Is LTX-2.3 Better Than Wan 2.2? Audio, Licensing & Hardware Compared
Is LTX-2.3 better than Wan 2.2? There is no single winner — the right choice depends on your hardware, licensing needs, and whether your video needs synchronized sound. This guide compares both models on native audio, licensing, GPU requirements, speed, and ecosystem, then tells you which one to run.
TL;DR
Neither model is universally "better" — the right pick depends on your hardware, license situation, and whether your project needs sound.
- LTX-2.3 wins on native audio. It is a 22B-parameter Diffusion Transformer that generates synchronized audio and video in one pass — something Wan 2.2 does not do in a single model.
- Wan 2.2 wins on licensing and ecosystem. It is Apache 2.0, uses Mixture-of-Experts (27B total, 14B active), and its TI2V-5B runs 720P at 24 FPS on 24GB GPUs.
- Pick LTX-2.3 for dialogue, lip-sync, and soundtracks from one model with fast distilled inference.
- Pick Wan 2.2 for unrestricted Apache-2.0 commercial use, the TI2V-5B low-VRAM path, or the largest LoRA/workflow library.
The two models, briefly
Wan 2.2 (Alibaba, weights shipped July 28, 2025) became the reference open-source model for 720P generation; LTX-2.3 (Lightricks, October 2025) handles video and audio in a single model. Both run in ComfyUI and in launchers on Pinokio such as Wan2GP.
LTX-2.3 is a DiT-based audio-video foundation model — a significant update to LTX-2 with improved audio and visual quality, per the Hugging Face model card. Verified specs: 22B checkpoints (dev, distilled, distilled v1.1; distilled runs in 8 steps at CFG 1), synchronized audio + video in one model, a Gemma 3 text encoder, and resolution rules (dimensions ÷ 32, frames 8k + 1). The LTX-2.x Community License is free for individuals and smaller companies, but requires a paid license for commercial use by entities with at least $10M in annual revenue. The ltx23.app product page adds positioning claims — native 9:16 portrait at 1080×1920, a rebuilt VAE, a 4x-expanded text connector — and states the LTX 2 series achieves roughly 18x the throughput of Wan 2.2 14B on H100-class hardware (vendor claims, not independent benchmarks).
Wan 2.2 is Alibaba's open-source family: a Mixture-of-Experts A14B series (27B total, 14B active per step) plus the dense TI2V-5B, which runs text-to-video and image-to-video at 720P, 24 FPS, generating a 5-second clip in under 9 minutes on a consumer GPU such as an RTX 4090. Versus Wan 2.1 it trained on 65.6% more images and 83.2% more videos, and it ships specialist variants (S2V-14B speech-to-video, Animate-14B character animation). Its license is Apache 2.0.
LTX-2.3 vs Wan 2.2: the comparison that matters
| Dimension | LTX-2.3 | Wan 2.2 |
|---|---|---|
| Developer | Lightricks | Alibaba (Wan-Video) |
| Architecture | 22B Diffusion Transformer, audio + video in one model | MoE A14B (27B total / 14B active) + dense TI2V-5B |
| Native audio | Yes — synchronized dialogue, ambience, music | No single-model equivalent; S2V-14B is a separate variant |
| Consumer GPU | Larger 22B model; FP8 and offload paths documented | TI2V-5B runs 720P @ 24 FPS on 24GB GPUs (e.g., RTX 4090) |
| License | LTX-2.x Community License — free below $10M annual revenue | Apache 2.0 |
| Ecosystem | Growing; official ComfyUI + Diffusers support | Large community LoRA/workflow library since July 2025 |
Where each model wins
LTX-2.3: synchronized audio and video in one model
Audio and video in one model is the biggest differentiator: talking heads, lip-sync, and music cues come out of the same pass, whereas Wan 2.2 needs a separate audio model or S2V-14B. LTX-2.3 also offers fast distilled inference (8 steps at CFG 1; Lightricks claims up to 50% lower compute) and one unified workflow.
Wan 2.2: Apache-2.0 licensing, proven consumer-GPU path, deeper ecosystem
Apache 2.0 with no revenue threshold is far easier for larger companies than the LTX-2.x $10M line. Its consumer-GPU path is the most proven (TI2V-5B at 720P on 24GB-class cards), and its ecosystem is deeper — a large community library of LoRAs and workflows.
Wan 2.1 vs Hunyuan Video vs LTX-2.3: the local-stack view
Running locally through Wan2GP on Pinokio, you can run them all side by side — Wan2GP (built for "GPU poor" machines with as little as 6GB of VRAM) supports Wan 2.1/2.2, Hunyuan Video, and LTX Video / LTX-2. LTX-2.3 is the only one generating audio and video together (best for dialogue); Wan 2.1/2.2 is the default for long, quality-first generations on modest hardware; Hunyuan Video (Tencent, an open-source 13B+ model) is a strong text-to-video baseline without native audio.
