Local AI Video Generation: Wan 2.2, LTX-Video & HunyuanVideo (2026)

Published April 23, 2026 - 22 min read

Short answer (July 2026): The current open local-video story is LTX-2 from Lightricks - the first open-weights model that generates synchronized video and audio in a single pass. The March 2026 update, LTX-2.3 (a ~22B-parameter DiT), ships with a free open-source desktop editor (LTX Desktop) so the whole pipeline runs on your own machine. For pure silent-video quality on a 24GB card, Wan 2.2 is still the leader; LTX-Video 0.9.5 (LTX-2's lighter predecessor) remains the fastest thing that fits on a 12GB GPU. The catch nobody talks about: which open model, at which precision, with which sampler, on which VRAM tier - get one wrong and you sit through a four-minute generation that returns mush.

A year ago, "local video generation" meant 16-frame, 256x256 jittery loops that took 20 minutes per clip. Now you can render a five-second 720p shot from text on a single RTX 4090 in 90 seconds with motion that does not embarrass you - and with LTX-2, the clip can come out of the model already carrying its own synced soundtrack.

I have spent the last three months running every credible open video model on three different rigs: a 4090 24GB, a 3090 24GB, and an RTX 5070 Ti 16GB. This guide is the distilled output - what actually works, what fails on a 16GB card, what to install, and the ComfyUI workflows that produce usable output instead of the demos cherry-picked on Reddit.

---

The Local Video Stack Right Now (July 2026) {#stack-july-2026}

Section added July 20, 2026. The full guide below still stands, but the practical stack has shifted since the June benchmarks were run - mostly in the direction of lower VRAM floors. Three models now define local video:

• Wan 2.2 (14B, FP8/GGUF) - still the silent-quality pick, and the "24GB or skip it" advice below is no longer the whole story. Community-reported numbers: the Q5_K_M GGUF runs 720p on 16GB cards, GGUF plus aggressive offloading reaches 6-8GB, and the smaller Wan 2.2 5B TI2V checkpoint fits on 8GB outright. Expect slower generations on those tiers, but Wan on a mid-range card is now a real option.
• HunyuanVideo 1.5 - Tencent slimmed the model to 8.3B parameters, and it runs in roughly ~14GB with offloading. That is a very different proposition from the original 13B model benchmarked below, which wanted 24GB at Q8.
• LTX-2.3 (Lightricks, ~March 2026) - the 22B DiT that generates synchronized video and audio in one model, shipped in dev and 8-step distilled variants. The Hugging Face model card now lists 52 quantizations, so there is almost certainly a build for your VRAM tier. The weights remain under the LTX-2 Community License (check the license text on the model card for the exact terms), and Lightricks also ships a unified LTX Trainer covering multiple training objectives.

A warning while you shop for models: you will run into pages offering "Wan 2.7 open weights" downloads. Those claims are SEO fabrications - official Wan open weights stop at 2.2, and anything advertising a "Wan 2.7" checkpoint is not distributing what it claims to be. We trace where the claim comes from in our Wan 2.7 fact-check.

Tooling note: ComfyUI v0.28.0 (July 15, 2026) added native SeedVR2 image and video upscaling, so the "generate at 720p, upscale for delivery" workflow no longer requires a third-party node pack.

Quick Start: 5 Minutes to Your First Local Video

If you already have ComfyUI installed:

1. Update ComfyUI: git pull && pip install -r requirements.txt
2. Download LTX-Video 0.9.5 (best speed/quality on 16GB+ GPUs): huggingface-cli download Lightricks/LTX-Video --local-dir models/checkpoints/ltx
3. Drop the official LTX-Video workflow into ComfyUI
4. Set 768x512, 65 frames (~5 sec at 13fps), 30 steps, then click Queue

On a 4090 you have a generated clip in ~90 seconds. On a 3060 12GB, expect ~7 minutes. Welcome to local video.

