Wan: the open-weights baseline on the video side, letting the whole ecosystem iterate on its own hardware
wan
Alibaba Tongyi's Wan releases video generation weights under Apache-2.0, making it the most important public baseline in the video domain. Its value differs from a closed service: not "best-looking output today" but that it is <strong>fine-tunable, reproducible and runnable on your own inference stack</strong> - the whole ecosystem of LoRAs, ControlNet-style conditioning, quantisation and few-step distillation builds on it. For an intelligence site, an open-weights baseline means the question "what can video generation do today" finally has a reference we can verify ourselves instead of trusting vendor-selected samples. The licence and the public repository we have verified; picture quality and physical plausibility remain C-grade vendor claims.
- CONFIDENCE
- Vendor Claim
- Official model card or keynote only, no independent re-test
- KEY METRIC
- 开放权重许可
- Vendor Claim · 2025-07
- MATURITY
- Product
- research → demo → product → production
Our take<p>Wan's role on this ladder is <strong>public reference point</strong>, not "best-looking output tool". The distinction matters: closed video products cannot be fairly compared with each other (each vendor publishes selected samples, prompts and sampling settings undisclosed), whereas an Apache-2.0 open-weights model can be run by anyone on the same hardware with the same prompt set. It therefore became the community's de facto control group - new acceleration, conditioning and distillation methods almost all prove themselves on it.</p><p>Verifiable A-grade facts: the repository is public and the licence is Apache-2.0 (confirmed via the GitHub API). C-grade: the vendor's claims about picture quality, motion range and physical plausibility - we have <strong>not self-hosted a shared prompt comparison</strong>, so those are not used as conclusions.</p><p>One threshold worth stating plainly: self-hosting a video model costs <strong>far more VRAM and time than an image model</strong>. This is not a "download the weights and go" scenario. Batch production either trades quality for quantisation and few-step distillation, or pays the time cost of a cloud service. As a baseline for research and method validation it is well suited; as a production tool, work out the unit cost first.</p>
Why open weights matter more in video than in images
The open-weights ecosystem on the image side is already mature, and the quality gap between closed services and open models has shrunk to the point where most production work cannot feel it. Video is the opposite: the gap between closed-product samples and actual open-weights performance is still visible, and no third party can verify it fairly - you cannot run your own controlled experiment on a closed service, only trust the few dozen selected clips the vendor releases. The practical consequence of that asymmetry is teams greenlighting projects on the strength of a demo, then discovering halfway in that the real water level is much lower.
An Apache-2.0 video baseline breaks this. Anyone can get their own reading, on their own hardware, with their own prompts. For an intelligence site this is decisive: we grade most conclusions in the video domain C not out of pessimism but because reproducible comparisons are missing - open weights are the only path that moves them to B or A.
The ecosystem that grew around it
- Conditioning: reference images, pose, depth and edge maps as extra inputs narrow the freedom of "how it moves" into a controllable range. This is the key layer on the way from lottery to scheduling.
- Few-step distillation and acceleration: compressing dozens of sampling steps into a handful cuts time cost by an order of magnitude at the price of detail and motion coherence - the most active direction right now.
- Style LoRAs: the same fine-tuning path as images, making a specific art direction reusable. Video LoRA training is far more expensive than image LoRA training, so there are fewer of them and each is worth more.
- Inference optimisation: memory tiling, sequence parallelism and quantisation turn "only runs on a large cluster" into "runs short clips on a single workstation GPU".
The real cost account of self-hosting
| Stage | Cost nature | Notes |
|---|---|---|
| VRAM | Hard threshold | Activations and KV memory dwarf an image model; resolution and frame count scale quadratically |
| Per-clip generation time | Linear cost | Dozens of sampling steps, each with full-sequence attention - a 5 second clip often takes minutes |
| Re-roll count | Multiplier | Three to ten attempts per shipped clip is normal; multiply the time cost accordingly |
| Distillation / quantisation | One-off investment | Buys an order-of-magnitude time reduction at a quality cost you must verify is acceptable |
| Storage and transfer | Easily overlooked | Video files are two to three orders of magnitude larger than images; disk and CDN costs matter at batch scale |
Boundaries and failure modes
- Open weights do not mean open capability ceiling: vendors typically keep the strongest tier inside the closed service and release a subset. Inferring a team's real water level from the open baseline underestimates it.
- Physics and temporal consistency: the same problem as the whole domain - fluids, collisions, object conservation and long-horizon identity all break down, worse as clips get longer.
- Prompt sensitivity: open-weights models lack the prompt-rewriting layer of closed products, so the quality of your natural-language description directly determines output quality; the same meaning phrased differently gives noticeably different results.
- Read the licence details: Apache-2.0 covers code and weights, but compliance of generated output and provenance statements about training data are a separate layer that needs legal review before commercial deployment.
Verification backlog (what we do next)
The path to B or A here is clear and expensive: on one fixed machine with one fixed prompt set, place Wan's open tier next to two or three closed services and blind-evaluate temporal consistency, motion range and physical plausibility; record peak VRAM and wall-clock time per clip; run a few-step distilled variant and quantify the quality loss. This is the only asset in our video domain that could plausibly reach an A-grade reading, so it has the highest priority - but it needs a dedicated compute budget.