The Best Open-Source Video Generation Stack in 2026

Open-weight video generation grew up fast. This guide maps the field as of August 2026: which models are worth your disk space, what they actually need in VRAM, how the ComfyUI stack fits together, and which model to pick for each job. It is written for developers and technical makers who run video models on hardware they control, and it covers the generation of models that replaced the early research era the directory already documents.

What changed since the AnimateDiff era

Two years ago, open video generation meant AnimateDiff motion modules bolted onto Stable Diffusion checkpoints, and the state of the art was Genmo's Mochi 1, a 10B Apache 2.0 model from October 2024 that maxed out at 480p and 5.4 seconds. Tencent's original HunyuanVideo closed out 2024 with 720p, 129-frame clips from a full attention transformer, and it signaled the shift that defines the current era: large Chinese labs treating open weights as a distribution strategy rather than a research afterthought.

Three things changed between then and now.

First, the architecture standardized. Every serious open model today is a diffusion transformer paired with a 3D causal VAE, and differentiation moved up a level: mixture-of-experts denoising in Wan 2.2, tile-based sparse attention in HunyuanVideo 1.5, dual-stream audio-video attention in LTX-2.

Second, consumer hardware became a first-class target. FP8 checkpoints, GGUF quantization, block offloading, and step-distillation LoRAs turned models that officially want 80 GB into things that run on a 12 GB card.

Third, audio arrived. Synchronized sound was a closed-model exclusive until Lightricks opened LTX-2 in January 2026, and MiniMax H3, an open-weights audio-video model, joined it in August 2026.

The other lesson of the period was about trust. Alibaba pre-announced Wan 2.5 with an open release implied, then shipped it API-only. The weights never came. Open-when-convenient is now a real risk you plan around.

Wan 2.2: still the center of gravity

Wan 2.2 (Alibaba, released July 28, 2025, Apache 2.0) remains the default answer to most questions in 2026, not because it is the newest but because everything plugs into it. The release has three core checkpoints. T2V-A14B and I2V-A14B are 27B-parameter mixture-of-experts models that split denoising between a high-noise expert and a low-noise expert, with roughly 14B parameters active per step, so you pay mid-size inference cost for big-model quality. TI2V-5B is a dense model that does both text-to-video and image-to-video at 720p and 24 fps on a single consumer GPU. Alibaba later added S2V-14B for audio-driven speech-to-video (August 2025) and Animate-14B for character animation and replacement (September 2025).

The official pipeline is honest about its appetite: the README commands for the A14B models assume an 80 GB card. Almost nobody runs it that way. The community stack of FP8 checkpoints, GGUF quants, and offload flags is the real interface, covered below. If you do run the reference code, the memory flags are what make smaller cards viable:

git clone https://github.com/Wan-Video/Wan2.2.git
cd Wan2.2
pip install -r requirements.txt
python generate.py --task t2v-A14B --size 1280*720 --ckpt_dir ./Wan2.2-T2V-A14B --offload_model True --convert_model_dtype --t5_cpu --prompt "a corgi surfing at golden hour"

The weaknesses are just as concrete. Native clips are about 5 seconds. The A14B MoE design means juggling two DiT files in every workflow. There is no audio. And the upstream faucet may be closed: as of August 2026 the official Wan base-model weights still stop at the 2.2 family: newer generations like Wan 2.5 live behind an API, though task-specific models still go open, most recently the Wan Animate 2 character animation release in August 2026.

LTX-2 and LTX-2.5: audio plus video, with a license asterisk

LTX-2 is the most complete single model in open video right now. Lightricks announced it in October 2025 and released the full package of weights, inference code, and training code in January 2026. It is a 19B model split into a 14B video stream and a 5B audio stream, and it generates synchronized video and audio in one pass: dialogue, ambience, and effects that match the picture. Output reaches 4K at up to 50 fps. It comes from the same team as LTX-Video, the speed champion of the previous generation, and inherits that focus: distilled FP8 builds run on a single consumer GPU, though 24 GB and up is where it gets comfortable.

LTX-2.5 followed in August 2026: a 22B asymmetric dual-stream diffusion transformer whose video and audio streams are linked by bidirectional cross-attention. It adds native 4K HDR output, automatic duration prediction through an optional duration head, and native multishot generation with character consistency across cuts. It ships in BF16, distilled, int8, and NVFP4 variants with day-one ComfyUI workflow templates.

The asterisk is the license. LTX weights are free for research and for commercial use by companies under 10 million USD in annual revenue. For an indie developer that is functionally free; for anyone larger it is a legal conversation, and it is not Apache or MIT no matter what the launch posts imply. The ecosystem is also younger than Wan's: fewer community LoRAs, fewer control adapters, less tribal knowledge to borrow.