---

Table of Contents

1. What Actually Runs Locally Right Now
2. LTX-2 & LTX-2.3: First Open Synced Audio+Video
3. Hardware Requirements
4. Installing ComfyUI for Video
5. Wan 2.2: Best Open Quality
6. LTX-Video 0.9.5: Best Speed (LTX-2's Predecessor)
7. HunyuanVideo: Cinematic Look
8. Mochi 1 & CogVideoX: Honorable Mentions
9. Head-to-Head Benchmark
10. Prompt Patterns That Actually Work
11. Pitfalls
12. FAQ

---

What Actually Runs Locally Right Now {#what-runs-locally}

Forget Sora, Veo 3, Runway Gen-3. Those are closed APIs. The open ecosystem as of July 2026 has five serious contenders, and one of them - LTX-2.3 - is the first that outputs synced audio with the video:

LTX-2.3 is the big shift of 2026: it is the first production-ready open-weights model that generates synchronized video and audio - dialogue, ambience, lip-sync - in a single diffusion pass. It is a ~22B model, so it is heavier than the silent options here, and we have left its max-resolution and clip-length cells as "see notes": those figures move between point releases and we have not independently benchmarked them, so check the model card for current numbers. On VRAM there is no single floor - the Hugging Face model card ships 52 quantizations, from full precision down to aggressive GGUF builds, so where it runs depends on which quant you pick and the resolution you target (more in the LTX-2 section below).

Among the silent models, LTX-Video 0.9.5 is the only one that fits comfortably on a 16GB card without gymnastics. Wan 2.2 and HunyuanVideo are the silent quality leaders, but they want a 4090, 3090, or A6000. Mochi 1 is excellent but slow - production use only if you have time to spare.

The Hugging Face video model leaderboard discussion tracks newer entries, and the field moves fast. By late 2026 expect at least two more open models in this tier.

---

LTX-2 & LTX-2.3: The First Open Synced Audio+Video {#ltx-2}

If you only track one thing in open video this year, make it LTX-2. Lightricks open-sourced it as, in their words, the first production-ready audio-video generation model with truly open weights. The headline claim - and it holds up - is that it generates synchronized video and audio in one pass: not a silent clip you dub later, but frames and an audio waveform produced together, frame-aligned, including lip-sync.

The March 2026 point release, LTX-2.3, is the one most people actually run now. It is a ~22B-parameter Diffusion Transformer, shipped in dev and 8-step distilled variants. It ships with LTX Desktop, a free, open-source local video editor that wraps the model in a project-based UI - visual prompt builder, batch queue, trimming and export - so the entire generate-and-edit loop runs on one machine without touching a cloud API.

The license

The LTX-2 model weights are released under the custom "LTX-2 Community License Agreement" - not a blanket permissive license like MIT or Apache-2.0. The terms are more involved than a plain "do anything" license, so read the exact conditions on the model card before you ship a commercial product on the weights. The takeaway: treat the LTX-2 weights as their own license, not as generically "open source."

Setup (ComfyUI)

# LTX-2.3 has official ComfyUI support. Pull the weights from the model card:
huggingface-cli download Lightricks/LTX-2.3 \
  --include "*.safetensors" \
  --local-dir models/checkpoints/ltx2

# Low-VRAM users: grab a GGUF or FP8/NVFP4 variant instead of the BF16 base,
# and use the ComfyUI-GGUF custom node to load it.

Or skip ComfyUI entirely and install LTX Desktop (Windows or Apple Silicon) from the Lightricks GitHub for a one-window local workflow. Note: on macOS the desktop app currently routes generation through the LTX API rather than running fully on-device.

The honest VRAM reality

This is the catch. LTX-2.3 is a ~22B model, so it is heavier than the silent options on this page, and there is no single "you need exactly X GB" number we can verify. What the model card does provide is breadth: 52 quantizations, from full precision down to aggressive GGUF builds. That is the honest answer to "will it run on my card" - the floor depends on which quant you pick, the resolution you target, and how much you offload to system RAM. Lower-precision community builds bring it onto smaller cards at reduced quality; the full-precision path wants a workstation-class GPU. Rather than trust a spec number off a blog, pull the quant that matches your VRAM from the model card and benchmark it on your own rig.

What LTX-2.3 is great at

• Synced audio out of the box - dialogue, ambience, foley, lip-sync - with no second tool in the loop
• One model for the whole job - video and its soundtrack generated together, frame-aligned
• A real local editor (LTX Desktop) instead of a bare inference script

Where it still struggles

• Hardware: as a ~22B model it is heavier than the silent options here, so the full-precision path wants a workstation-class GPU
• Audio quality is impressive for an open model but not yet studio-clean; expect to re-touch in a DAW for client work
• Like every model here, fast discrete motion and hands remain weak spots

LTX-Video 0.9.5 (covered below) is still in this guide on purpose: it is LTX-2's lighter, silent predecessor and remains the best fast option on a 12GB card. Think of 0.9.5 as the daily driver for quick silent clips and LTX-2.3 as the finals tool when you want synced sound and have the VRAM for it.

