of 8); Wan 2.2 LoRA mostly turns it into a few cuts (1 of 7).
• An adult LoRA moves the line on H3. With a community adult LoRA loaded, H3 LoRA delivered our L4 probe on all three endpoints.

The H3 LoRA and Singularity rows come from 26 and 27 September; the others include runs up to 5 October. A handful of runs per cell is a sample, not a rate.

Same prompt, four open models

Scores do not show you a model's look. On 6 October 2026 we sent two clothed, suggestive prompts to the three open models added on 5 October (HunyuanVideo 1.5, LTX-2 19B and Wan 2.2 LoRA) plus the top scorer, MiniMax H3 LoRA: text-to-video, no LoRA, 480p, 16:9, 5 seconds, seed 61006, no negative prompt. Only the model changes.

The first prompt is one person and one beat of action:

An adult woman in her early thirties with dark curly hair, wearing a rust-coloured satin slip dress under an open oversized tweed blazer, sits on the arm of a worn leather chair in an old map library at night. She closes the large map book on her lap, lets the blazer slide off one bare shoulder and looks up at the camera with a slow, knowing smile. Green-shaded reading lamps and warm tungsten light, dust drifting in the air, slow push-in, 35mm lens, film grain, cinematic, photorealistic.

Same prompt and seed, text-to-video, 480p, 5 seconds, frame at 4.3 s. Top left HunyuanVideo 1.5, top right LTX-2 19B, bottom left Wan 2.2 LoRA, bottom right MiniMax H3 LoRA.

Three of the four did what was asked. HunyuanVideo and Wan 2.2 LoRA slid the blazer off and ended on a smile to camera, Wan with the book and a wall map exactly as written. H3 moved her to a reading table but played the beat cleanly, in the most detailed room of the four. LTX-2 rewrote the wardrobe: a strapless dress, the blazer gone and her head tilted back instead of a smile. That fits its blind score: quick and atmospheric, loose with details.

The second prompt puts two adults close together, where video models usually come apart:

A tailor's atelier after hours. An adult woman in her thirties in a fitted black silk wrap dress stands on a low fitting platform; an adult man in his thirties with rolled shirtsleeves and an open waistcoat stands behind her and adjusts the dress at her waist. She turns her head toward him until their faces are almost touching, and neither of them looks away. Warm work lamps, bolts of fabric and dress forms in soft focus, slow push-in, 50mm lens, film grain, cinematic, photorealistic.

Same prompt and seed, text-to-video, 480p, 5 seconds, frame at 4.3 s. Same order: HunyuanVideo 1.5, LTX-2 19B, Wan 2.2 LoRA, MiniMax H3 LoRA.

Only H3 put his hands at her waist and brought their faces together as written. HunyuanVideo got the closeness but gave it to him: he leaned in cheek to cheek while she looked at the camera. Wan 2.2 LoRA had her turn to him, his hand on her shoulder. LTX-2 framed the pair wide and apart. One seed per model is a single sample, but it lines up with the blind-test order.

Two more findings. Sound: H3 LoRA put an audible track on both clips unasked; LTX-2's library clip came back near-silent and its atelier clip carried only a quiet track, so always write an "Audio:" line on LTX-2. Speed: model time for a 5-second 480p clip was 7 to 23 seconds on LTX-2, 27 to 38 on H3 LoRA, 38 to 70 on Wan 2.2 LoRA and 39 to 144 on HunyuanVideo.

Why open weights matter for NSFW work

Open weights do not make a model uncensored by themselves: LTX-2 19B is fully open and still softened more than half of its 50 runs. What they change is who decides.

No hosted app in the way. A publisher serving its own model usually adds its own policy. The plain Wan 2.2 family in our catalogue carries the publisher's own policy behaviour: its text-to-video refused all 19 of our ladder runs, from L1 up, for a Freedom score of 43.1. Wan 2.2 LoRA runs the released weights and refused none of its 43 runs. H3 Max, which is not among the checkpoints on MiniMax's model card, is labelled Filtered on text-to-video, while every open H3 build is Uncensored.

Community LoRAs. Public weights get adapters: styles, characters, camera moves, adult fine-tunes. On the LoRA families you pass a direct link to the .safetensors file and it loads per request, with no upload and no review of its contents. Here is a published Wan 2.2 LoRA example with and without its LoRA, same prompt and seed:

Wan 2.2 LoRA text-to-video, 720p, 5 seconds, seed 510531. Left: no LoRA. Right: a community vintage film-colour LoRA at scale 1, trigger words first in the prompt.

The wrong base fails quietly. On Wan 2.2 LoRA, a LoRA trained for another base model loaded without an error and produced a file identical to the no-LoRA run; on LTX-2 19B LoRA a Wan 2.1 file changed nothing visible. Run the same seed with and without a file before you trust it.

Community knowledge carries over. Trigger words and prompt habits from model cards and forums apply here. One commenter on the H3 open-weights launch thread called it the "Least-censored model of any type I've ever seen from a major lab." Our ladder agrees for the H3 builds.

The families one by one

MiniMax H3 LoRA, Singularity and plain H3

MiniMax released H3 as open weights in August 2026, a 33-billion-parameter model that makes picture and stereo sound together, per its model card. The LoRA build runs those weights with up to three LoRA files per request, has the best blind score here and delivered L4 in 11 of 13 runs with no LoRA loaded.

MiniMax H3 LoRA text-to-video with a community VHS-tape LoRA at scale 1, trigger word first, 480p, 4:3, 5 seconds. The tape look comes from the LoRA, not from editing.

