Cloud video generators charge by the second. Runway, Kling, Pika, and the rest run roughly $10-$30/month for credit bundles that vanish fast under real production work, and the frontier is metered per second of output. Meanwhile the open-weight video models got genuinely good, and they run on hardware you already own - no API keys, no per-second billing, no uploading your prompts to someone else’s servers.
This is the honest mid-2026 guide to setting up a local video studio: what actually runs on your GPU, where open models still trail the closed leaders, and how to get from zero to a generated clip this weekend.
What You’re Choosing Between
Three open-weight families lead local video generation right now. Each has a different sweet spot.
Wan 2.2 (Alibaba) - the all-round quality leader among fully-open models. It ships in two shapes: a 5B dense text-and-image-to-video model (TI2V-5B) that runs on a single 24GB card such as an RTX 4090 at 720p/24fps, and a much larger 27B mixture-of-experts pair (T2V-A14B / I2V-A14B, 14B active per step) whose official requirement is around 80GB, so the MoE is really a multi-GPU or workstation model. Apache 2.0, no audio. Repo and requirements.
One correction worth making, because the wider web gets it wrong: Wan 2.5, 2.6, and 2.7 are Alibaba Cloud API products, not open weights. Despite many blog posts claiming “Wan 2.7 is open source under Apache 2.0,” the official Wan-Video GitHub and Wan-AI Hugging Face orgs list only Wan 2.1 and Wan 2.2 as downloadable models. If you want to self-host, Wan 2.2 is the current ceiling; the 2.5+ line is only reachable through Alibaba’s paid API.
LTX-2.3 (Lightricks) - the audio-video model. Open-sourced in January 2026 as LTX-2, it was the first open model to generate synchronized audio and video in one pass (footsteps, ambient sound, dialogue). It is a 22B model: the distilled build runs at 720p on about 12GB, an fp8 build fits comfortably on 16-24GB consumer cards, and the full pipeline officially wants 32GB+. It targets 4K and high frame rates; reported maximum clip length has been quoted at both 10 and 20 seconds at different points, so treat exact duration as version-dependent and check the current model card. Apache-family license.
HunyuanVideo-1.5 (Tencent) - the new low-VRAM pick. Tencent cut the original 13B HunyuanVideo down to 8.3B and dropped the floor to about 14GB with model offloading enabled, generating 480p/720p (upscalable to 1080p) at a 121-frame default (~5s at 24fps). It is the model that made “runs on a single mid-range card” real for local video. Repo.
Where Open Still Trails Closed
Be realistic. On the Artificial Analysis text-to-video arena, the best open-weight model (LTX-2.3) scores around 977 Elo, while the closed leaders - ByteDance Seedance 2.0 (~1226), Kling 3.0 (~1112), and Alibaba’s own API-only Wan 2.7 (~1158) - sit clearly ahead, a gap of roughly 150-260 Elo points in blind human preference. Open models have closed most of the spec gap (LTX-2.3 does 4K and synchronized audio, which some closed models do not), but on raw output quality the paid frontier is still ahead. Self-hosting wins on cost, privacy, and control, not on topping the leaderboard.
Hardware Reality Check
| Setup | What you can run | Quality |
|---|---|---|
| 12GB VRAM (RTX 4070) | LTX-2.3 distilled at 720p; HunyuanVideo-1.5 with offloading | Usable for social, drafts |
| 16GB VRAM (RTX 4080, RTX 5080) | LTX-2.3 fp8 with audio; HunyuanVideo-1.5 comfortably | Broadcast-quality shorts |
| 24GB VRAM (RTX 3090, RTX 4090) | Wan 2.2 5B dense; LTX-2.3 fp8 at higher settings | Professional production |
| 32GB+ / multi-GPU | LTX-2.3 full pipeline; Wan 2.2 A14B MoE (needs ~80GB) | Maximum quality |
System RAM matters too - models offload to regular memory when VRAM runs short. 32GB RAM is the practical minimum; 64GB gives headroom.
Option 1: ComfyUI (Most Flexible)
ComfyUI is the standard. It is a node-based interface that supports all three families, and as of the current build (v0.27.1, July 2026, per its changelog) it has native template support for LTX-2 (including the audio+video sampling), Wan 2.2, and HunyuanVideo.