HunyuanVideo 1.5: the most model per gigabyte

HunyuanVideo 1.5 (Tencent, November 20, 2025) is an 8.3B diffusion transformer with a 3D causal VAE and a selective and sliding tile attention (SSTA) mechanism that prunes redundant spatiotemporal computation. It generates 480p and 720p text-to-video and image-to-video natively, with 1080p available through a super-resolution stage, at a default 121 frames. The headline number: the full pipeline runs in 14 GB of VRAM with model offloading enabled. Tencent also ships step-distilled 480p checkpoints that cut inference to 8 or 12 steps for roughly a 75 percent speedup, plus CFG-distilled and sparse-attention variants.

For quality per gigabyte nothing else comes close, and the distance from the original HunyuanVideo, a datacenter-card model at launch, shows how fast the efficiency frontier moved in one year. The catch is the license: custom Tencent terms rather than an OSI-approved license. Read it before building a product on it.

Long video and the rest of the field

LongCat-Video (Meituan, 13.6B, MIT license) unifies text-to-video, image-to-video, and video continuation in one diffusion transformer, and the continuation pretraining is the point: it generates minute-scale 720p, 30 fps video without the quality collapse that plagues naive autoregressive stitching. The MIT license is the cleanest in the top tier, full stop.

SkyReels-V2 (Skywork, April 2025) took a different route to long video: diffusion forcing, where each frame carries its own noise level, enabling unbounded-length generation from 1.3B and 14B checkpoints at 540p and 720p. Magi-1 (Sand AI) generates autoregressively in chunks through a FIFO denoising pipeline, which makes it the interesting one for streaming and interactive uses. ByteDance's Waver 1.0, a 12B rectified flow transformer, handles text-to-video, image-to-video, and text-to-image in one model with flexible resolution and duration, though its licensing is thinly documented and deserves a careful read before commercial use.

FramePack (from lllyasviel of ControlNet fame) attacks a different axis entirely. It is a next-frame-section prediction architecture that compresses input context to a constant length, so generation cost does not grow with video length. It runs 13B-class models in 6 GB of VRAM and generates clips up to 60 seconds, 1800 frames at 30 fps, on laptop GPUs. It is image-to-video centric and slower per frame than the distilled big models, but nothing else touches its VRAM floor.

VRAM realities: what actually runs on your card

The official requirements and the practical ones are different documents. Quantization and offloading define the real tiers.

6 to 8 GB. FramePack is the honest option at 6 GB. Wan 2.2 TI2V-5B fits 8 GB cards under ComfyUI's native offloading and still delivers 720p at 24 fps. GGUF quants of the A14B models technically go this low (Q2_K is a 5.3 GB file) but artifacts mount fast below 4-bit; treat Q4 as the sane floor.

12 to 16 GB. The sweet spot in 2026. HunyuanVideo 1.5 runs its full pipeline in 14 GB with offloading. QuantStack's GGUF conversion of the Wan 2.2 T2V DiT spans 5.3 GB (Q2_K) to 15.4 GB (Q8_0) per model file, with Q4_K_M at 9.65 GB, and the two noise experts load sequentially rather than together, so a Q4 build with the text encoder offloaded is workable here. LTX-2's distilled FP8 build targets this class of card, with offloading picking up the slack.

pip install -U "huggingface_hub[cli]"
huggingface-cli download QuantStack/Wan2.2-T2V-A14B-GGUF --include "*Q4_K_M*" --local-dir ComfyUI/models/unet

24 GB. Everything above runs comfortably, FP8 A14B workflows become routine, HunyuanVideo 1.5 runs without aggressive offloading, and video LoRA training becomes possible on the same card.

48 GB and up. The official pipelines: Wan A14B wants 80 GB on a single GPU, with multi-GPU FSDP and sequence-parallel setups as the alternative. Rent it by the hour unless you generate daily.

Two budget rules save the most grief. Attention cost grows superlinearly with token count, and tokens scale with resolution times frames, so 720p at 121 frames is drastically heavier than 480p at 81 frames, not proportionally so. And the DiT is not the whole budget: the text encoder and the VAE decode have their own memory peaks, which is why flags like --t5_cpu exist at all.

If you would rather not assemble any of this yourself, Wan2GP packages the low-VRAM path as a web UI: Wan 2.1 and 2.2, HunyuanVideo, LTX-Video, and LTX-2 alongside image models, with tuned profiles down to 6 GB of VRAM and support for cards as old as the GTX 10XX series plus AMD RDNA 2 through 4.

git clone https://github.com/deepbeepmeep/Wan2GP.git
cd Wan2GP
pip install -r requirements.txt
python wgp.py

The ComfyUI layer, and how to make it fast

ComfyUI is the de facto runtime for all of this. Wan 2.2, HunyuanVideo 1.5, and LTX-2.5 each had day-one native support with official workflow templates, and the native FP8 checkpoints plus built-in offloading are what turn paper requirements into consumer ones. ComfyUI-GGUF adds the quantized DiT loader: drop the file into models/unet and swap the loader node. Kijai's wrapper nodes, which usually land new video features days before native support, are the other half of the ecosystem worth following.

