You can't run Seedance 2.0 locally, and that holds for NSFW video as much as for any other kind. ByteDance has not released the model's weights, so there is nothing to download, and no setup guide can change that. What you can run at home is an open-weight video model such as Wan 2.2 or HunyuanVideo 1.5, which add no moderation layer when you run them yourself. Below: how we know, why the "local Seedance" ComfyUI tutorials are misleading, which open models fit your graphics card, and what five hosted endpoints did with the same still and prompt.
Short answer: Seedance 2.0 is only available through ByteDance's apps and API.
- Have a 24 GB card and want to run things yourself: Wan 2.2 5B, or HunyuanVideo 1.5 if you have 14 GB.
- Want Seedance 2.0 quality: use it hosted. Seedance 2.0 image-to-video is $0.35 for a 5-second 480p clip in our test, and the Spicy edition is $0.57.
- Want the cheapest and fastest clip without a card: Wan 2.2 Spicy took 46.5 seconds and $0.095.
Why there is no local Seedance 2.0
The evidence is on ByteDance's own pages and accounts, checked on 24 October 2026.
- The official Seedance 2.0 page and the launch post of 12 February 2026 point to a chat app (Dola), the BytePlus Playground and an API. Neither has a download link, a GitHub or Hugging Face repository, or a model licence.
- ByteDance does open-source other models. Its official Hugging Face accounts (ByteDance-Seed and ByteDance) list over a hundred models between them, including Seed-OSS, SeedVR2, UI-TARS and BAGEL. None of them is Seedance. Nor is anything on its two official GitHub organisations.
- The repositories that do say "Seedance" on Hugging Face belong to individual accounts, are mostly a README and a promo link, and have zero downloads.
- There is no licence to use. A model licence needs a model file. Seedance 2.0 comes only under app and API terms (Dola, Dreamina and CapCut, BytePlus ModelArk).
If a site offers "Seedance 2.0 weights", an "offline installer" or a "cracked local build", treat it as fake. At best it is a wrapper around a hosted service, at worst it is malware. As of 24 October 2026 ByteDance has announced no plan to release weights, and we won't guess.
The "local Seedance" ComfyUI nodes are API clients
Search "Seedance ComfyUI" and you will find node packs that look local. We read the code of two.
- The community pack
seedance2-comfyuihas a hard-coded remote API address in its main file. It submits a job with an HTTP request, polls for the result and needs an API key. The only thing it does with PyTorch is turn the downloaded frames into tensors. No model is ever loaded. - ComfyUI's own Seedance nodes (in its
comfy_api_nodesfolder) send requests through Comfy's proxy to ByteDance's cloud, using your Comfy account's key. These are paid API nodes.
So the node sits in your ComfyUI graph, but the video is generated on someone else's server, you pay per clip, and the content rules are that service's. That is fine if you want Seedance inside a ComfyUI workflow, just don't mistake it for a local setup. The same goes for any tutorial that says "no GPU needed": that is the tell.
Open models you can actually run at home
Figures below are from each model's official repository unless marked. Download the weights from the official pages, not from re-uploads.
| Model | Size | Minimum VRAM | Licence | 5-second clip on a home card |
|---|---|---|---|---|
| Wan 2.1 (1.3B) | 1.3B | 8.19 GB (official) | Apache 2.0 | About 4 min at 480p on an RTX 4090 (official) |
| HunyuanVideo 1.5 | 8.3B | 14 GB with offloading (official) | Tencent Hunyuan Community License: not valid in the EU, UK or South Korea | 480p image-to-video, step-distilled: within 75 s on an RTX 4090 (official) |
| Wan 2.2 TI2V-5B | 5B | 24 GB, e.g. RTX 4090 (official) | Apache 2.0 | 720p in under 9 min on one consumer card (official) |
| Wan 2.2 A14B | about 27B total, 14B active | 80 GB official; 16-24 GB with community quantisation | Apache 2.0 | About 4 min 20 s at 720p on an RTX 4090 (community report) |
| LTX-2.3 | 22B | 32 GB+ per Lightricks docs (they now list LTX-2.5); about 16 GB with FP8 and offloading (community) | LTX Community License: companies above $10M annual revenue need a commercial licence | No official figure |
How to read it:
- 8-16 GB card: HunyuanVideo 1.5 is the one built for it. Check the territory clause first; if you live in the EU, UK or South Korea you may not use it.