---

Hardware Requirements {#hardware}

Minimum (LTX-Video only)

Recommended (all models, full quality)

Mac Apple Silicon

LTX-Video runs on M2/M3 Max with 64GB unified memory at roughly 1/8 the speed of a 4090, which is borderline usable for previews. Wan 2.2 and HunyuanVideo do not have stable MPS paths yet. If you are on Mac, see our Mac local AI setup guide first to confirm your driver state.

---

Installing ComfyUI for Video {#install-comfyui}

ComfyUI is the de facto frontend for local video generation. Most workflows are shared as JSON files you drop directly onto the canvas.

# Clone ComfyUI
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI

# Create a clean Python 3.11 venv
python3.11 -m venv venv
source venv/bin/activate  # (Windows: venv\Scripts\activate)

# Install PyTorch matching your CUDA (12.4 example)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

# Install ComfyUI deps
pip install -r requirements.txt

# Install ComfyUI Manager (essential for video custom nodes)
cd custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Manager
cd ..

# Launch
python main.py --listen 127.0.0.1

Open http://127.0.0.1:8188. From the Manager, install:

• ComfyUI-VideoHelperSuite (frame export, video preview)
• ComfyUI-LTXVideo (LTX nodes)
• ComfyUI-HunyuanVideoWrapper (Hunyuan nodes)
• ComfyUI-WanVideoWrapper (Wan 2.2 nodes)

Restart ComfyUI after each install.

---

Wan 2.2: Best Open Quality {#wan-22}

Alibaba's Wan 2.2 is the highest-quality open video model I have benchmarked. It produces stable subjects, coherent motion across 5 seconds, and surprisingly clean text rendering inside frames. It also wants every byte of a 24GB card.

Setup

# Inside ComfyUI/models
mkdir -p diffusion_models text_encoders vae

# Download Wan 2.2 14B (FP8 for 24GB cards)
huggingface-cli download Wan-AI/Wan2.2-T2V-A14B \
  --include "*.safetensors" \
  --local-dir models/diffusion_models/wan22

# T5 text encoder (shared with Hunyuan)
huggingface-cli download Wan-AI/Wan2.2-T2V-A14B \
  --include "text_encoder/*" \
  --local-dir models/text_encoders/wan22_t5

# VAE
huggingface-cli download Wan-AI/Wan2.2-T2V-A14B \
  --include "vae/*" \
  --local-dir models/vae/wan22

Total disk: ~38GB.

Recommended Settings (RTX 4090 24GB)

• Resolution: 1280x720
• Frames: 81 (~5.4s at 15fps)
• Sampler: Euler (Wan-tuned)
• Steps: 30
• Guidance: 5.0
• Precision: FP8 e4m3fn

Generation time on a 4090: 4 minutes 20 seconds. On a 3090: 7 minutes 10 seconds. The 5070 Ti 16GB cannot fit Wan 2.2 even at FP8 without aggressive offloading - skip it.

What Wan 2.2 is great at

• Stable human subjects across a full 5-second shot
• Camera motion (push-in, pan, dolly) that holds geometry
• Text inside the frame ("hello" on a sign reads correctly more often than not)
• Realistic lighting, especially golden-hour outdoors

Where it fails

• Hands. Still hands. Always hands.
• Anything with discrete fast motion (juggling, sports). It blurs to mush.
• Frames longer than 81 - quality cliff.

---

LTX-Video 0.9.5: Best Speed (LTX-2's Predecessor) {#ltx-video}

LTX-Video 0.9.5 is LTX-2's lighter, silent predecessor, and it is still the model I reach for 80% of the time for fast clips. On a 4090 it generates a 5-second 768x512 clip in 90 seconds. On a 3060 12GB it takes about 7 minutes - slow, but it actually fits. If you have the VRAM and want sound or length, step up to LTX-2.3 (see the LTX-2 section); for quick silent iteration on a 12GB card, 0.9.5 is still the right tool.

Setup

huggingface-cli download Lightricks/LTX-Video \
  --include "*.safetensors" \
  --local-dir models/checkpoints/ltx

# T5 encoder (FP16)
huggingface-cli download PixArt-alpha/PixArt-Sigma-XL-2-2K-MS \
  --include "text_encoder/*" \
  --local-dir models/text_encoders/ltx_t5

The official LTX workflow is on the Lightricks GitHub. Drag the JSON onto the ComfyUI canvas.

