Lightricks put a 22B video model on Hugging Face in August with a permissive licence, day-one ComfyUI support and no content filter. In a month where Alibaba and ByteDance shipped stronger but closed models, this is the release that changes what adult creators can run themselves.
Short answer: LTX-2.5 is the most important video release of August 2026 for adult creators, and not because it is the best-looking model; Wan 3.0 and Seedance 2.5 both beat it on cinematic quality. It matters because it is open weights: Lightricks published the full 22B model on Hugging Face on 11 August under a permissive licence, it runs quantised on a 24–32 GB consumer card, ComfyUI supported it on day one, and community NSFW LoRAs trained for LTX-2.3 load on it without changes. Since Alibaba stopped publishing Wan weights after 2.2, LTX is now the strongest video model you can run without anyone’s content policy.
What shipped
LTX-2.5 | |
|---|---|
Released | 11 August 2026 |
Size | 22B diffusion transformer, Gemma 4 12B text encoder |
Weights | Hugging Face, Lightricks organisation |
Licence | Permissive; free for organisations under $10M ARR, no branding requirement |
Clip length | Auto-duration, 10 s image-to-video demonstrated |
Resolution | Up to 4K HDR |
Speed | 10-second 720p clip in 6.8 s on two GB200s; minutes on a consumer card |
New in 2.5 | Diffusion video decoder (sharper faces, text, fast motion), fidelity rendering that spends compute on hard frames, native multishot with consistent character and voice across cuts |
Tooling | Day-one ComfyUI templates, LTX API for managed hosting |
The headline number Lightricks pushed was speed. The number that matters here is the licence.
Why open weights matter again
For a year the adult video conversation has been “which hosted model can be run with adult content enabled”. That is a workable route, and it is how MyBabes runs Wan 2.7 and Seedance, but it depends on someone else’s willingness to host the model that way. Open weights remove the dependency:
No filter to negotiate. The model on your disk has no prompt classifier, no upload classifier, no output scan.
LoRAs. Act, position and camera adapters can be trained, shared and stacked. LTX has had an NSFW LoRA ecosystem since 2.3, and Lightricks confirmed that LoRAs and IC-LoRAs trained for 2.3 run on 2.5 unchanged. The catalogue is live on day one, which never happened with a Wan release.
Policy cannot change under you. Wan 2.7 users learned in April what “API-only” means; Sora users learned it harder.
Our Wan 2.7 analysis explains the hosted route; this article is about the self-hosted one, which LTX-2.5 just made far more attractive.
What it takes to run it
Setup | VRAM | Notes |
|---|---|---|
Full precision | Well over 48 GB | Datacentre only |
FP8 (native) | ~30–35 GB | 48 GB cards comfortable; 32 GB workable with the distilled checkpoint |
INT8 | ~24 GB | RTX 4090/5090 class, the community’s practical floor |
NVFP4 | ~11–16 GB | Runs, with visible quality loss; fine for previews |
Practical notes from the launch-week guides: CUDA 13.2 or newer is required, the distilled checkpoint is the recommended starting point on anything below A100 class, and generation on a 24 GB card is minutes per clip rather than seconds. That is still a different world from Wan 2.2‘s 480p self-hosted ceiling.
NSFW on LTX: what to expect
LTX has historically trailed Wan and Hunyuan on anatomy in adult scenes, and the base 2.5 model was not trained for explicit content. Three things make the picture better than the reputation:
The 2.3 NSFW LoRAs carry over. The multi-concept adapters that appeared for 2.3 in July load on 2.5. The community does not start from zero.
The new decoder helps exactly where adult clips fail. Faces and fast motion are the two things that degraded on earlier LTX versions; the diffusion decoder targets both.
Multishot with consistent character is the feature adult series creators have been asking every model for.
Where it will still lag, at least until adapters mature: close-contact physics (Wan’s speciality) and very long single shots (Seedance 2.5 and Wan 3.0 territory).
LTX-2.5 versus the alternatives
LTX-2.5 | Wan 2.2 (self-hosted) | Wan 2.7 (hosted) | Wan 3.0 (hosted) | |
|---|---|---|---|---|
Weights | Open | Open | Closed | Closed |
Adult content | Self-hosted, no filter | Self-hosted, no filter | Only via uncensored host | Vendor filtered |
Resolution | Up to 4K | 480–720p | 720/1080p | 480–1080p |
Speed | Fast | Slow | Fast | Fast |
LoRA ecosystem | Growing, 2.3 compatible | Largest for adult | None | None |
Contact physics | Moderate | Good | Best | Good |
For a self-hoster the choice is now LTX-2.5 for quality and speed, Wan 2.2 when a specific act LoRA exists only there. For everyone else, the hosted uncensored route with Wan 2.7 and Seedance remains simpler; we expect to add LTX-2.5 to the MyBabes video stack once our own adapters are trained, and this page will say when.
Related in the Lab
Key takeaways
LTX-2.5 is open weights with a permissive licence, day-one ComfyUI support and no content filter.
It runs on a 24–32 GB card quantised; NSFW LoRAs from LTX-2.3 work unchanged.
It is the strongest self-hostable video model now that Wan is closed past 2.2.
Hosted Wan 2.7 and Seedance still win on cinematic quality and contact physics.
FAQ
Is LTX-2.5 uncensored?
The weights have no content filter. Whatever you run locally generates what you prompt; the LTX API applies Lightricks’ terms.
Can LTX-2.5 run on an RTX 4090?
Yes, in INT8 or FP8 with the distilled checkpoint, at minutes per clip. NVFP4 fits smaller cards with quality loss.
Do LTX-2.3 NSFW LoRAs work on 2.5?
Lightricks states that LoRAs and IC-LoRAs trained on 2.3 run on 2.5 without changes.
Is LTX-2.5 better than Wan for NSFW?
For self-hosting, it beats Wan 2.2 on resolution and speed. Hosted Wan 2.7 still leads on close-contact physics.
Published: August 2026. We will update as adapters for 2.5 appear.
Skip the filter
Run these models with adult content enabled
MyBabes runs Wan 2.7, Seedance 2.0 and Krea inside its own generation stack, so the prompts refused on official apps generate as written. No GPU, no setup.
Open MyBabes
Editorial team
The MyBabes Lab team tests AI image, video and chat models for adult use cases: what they allow, what they block, and how to get the best output. Every note here is based on hands-on runs, not vendor marketing.
Follow on X
More from the Lab
Latest articles
All articles
Analysis
AI Porn and the Law in 2026: What's Actually Illegal, What's Regulated, and Where the Lines Sit
2025 and 2026 turned AI adult content from a legal grey zone into a regulated space with real enforcement: a US federal crime for non-consensual imagery, deepfake laws in 48 states, UK fines for missing age checks, EU labeling duties. Here is the map — and the one principle that organises all of it.
·
5 min read
How-To
Why AI Porn Looks Off: Every Artifact Explained, What Causes It, and How to Fix Each One
Everyone can tell when an AI image is off, few can say why. Each artifact has a specific technical cause and a specific fix — hands, limbs, faces, skin, teeth, text, seams, smear, flicker. This is the field guide, for images and video.
·
5 min read
Analysis
ChatGPT and NSFW in 2026: The Adult Mode That Never Shipped, What's Actually Allowed, and Where People Went Instead
"Treating adults like adults" was the promise in October 2025. Ten months later the adult mode is paused with no timeline, ChatGPT still refuses explicit content, and the search interest hasn't gone anywhere. Here is the full story, the current policy line in practice, and what the alternatives actually are.
·
4 min read