The speed stack matters as much as the models:

• Lightning LoRAs. The lightx2v Wan2.2-Lightning LoRAs distill the default 40-step, CFG-guided schedule down to 4 steps with no CFG, a claimed 20x speedup; common ComfyUI workflows run them at 4 steps per expert. A 480p, 5-second clip that takes tens of minutes stock on an RTX 4090 drops to a few minutes. The cost is real but modest: some detail and motion dynamics flatten out.
• FastVideo. The FastWan builds (August 2025) use sparse distillation, training video sparse attention and distribution matching distillation jointly. FastWan2.2-TI2V-5B generates in 3 denoising steps for up to a 50x denoising speedup at 720p, and the 1.3B variant clears its denoising loop in about a second on an H200.
• First-party distills. HunyuanVideo 1.5's 8 and 12 step checkpoints and LTX-2's distilled variants come from the original labs, so the quality tradeoff is better controlled than with community distillation LoRAs.

Control, editing, and character consistency

Raw text-to-video is a demo; control is a workflow, and this is where Wan's ecosystem lead is widest. VACE is an all-in-one creation and editing layer over Wan (1.3B and 14B variants) covering subject replacement from reference images, first-and-last-frame control, object modification inside an existing video, and outpainting. Phantom handles subject-consistent generation from one or several reference images, the closest open equivalent to the consistent-character features closed platforms sell. Wan's own S2V-14B (audio-driven video) and Animate-14B (motion and expression transfer onto a character) fill out the set.

For custom styles and characters, musubi-tuner from kohya-ss is the standard Wan 2.2 LoRA trainer: 12 GB of VRAM is enough for image-based training at 960x544 or below using the blocks_to_swap and fp8 options, 24 GB or more is recommended for video training, and rank 16 is the common default for character LoRAs. diffusion-pipe is the other established trainer for video DiTs. Training a LoRA on a few dozen images of a character, then driving it through VACE or an image-to-video workflow, is the bread-and-butter pattern of 2026 production work.

Picks per use case

Best overall: Wan 2.2 A14B. The deepest ecosystem in open video: LoRAs, VACE, Phantom, GGUF quants, lightning LoRAs, and every UI supports it first. Weaknesses: about 5 seconds per native clip, no audio, two expert files to manage, and no open successor to the base models in sight.

Best when audio matters: LTX-2, or LTX-2.5 if your hardware is current. Single-pass synchronized audio and video at up to 4K exists nowhere else in open weights; MiniMax H3, the one other open audio-video model as of August 2026, tops out at 2K. Weaknesses: the sub-10M revenue license needs reading, full-precision weights are heavy, and the community ecosystem is a year behind Wan's.

Best under 16 GB: HunyuanVideo 1.5. 8.3B parameters, 14 GB with offloading, first-party distills, 1080p via super-resolution. Weaknesses: custom license, no audio, and a lower ceiling than the 20B-class models.

Best under 8 GB: FramePack, then Wan 2.2 TI2V-5B. FramePack generates minutes of video in 6 GB; the 5B Wan runs 720p at 24 fps in 8 GB under ComfyUI. Weaknesses: FramePack is image-to-video centric and slow per frame, and the 5B Wan clearly trails the A14B on complex motion.

Best license for a commercial product: LongCat-Video. MIT, 13.6B, minute-scale output. Weaknesses: smaller community, fewer control adapters, and out-of-the-box aesthetics trail a LoRA-tuned Wan.

Best long-form: LongCat-Video and SkyReels-V2. Continuation pretraining and diffusion forcing are the two credible answers to the 5-second wall. Weakness: coherence over minutes still degrades in busy scenes, so plan on curation and retries.

And the caveat that applies to every pick: open models still trail the Veo 3 and Seedance class of closed APIs on prompt adherence and physics, hands and object interactions still break, and nothing open sustains a coherent multi-minute narrative without a human in the loop. The gap narrowed dramatically since 2024; it did not close.

How to choose, and what to watch

Choosing is mostly three questions. Does the output need sound? If yes, LTX-2 is the main single-model answer, license permitting, with MiniMax H3 the newer open alternative. Is this going into a commercial product at scale? If yes, Apache 2.0 (Wan 2.2) and MIT (LongCat-Video) beat the custom licenses, and your lawyer will agree. What is the VRAM budget? 6 to 8 GB points to FramePack or the 5B Wan, 12 to 16 GB to HunyuanVideo 1.5 or a quantized Wan A14B, 24 GB opens everything quantized, and the full-precision official pipelines stay rental-GPU territory.

Three things are worth watching for the rest of 2026. Whether Alibaba ever opens a post-2.2 Wan base model: the largest ecosystem in open video is coasting on a July 2025 base model, and the Wan 2.5 precedent says not to assume. Sparse attention plus distillation: the techniques behind FastWan and SSTA took 720p clips from tens of minutes to a couple of minutes in a year and are pointed squarely at real time. And the LTX-2.5 quant ecosystem: a 22B audio-video model that runs well on a 24 GB card would reset the defaults the way Wan 2.2 did last summer.