Recommended Settings

• Resolution: 768x512
• Frames: 65 (5 sec at 13fps)
• Sampler: Euler
• Steps: 30
• Guidance: 3.5
• Precision: BF16 (FP8 also works but adds shimmer)

What LTX is great at

• Speed. It is 3-4x faster than Wan 2.2 and Hunyuan.
• Lower VRAM footprint - the only model that runs on 12GB cards.
• Image-to-video extensions (animate a still photo)

Where it fails

• Motion at high speed (anything faster than walking)
• Detailed text in-frame
• Photorealistic faces - they hold for 2-3 seconds, then drift

---

HunyuanVideo: Cinematic Look {#hunyuan}

Tencent's HunyuanVideo has the most distinctive aesthetic of the four - rich color, real bokeh, surprisingly cinematic compositions. It is also the slowest of the three "premium" models.

Setup

# 13B model in Q8 quantization fits on 24GB
huggingface-cli download tencent/HunyuanVideo \
  --include "transformers/mp_rank_00_model_states.pt" \
  --local-dir models/diffusion_models/hunyuan

# CLIP + LLaVA text encoder bundle
huggingface-cli download Kijai/HunyuanVideo_comfy \
  --include "*.safetensors" \
  --local-dir models/text_encoders/hunyuan

Use Kijai's HunyuanVideoWrapper nodes (search ComfyUI Manager for "HunyuanVideo"). It is the most actively maintained wrapper.

Recommended Settings (24GB GPU)

• Resolution: 720x1280 (vertical) or 1280x720 (horizontal)
• Frames: 73 (5 sec at 14.6fps)
• Sampler: dpmpp_2m_sde
• Steps: 30
• Guidance: 6.0
• Precision: FP8 e4m3fn

Generation time on 4090: 5 minutes 50 seconds.

What Hunyuan is great at

• Cinematic feel out of the box - color, depth, film-like motion
• Detailed environments (forests, cityscapes, water)
• Vertical format for social

Where it fails

• Slower than Wan and LTX
• Memory hungry - any drift from recommended settings tips into OOM
• Less prompt control - it has a "house style" you cannot fully suppress

---

Mochi 1 & CogVideoX: Honorable Mentions {#mochi-cogvideox}

Mochi 1 (Genmo) was state of the art in late 2025. It still produces excellent motion but takes 8+ minutes per 5-second clip on a 4090, and the quality has been matched or exceeded by Wan 2.2. Use it if you specifically like the Mochi aesthetic.

CogVideoX-5B (Tsinghua) was the open standard for most of 2024. It runs on 12GB VRAM but the output quality is now clearly behind LTX-Video. Skip unless you have a specific compatibility need.

---

Head-to-Head Benchmark {#benchmark}

I generated the same 12 prompts on each model on a 4090. Identical seed, comparable settings. Output was rated by three independent reviewers on prompt adherence, motion stability, and aesthetic quality (1-10).

Composite winners by use case:

• Best overall quality: Wan 2.2
• Best for fast iteration: LTX-Video
• Best look: HunyuanVideo
• Best for 12GB cards: LTX-Video (only option)

---

Prompt Patterns That Actually Work {#prompts}

Video models reward different prompts than image models. After 600+ test generations across all four models, these patterns produced the most reliable output:

1. Camera direction first

Bad: A woman walks through a forest Good: Cinematic wide shot, slow dolly forward, a woman walks through a sunlit pine forest, golden hour, shallow depth of field, 35mm film grain

2. Specify motion explicitly

Bad: A cat Good: A black cat slowly turns its head from left to right, then blinks, photorealistic, soft window light

3. Anchor with a still-image vibe

Bad: A futuristic city at night Good: Wide aerial shot of a neon-lit Tokyo street at night, light rain on asphalt, gentle camera drift to the right, cinematic, anamorphic lens, color grade like Blade Runner 2049

4. Avoid contradictory motion

Models break when you ask for "fast running, slow camera, smooth transition." Pick one tempo and stick with it.

5. Use the model's strengths

• Wan 2.2: subjects, dialogue scenes, in-frame text
• LTX-Video: nature, water, atmosphere, animal close-ups
• HunyuanVideo: environments, cinematic establishing shots, vertical social content

For longer-form workflows (image-to-video, video-to-video, style transfer), our Flux local image generation guide walks through the still-image side of the pipeline that feeds into image-to-video nodes.

---

Common Pitfalls {#pitfalls}

1. Running BF16 on a 16GB card. All three premium models OOM at BF16 on 16GB. Use FP8 or Q8.

2. Skipping the model offload settings. Wan and Hunyuan have explicit offload nodes - configure CPU offload for the text encoder and you reclaim 4-6GB VRAM.