Install ComfyUI
Windows: download the installer from comfy.org; it handles Python, CUDA, and dependencies.
Linux:
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
pip install -r requirements.txt
python main.pyMac (Apple Silicon): same clone, then python main.py --force-fp16. Expect performance well below an NVIDIA card.
The easy path: ComfyUI Manager templates
Rather than hand-placing checkpoints, use ComfyUI Manager. Open the Manager, search the model name (“LTX-2.3”, “Wan2.2”, “HunyuanVideo-1.5”) in Templates, and it downloads the matching checkpoint, VAE, and text encoder and drops in a ready workflow graph. For VRAM-constrained cards, enable an fp8 or quantized cast in the loader node. This avoids the stale-filename problem that breaks older written guides when a repo reorganizes.
A note on LTX audio
LTX-2.3’s distinguishing feature is native, synchronized audio: generated clips come with sound tied to the picture, no separate audio-generation step. If audio-video is your goal, LTX is the family to start with.
Option 2: Wan2GP (Easiest Setup)
If ComfyUI’s node graph feels intimidating, Wan2GP is built to make local video simple. It describes itself as “a fast AI video generator for the GPU poor.”
git clone https://github.com/deepbeepmeep/Wan2GP.git
cd Wan2GP
pip install -r requirements.txt
python app.pyIt downloads models on first run and gives you a clean web interface. Wan2GP leans on aggressive memory management - loading and unloading model components between VRAM and system RAM - so it stretches lower-VRAM cards further than a naive ComfyUI setup, at the cost of ComfyUI’s flexibility (no workflow chaining, no ControlNet). For straight text-to-video or image-to-video, it is the fastest path from zero to a clip.
The Cost Math
Say you generate 50 one-minute videos per month - a moderate creator or small-marketing workload.
- Metered cloud (per-second frontier models): at roughly $0.15/second, 50 minutes of output is on the order of $450+/month.
- Subscription cloud (Kling, Pika, etc.): $10-$30/month plans cap you at a fraction of that volume; heavy use pushes you to enterprise tiers.
- Self-hosted (Wan 2.2 or LTX on your own GPU): electricity only. An RTX 4090 draws ~450W under load; at ~$0.17/kWh, 40 hours of generation a month is about $3.
The upfront GPU cost is real - an RTX 4090 is around $1,600 - but against metered cloud it pays back in months, and you own the card for everything else. An RTX 4070 (12GB) or 4080 (16GB) gets you started at lower resolution.
Privacy Matters
Every cloud video service processes your prompts on their servers. That means creative ideas and scripts pass through third-party infrastructure, generated content may be stored or used for training, and unreleased-product or proprietary work creates IP exposure. Self-hosted generation keeps prompts and output on your machine, under no terms of service governing your results. For businesses generating product videos, prototypes, or internal training content, that is often a compliance requirement, not a convenience.
Tips for Better Results
- Prompting matters more than model size. Describe camera movement, lighting, and composition: “a woman walking through a sunlit forest, medium tracking shot, golden-hour light, shallow depth of field” beats “woman in forest.”
- Draft at low resolution, refine high. Test prompts fast at 480p, then regenerate the keepers at 720p/1080p. It saves hours of waiting on bad generations.
- Use image-to-video for consistency. Generate or source a starting frame, then use I2V mode for far more control over look and composition.
- Negative prompts help. “blurry, distorted hands, text overlay, watermark, low quality” noticeably improves output across all three families.
What You Can Do Today
- On a 12-16GB card, start with HunyuanVideo-1.5 (lowest VRAM floor) or LTX-2.3 distilled/fp8 (if you want audio), via ComfyUI Manager or Wan2GP.
- On a 24GB card, add Wan 2.2 5B dense for its all-round quality, and run LTX at higher settings.
- Want audio-video in one pass? Start with LTX-2.3 - it is the only open family here that generates synchronized sound.
- Remember the ceiling: for self-hosting, Wan 2.2 is the current top open model. Anything labelled “Wan 2.5/2.6/2.7 open source” is mislabelled - those are Alibaba’s paid API.
The gap between cloud and local video narrowed fast. Local generation is a genuine production tool now, and for anyone doing volume, the pricing of metered cloud services is the opportunity.