Wan 2.2 vs LTX-2.3 vs HunyuanVideo 1.5

Specs side by side

Everything below comes from each developer's model card, repo, config files or licence. Wan 2.2 appears twice because its two sizes behave like different models.

Architecture and size

Wan 2.2 A14B is a mixture of experts. A high-noise expert handles the early denoising steps, where the layout is decided, and a low-noise expert handles the later detail steps. Each is about 14B parameters, so 27B in total with 14B active per step. That's what the "high" and "low" model files in ComfyUI workflows are. TI2V-5B is a dense 5B model on a new VAE that compresses 4 × 16 × 16, reaching 4 × 32 × 32 with an extra patchification layer.

HunyuanVideo 1.5 is an 8.3B diffusion transformer. Tencent pairs it with a selective and sliding tile attention scheme (sparse attention, 720p models only, H-series GPUs only) and separate step-distilled super-resolution models for 720p and 1080p.

LTX-2.3 is a 22B audio-video diffusion transformer. One model produces both streams, which is why its card lists tasks such as audio-to-video and video-to-audio. It uses a Gemma 3 text encoder, downloaded separately.

Resolution, length and frame rate

Wan 2.2 A14B accepts exactly four sizes: 1280 × 720, 832 × 480 and their portrait versions. For image to video, --size sets the pixel area and the aspect ratio follows the input image. The default is 81 frames at 16 fps, about five seconds, and frame counts must be 4n+1. TI2V-5B is 720p only, which for this model means 1280 × 704, at 121 frames and 24 fps.

HunyuanVideo 1.5 generates at 480p or 720p in 16:9, 4:3, 1:1, 3:4 or 9:16, and Tencent calls 121 frames the optimal length. Super resolution is on by default and lifts 480p to 720p or 720p to 1080p.

LTX-2.3's open repo doesn't set a ceiling; width and height must be divisible by 32 and the frame count must be a multiple of 8 plus 1. Spatial upscalers (x2 and x1.5) raise resolution, and a temporal x2 upscaler raises frame rate. For a sense of the intended range, Lightricks' hosted API serves LTX-2.3 from 720p to 4K at 24 to 50 fps, in clips of 6 to 20 seconds.

Audio

Only LTX-2.3 generates sound. Wan 2.2 and HunyuanVideo 1.5 output silent video, so with either of them you'd add sound in a separate step. Wan has an audio-driven sibling, Wan2.2-S2V-14B, that animates a reference image to an audio track, but it takes audio in rather than generating it.

The LTX-2.3 card notes that audio generated without speech "may be of lower quality". If native audio is the deciding factor, also look at LTX-2.5 and MiniMax H3 below.

Distilled variants and step counts

Wan has no official distilled checkpoint. Its README points to LightX2V, a community project with step-distilled and quantised Wan models, and to FastVideo for distilled models with sparse attention. LTX's distilled pipeline runs 8 steps in stage one and 4 in its upscaling stage two, and Lightricks also publishes a distilled LoRA for the full dev model.

# Wan 2.2 TI2V-5B, image to video, 24 GB GPUpython generate.py --task ti2v-5B --size 1280*704 --ckpt_dir ./Wan2.2-TI2V-5B \  --offload_model True --convert_model_dtype --t5_cpu \  --image start.jpg --prompt "The camera pushes in slowly..."# HunyuanVideo 1.5, 480p image to video with the step-distilled modeltorchrun --nproc_per_node=1 generate.py --prompt "..." --image_path start.png \  --resolution 480p --aspect_ratio 16:9 --enable_step_distill true \  --output_path ./outputs/output.mp4 --model_path ./ckpts

VRAM

HunyuanVideo 1.5 has the lowest official floor: 14 GB, measured with offloading on, on Linux with an NVIDIA GPU. Wan 2.2 TI2V-5B needs 24 GB with offloading, dtype conversion and the T5 encoder on CPU. Wan 2.2 A14B needs 80 GB for the single-GPU 720p command, or multiple GPUs with FSDP and Ulysses.

Lightricks doesn't publish a VRAM figure for LTX-2.3. It suggests FP8 casting and CPU or disk offload for constrained GPUs, and publishes an FP8 build of 2.3. Community tools report lower numbers for all three, but those aren't the developers' figures. Our hardware requirements page lists what each developer publishes.

Licence

• Wan 2.2: Apache 2.0. Commercial use anywhere. The Wan team claims no rights over outputs.
• HunyuanVideo 1.5: Tencent Hunyuan Community License. Not licensed in the EU, UK or South Korea. A licensee with more than 100 million monthly active users in the month before the November 21, 2025 release needs a separate licence. Services must name the actual provider and state that Tencent doesn't endorse them, and outputs can't be used to improve other AI models.
• LTX-2.3: LTX-2 Community License. Free for entities under $10M in annual revenue. At $10M or more, you need a paid commercial licence from Lightricks.

For a business in Europe, that rules out HunyuanVideo 1.5. For a company past $10M revenue, LTX costs money. Wan 2.2 has neither condition. More on this in commercial use of open models.

LoRA training

Tencent and Lightricks ship their own training code. HunyuanVideo 1.5's train.py has a --use_lora option (rank 8, alpha 16 by default) and Tencent asks that you fine-tune with its Muon optimizer. Lightricks' ltx-trainer covers LoRA, full fine-tuning and IC-LoRA on the LTX dev model, and Lightricks publishes its own LTX-2.3 LoRAs and IC-LoRAs on Hugging Face.

Wan's official repo has no training code; it points to DiffSynth-Studio. In practice Wan 2.2 is covered by the community trainers: musubi-tuner supports Wan 2.2 and HunyuanVideo 1.5, and ai-toolkit supports Wan 2.2 and LTX-2.3. One caveat from the Wan team: LoRAs trained on Wan 2.2 aren't recommended for Wan-Animate.