3. Asking for 10-second clips. Every model degrades hard past their training length. Stitch two 5-second clips with a video editor instead.

4. Wrong CFG/guidance scale. Each model has a sweet spot - LTX wants 3.0-3.5, Wan wants 5.0-5.5, Hunyuan wants 6.0-7.0. Defaults from one model do not transfer.

5. Outdated ComfyUI Manager versions. Video custom nodes change weekly. Update before each session.

6. Using a slow SSD. Models are 30-80GB and reload between runs if you swap. SATA SSDs add 30-90 seconds per generation. NVMe is not optional.

7. Trying these on a Mac without MPS support compiled in. Most "Mac compatible" claims for video models in early 2026 still depend on PyTorch builds with MPS fallback that silently revert to CPU. Validate with torch.backends.mps.is_available() before pulling models.

For the broader hardware question, our best GPUs for AI 2025 breakdown covers the price-per-performance reality of 4090 vs 3090 vs newer cards.

---

Frequently Asked Questions {#faq}

Can I generate video on a 12GB GPU?

Yes, but only LTX-Video. A 3060 12GB or 4060 Ti 12GB will run LTX-Video at 768x512 in roughly 7-9 minutes per 5-second clip. Wan 2.2, HunyuanVideo, and Mochi 1 all require 24GB or aggressive multi-GPU sharding.

How long does a 5-second clip take to render?

On an RTX 4090: LTX-Video ~90 seconds, Wan 2.2 ~4.5 minutes, HunyuanVideo ~6 minutes, Mochi 1 ~8 minutes. Halve the speed roughly for a 3090, double it for a 3060 12GB on LTX.

Can I do image-to-video locally?

Yes. LTX-Video has a built-in image-to-video mode, and CogVideoX-I2V has a dedicated image-to-video variant. Wan 2.2 has an experimental I2V workflow as of March 2026 but is still rough.

Do these models support audio?

LTX-2.3 does - it is the first open model that generates synchronized audio and video in a single pass (dialogue, ambience, lip-sync), as of its March 2026 release. The other four here (Wan 2.2, LTX-Video 0.9.5, HunyuanVideo, Mochi 1) are silent. For those, generate the video first, then use Bark, MeloTTS, or our local AI voice clone workflow to add narration, and sync in DaVinci Resolve or CapCut.

Will these run on AMD GPUs?

LTX-Video runs on a 7900 XTX 24GB via ROCm 6.1+, slowly. Wan 2.2 and HunyuanVideo have unstable ROCm support as of April 2026. NVIDIA is the path of least pain for video.

How do I avoid the warped-face problem?

Three things: (1) use Wan 2.2 not LTX for face-heavy shots, (2) keep clips at exactly 5 seconds or less, (3) do not request fast head movement. Faces are the hardest thing for any video model.

Can I train LoRAs on top of these models?

Yes for HunyuanVideo (community has shipped LoRA trainers for the wrapper) and partial support for Wan 2.2. LTX-Video LoRA training is experimental. Expect 4-12 hours of fine-tuning on a 4090 for a single subject LoRA.

What is the upper time limit for a single shot?

Practically, 5 seconds. The architectures are trained on roughly that length. Past 5-6 seconds, motion incoherence becomes obvious. To make a 30-second piece, generate six clips with consistent prompts and a fixed seed strategy, then edit.

---

The Honest Takeaway

Local video generation in 2026 is roughly where local image generation was in early 2023 - usable, exciting, occasionally embarrassing - but the LTX-2 release moved the goalposts: open, local, synced-audio video is now a real thing, not a demo. You will not replace Sora for a hero ad spot. You will absolutely replace stock-footage subscriptions for B-roll, blog headers, social loops, and concepting. That is a real workflow shift.

The right starter stack for silent daily work: a used RTX 3090 24GB, 64GB system RAM, ComfyUI, and LTX-Video 0.9.5 for fast iteration plus Wan 2.2 for finals - roughly $900-1,100 used GPU plus existing PC. If you specifically want LTX-2.3's synced audio, plan for a bigger card, or pick one of its lower-precision community quantizations to fit a smaller GPU at reduced quality. Per-clip cost either way: electricity, which is 1/200th of what Runway charges.

If you are not on hardware that fits, the most cost-effective path is to wait six months. The model quality is roughly doubling every two quarters and the VRAM floor is dropping. By Q3 2026, expect Wan-class quality on 16GB cards. By Q1 2027, on 12GB.

For now: get ComfyUI running, install LTX-Video, render your first clip, and decide if the workflow earns the GPU.