Singularity is a community fine-tune of H3 trained for sharper fast motion and cleaner distant faces. It delivered all 38 ladder runs, the only 100 here, but has no text-to-video and took about 80 seconds per 5-second 480p clip, against 25 to 50 on H3 LoRA.

MiniMax H3 Singularity LoRA image-to-video, 768p, 5 seconds, realism LoRA at 0.6: one spin-and-cut beat with a whip-around camera.

Plain MiniMax H3 has no LoRA field but costs well under half as much per second at 768p ($0.037 against $0.10) and adds 2K. With no LoRA to load, start there.

HunyuanVideo 1.5

Tencent released HunyuanVideo 1.5 in November 2025, an 8.3-billion-parameter model built for consumer GPUs. It ties Wan 2.2 LoRA on Freedom and is the strongest of the 5 October additions at gore. Its strength is one person on screen: natural skin, and a face that performs on image-to-video.

HunyuanVideo 1.5 image-to-video from a 16:9 still, 720p, 5 seconds, frames near the start, middle and end.

Skip it for L4 scenes, which it softened every time, and for anything that needs sound. Our HunyuanVideo 1.5 test has the prompts and sizes.

Wan 2.2 LoRA

Alibaba's Wan 2.2, released in July 2025, splits generation between a high-noise expert that sets layout and motion and a low-noise expert that refines detail. The LoRA build exposes that split as three slots, which matters because community Wan 2.2 LoRAs usually ship as a high-noise and low-noise pair. Image-to-video costs $0.024 per second at 480p ($0.0204 with a 15% offer running at the time of writing); text-to-video, added on 5 October, costs $0.048. Our text-to-video test and slot guide go deeper.

LTX-2 19B and its LoRA build

Lightricks released LTX-2 in January 2026 as an audio-video model that makes picture and sound in one pass. It is the price and speed pick, with a 720p clip in under half a minute in our runs, and the one that held back the most.

LTX-2 19B text-to-video, 720p, 6 seconds, with an "Audio:" line listing wheels, wind, a crossing bell and a horn in order.

The LoRA build takes up to three files, including Lightricks' official camera LoRAs. Files over 2 GiB fail to load, which rules out many community LTX-2 LoRAs; our camera LoRA test lists which load.

LTX-2 19B LoRA text-to-video with the official Dolly-In camera LoRA at scale 1, 720p, 5 seconds.

LTX 2.5

Lightricks also published LTX-2.5 as open weights, a 22-billion-parameter audio-video model. Ours runs up to 20 seconds at up to 4K with sound and went much further than LTX-2 19B on our ladder (5 of 6 at L4). The catch is unevenness: in our review two everyday scenes dissolved into colour artefacts, and community testers report faces drifting on image-to-video.

Also open, but outside this comparison

Wan 2.2 Animate is open weights too but does a different job: it moves the character in your still with the motion of a reference clip (17 of 19 ladder runs delivered). The Spicy builds of Wan 2.2 and LTX 2.3 are adult-tuned builds on open bases rather than a publisher's release; our Wan and LTX 2.5 tests cover them.

Speed and price for a 5-second draft

Model time is from our 5 and 6 October runs unless noted.

Loading LoRAs adds no separate charge, and failed requests are not charged. Each family page shows the live price for every tier.

Which one to pick

Prompting habits behind the prompts above

• Say "adult" with an age band for every person, and name the wardrobe in full.
• One action per clip. A jacket sliding off and a smile fit in five seconds; a sequence does not.
• Put the camera move last, as in "slow push-in".
• On the sound models, add an "Audio:" line listing what should be heard, in order.
• On image-to-video, describe what happens next in the photo, not a new place.
• Put LoRA trigger words first, start style LoRAs around 0.5 to 0.8, and compare against a no-LoRA run.

What this comparison does not cover

• We did not score sound quality.
• The same-prompt run is one seed per model at 480p: a look, not a ranking.
• LTX 2.5 has no gore runs yet, and plain H3 has untested cells.
• Licensing is between you and each publisher; read the model cards before you ship.

FAQ

What is the best open-source NSFW video model?

MiniMax H3, through the H3 LoRA build, on our data: 68.7 in our blind test, 98.3 on Freedom and no refusals. Among the open models added on 5 October, HunyuanVideo 1.5 and Wan 2.2 LoRA tie on Freedom at 88.1.

Which is better, LTX or Wan?

For uncensored work, Wan 2.2 LoRA beats LTX-2 19B: 88.1 against 76.1 on Freedom and 58.5 against 47.5 in our blind test. LTX wins on sound, length and speed; LTX 2.5 combines sound with an 82.3 Freedom score.

Is Wan 2.2 still the best open model for NSFW video?

Not on our numbers: MiniMax H3 scored higher on quality and range. Wan 2.2 LoRA remains the pick if you rely on Wan LoRAs. Avoid the plain Wan 2.2 family for adult work; its text-to-video refused every ladder run.

Which LTX model is best?

LTX 2.5 for range and resolution, LTX-2 19B for cost and speed, and the LTX-2 19B LoRA build when you need a camera LoRA.

Is HunyuanVideo better than LTX and Wan for realistic people?

HunyuanVideo 1.5 (57.7) and Wan 2.2 LoRA (58.5) were close in our blind test, both well ahead of LTX-2 19B (47.5). HunyuanVideo is strongest on one person's face and less reliable with two people in close contact.

Do I need a GPU to run open-source video models?

Not on SpicyAPI. Every model here runs at a per-second price on hosted hardware, and the LoRA families fetch your files from a link you host.

Are open-weights video models always uncensored?

No. Open weights remove a hosted app's filter, but the model's training still decides how far it goes: LTX-2 19B softened more than half of our runs.