The head-to-heads

HunyuanVideo 1.5 vs Wan 2.2

HunyuanVideo 1.5 runs on less memory (14 GB against 24 GB for TI2V-5B and 80 GB for A14B), renders 24 fps where A14B defaults to 16, and has an official 1080p super-resolution path and an 8 to 12 step image-to-video model. Wan 2.2 wins on licence, with no territory or user limits, and on tool support. If you're outside the EU, UK and South Korea and memory is the constraint, HunyuanVideo 1.5. Otherwise Wan 2.2.

Wan 2.2 vs LTX-2.3

LTX-2.3 generates audio, goes above 720p and samples in 8 steps with its distilled model. Wan 2.2 is Apache 2.0 with no revenue threshold and has a published VRAM figure. Pick LTX if you need sound or resolution and are under $10M revenue; pick Wan for silent clips at 480p or 720p with no licence conditions.

HunyuanVideo 1.5 vs LTX-2.3

Both have official training code and fast distilled variants. LTX-2.3 adds audio and higher resolution; HunyuanVideo 1.5 has the published 14 GB floor. Their licences fail in different places: Hunyuan by territory, LTX by revenue.

LTX-2.3 vs LTX-2.5

LTX-2.5 arrived on August 11, 2026 and the LTX-2 repo now calls 2.3 legacy. It keeps the 22B size and adds native multi-shot generation (several connected shots in one pass, where earlier versions made one continuous shot), a diffusion video decoder that replaces the VAE reconstruction stage, a Gemma 4 12B text encoder, a prompt enhancer, an optional duration predictor and what Lightricks calls a substantially improved distilled model. The 2.5 card says most LoRAs and IC-LoRAs trained on 2.3 run on 2.5 unchanged. The licence moves to the LTX-2.x Community License, with the same $10M threshold but stricter terms above it: a larger company can only use it free for personal or non-production evaluation. Unless you depend on a 2.3-only workflow, start on 2.5.

Wan 2.2 vs MiniMax H3

MiniMax H3 is a 33B dense model that generates video with 32 kHz stereo audio, 4 to 15 seconds at 24 fps, with first and last frame or multi-reference input. Its open weights stop at a 768 px short side; the 2K stage is API-only. The bigger difference is the licence: the H3 Community License excludes the US, EU, UK and South Korea, requires authorisation above $20M in yearly revenue and makes commercial products show "MiniMax H3" in the UI. If you're in one of those regions, it isn't an option for self-hosting, and Wan 2.2 is.

LongCat-Video vs Wan 2.2

LongCat-Video is a 13.6B dense model under MIT, doing text to video, image to video and continuation in one checkpoint, with a 720p 30 fps output from a refinement pass and long clips built by chaining continuations. In Meituan's own internal mean-opinion-score test it scored 3.38 overall against Wan 2.2 T2V-A14B's 3.35 for text to video, and 3.17 against Wan 2.2 I2V-A14B's 3.26 for image to video. Those are the developer's numbers on its own prompts. Meituan doesn't publish a VRAM figure or training code, so Wan 2.2 is the easier model to run and fine-tune; LongCat is the one for long or extended clips.

Wan 2.2 vs Wan 2.1

Wan 2.2 added the mixture-of-experts design to the 14B models, trained on 65.6% more images and 83.2% more videos than 2.1, labelled the data for lighting, composition, contrast and colour tone, and introduced the high-compression VAE used by TI2V-5B. Wan 2.1 is still the source of models 2.2 doesn't replace: the T2V-1.3B model (8.19 GB VRAM per the Wan team), the first-last-frame FLF2V-14B and the VACE editing models. Both are Apache 2.0.

What about Wan 2.7 and Wan 3.0?

Neither has open weights. Alibaba launched Wan 2.7 in 2026 as API-only models on its own cloud: Wan2.7-Image in April, and Wan 2.7 video with 2 to 15 second clips and video editing. Wan 3.0 followed in August 2026, also API-only, with clips up to 30 seconds at up to 1080p and native audio. There are no Wan 2.7 or Wan 3.0 weights in the Wan-AI Hugging Face organisation.

So for self-hosting, Wan 2.2 is still the newest open Wan release. Its open family also grew past the T2V, I2V and TI2V models compared here: Wan2.2-Animate-14B for character animation and replacement, and Wan2.2-S2V-14B for audio-driven video, both Apache 2.0. If Alibaba publishes Wan 2.7 weights, we'll add it to this comparison.

Which one to pick

• You need an unrestricted commercial licence: Wan 2.2 (or LongCat-Video, MIT).
• You're in the EU, UK or South Korea: Wan 2.2, LongCat-Video or LTX. HunyuanVideo 1.5 and MiniMax H3 aren't licensed there, and H3 also excludes the US.
• You need audio in the same pass: LTX-2.5, or LTX-2.3 if your workflow is built on it. MiniMax H3 if its territory terms work for you.
• You have one GPU with 16 to 24 GB: HunyuanVideo 1.5 (14 GB floor) or Wan 2.2 TI2V-5B (24 GB).
• You need more than 720p: LTX, or HunyuanVideo 1.5 with its 1080p super-resolution model.
• You'll train LoRAs: Wan 2.2 for community trainer support, or LTX and HunyuanVideo 1.5 for official training code.
• You need long clips: LongCat-Video, built on continuation.

Windpaint doesn't serve video yet. Wan 2.2 image-to-video, HunyuanVideo 1.5, Wan 2.2 text-to-video, TI2V-5B and LongCat-Video are in its catalogue as coming soon.