How to load a custom .safetensors model for LTX-2.3 in Pinokio Wan2GP
A recurring community question is how to load a custom LTX-2.3 checkpoint — say, a finetune named 10eros_v1.3_bf16.safetensors — into the Wan2GP app on Pinokio. The general workflow, based on the official Wan2GP repo:
- Install Wan2GP from the Pinokio store — a one-click launcher for Wan2GP, a Gradio UI that auto-downloads models.
- Open the LTX Video / LTX-2 tab and let Wan2GP download a base LTX-2.3 checkpoint first.
- Add your finetune via Wan2GP's Models/Checkpoints Manager plugin (you can add checkpoints from Hugging Face or CivitAI; see
docs/FINETUNES.md). - Put LTX-2.x LoRAs in the shared
loras/ltx2folder, where Wan2GP reads them. - Confirm compatibility — a finetune only works with the model it was trained on, and the
_bf16suffix alone tells you nothing.
We could not verify that specific community file — treat it like any community finetune and confirm its base model.
Which one should you run?
| Your situation | Pick |
|---|---|
| Dialogue, lip-sync, or a soundtrack from one model | LTX-2.3 |
| Large company (at least $10M revenue) needing simple licensing | Wan 2.2 (Apache 2.0) |
| One 24GB-class consumer GPU, 720P @ 24 FPS | Wan 2.2 (TI2V-5B) |
| Fastest distilled loop, license threshold acceptable | LTX-2.3 |
| Biggest community LoRA/workflow library | Wan 2.2 |
| No GPU — just want finished video | LTX-2.3 via the ltx23.app cloud generator |
Rule of thumb: if your project involves people talking, pick LTX-2.3; for scale, licensing, or a modest single GPU, pick Wan 2.2.
FAQ
Is LTX-2.3 better than Wan 2.2?
Not universally — they win on different axes. LTX-2.3 is better for synchronized audio-video generation and fast distilled inference; Wan 2.2 is better for Apache-2.0 licensing and 720P on a 24GB consumer GPU. Better for what?
Is LTX-2.3 free?
The weights are free, but the LTX-2.x Community License requires a paid license for commercial use by entities with at least $10M in annual revenue. Individuals and smaller companies can use it freely; Wan 2.2 is Apache 2.0 with no threshold.
Does LTX-2.3 generate audio?
Yes — it generates synchronized dialogue, ambience, and music together with the visuals in one pass. That is its headline advantage over Wan 2.2.
Can I run Wan 2.2 and LTX-2.3 on the same PC?
Yes. Wan2GP on Pinokio supports Wan 2.1/2.2, Hunyuan Video, and LTX Video / LTX-2 in one app; both also work in ComfyUI.
Do I need a powerful GPU?
Only if you run them locally. On ltx23.app rendering happens in the cloud — no local GPU required — and new accounts get free credits to try LTX-2.3.
Try LTX-2.3 without a GPU
If synchronized audio, one unified workflow, and fast generation match what you need, you don't have to build a local setup to test it. The LTX-2.3 cloud generator at ltx23.app runs the model for you: type a prompt, optionally upload an image or audio track, and download the finished video. New accounts start with free credits, and outputs include commercial use rights. Start your first LTX-2.3 video at ltx23.app.
Sources
- Lightricks/LTX-2.3 — Hugging Face model card — official LTX-2.3 checkpoints, 22B spec, distilled 8-step config, Gemma 3 encoder, and limitations.
- Lightricks/LTX-2 — official GitHub repository — LTX-2.3 pipelines, LTX-2.3 legacy model notes, and the LTX-2.x Community License text.
- Wan-Video/Wan2.2 — official GitHub repository — Wan 2.2 MoE architecture, TI2V-5B specs, release timeline, Apache 2.0 license, and training-data figures (65.6% more images, 83.2% more videos vs Wan 2.1).
- Wan2GP (DeepBeepMeep) — GitHub repository — Wan2GP local launcher, supported models (Wan 2.1/2.2, LTX-2, Hunyuan Video), finetune and LoRA handling.
- Tencent/HunyuanVideo — official GitHub repository — the open-source 13B base model cited in the local-stack comparison.
- LTX-2.3 — ltx23.app product page — product positioning, vendor performance claims, native portrait claim, and cloud access.
Notes on sourcing: performance figures such as "18x throughput vs Wan 2.2 14B", "up to 50% lower compute", and native 9:16 portrait come from the ltx23.app product page and are vendor claims, not independent benchmarks; architecture, checkpoint, and license details are sourced from the official Lightricks, Wan-Video, and Tencent repositories. License terms change — verify on the official pages before you rely on them. This guide is published by the LTX-2.3 Editorial Team, and its call-to-action promotes the LTX-2.3 cloud generator at ltx23.app; LTX-2.3-favorable claims are model capabilities or labeled vendor claims, not independent benchmarks.