- 24 GB card: Wan 2.2 5B is the officially supported choice and the most permissive licence here.
- Wan 2.2 A14B and LTX-2.3 need quantised builds on consumer cards, and quantised builds are community work you have to trust and tune.
- Sound: LTX-2.3 generates its own soundtrack. The Wan models are silent.
Two cautions. Our "Spicy" editions of Wan 2.2 and LTX 2.3 are hosted endpoints, so you can't download them. The open weights are the base Wan 2.2 and LTX-2.3 models. And we haven't run any of these on a local card for this article, so the timings above are the makers' and community numbers, not ours.
Same still, same prompt: Seedance 2.0 against the open models
To see what you give up by going local, we ran one still and one test prompt through five hosted endpoints on 24 October 2026. These are hosted results, not local runs. The still is a generated adult character in a cafe. Settings were 480p, 5 seconds, seed 20261024, and one run per endpoint. The prompt asks for four actions at a Moderate intensity level with a fixed camera. The test prompts are not published.
| Endpoint | Result | Server time | Cost | Notes on motion and quality |
|---|---|---|---|---|
| Seedance 2.0 Spicy | Delivered, followed the prompt | 4 min 46 s | $0.57 | Did three of the four actions. Face closest to the still, most natural cloth movement; slowest |
| Seedance 2.0 | Delivered | 2 min 41 s | $0.35 | Close to the Spicy result, with about 1.5 s of stillness before the motion starts |
| Wan 2.2 Spicy | Delivered, followed the prompt | 46.5 s | $0.095 | Face matched, smooth motion; skipped one action |
| HunyuanVideo 1.5 | Delivered, followed the prompt | 81.6 s | $0.12 | Skipped one action; cloth movement slightly messy in one moment |
| LTX 2.3 Spicy | Delivered, drifted from the prompt | 38.8 s | $0.095 | Weakest prompt-following, skipped two actions; comes with ambient sound |
What the table says:
- None of the five refused or toned the prompt down at Moderate. They differ in how closely they follow the text, how well they keep the face, how long they take and what they cost.
- Seedance 2.0 and 2.0 Spicy kept the face closest to the still and moved fabric best, at $0.35-$0.57 and 2 min 41 s to 4 min 46 s. The open models cost $0.095-$0.12 and took 39-82 seconds.
- LTX 2.3 Spicy was the fastest and also the loosest: it drifted furthest from the prompt and ignored two of the four actions.
- It's one run each, so treat the ranking as a first impression. Further up the scale our earlier tests show the real gap. Seedance 2.0 Spicy delivered all nine runs from Mild to Strong and one of three at Very strong; Wan 2.2 Spicy delivered every run from Mild to Very strong. Details are in our Seedance versions test, the HunyuanVideo 1.5 test and LTX-2 vs Wan 2.2.
Local GPU or hosted: how to choose
- You own a 24 GB card and like tinkering: run Wan 2.2 5B or, on 14 GB, HunyuanVideo 1.5. Expect to spend time on drivers, ComfyUI graphs and offloading before the first clip, and minutes per clip afterwards.
- You want Seedance 2.0's face consistency and fabric motion: there is no local route. Use Seedance 2.0 or the Spicy edition hosted. Prompt wording is covered in the Seedance 2.0 prompt guide.
- You want open-model results without the card: Wan 2.2 Spicy is $0.095 per 5-second clip at 480p, so a graphics card costing hundreds of dollars equals thousands of hosted clips. If you make a few clips a month, hosted is cheaper; if you make hundreds and already own the card, local wins.
- You need LoRAs: Seedance takes none, and LoRAs don't transfer between model families. The Wan and LTX families are where LoRAs live.
Prices are the public prices on 24 October 2026 and can change. Use only fictional adult characters; don't upload real people's photos.
FAQ
Can you run Seedance 2.0 locally?
No. ByteDance has published no weights, code or licence for local use. Seedance 2.0 is available only through ByteDance's apps and its API, and through hosted services built on that API.
Can you run Seedance 2.0 locally for NSFW video?
No, for the same reason: there are no weights. For unfiltered local video, run an open-weight model such as Wan 2.2 or HunyuanVideo 1.5, which add no moderation layer when you run them on your own card. Hosted Seedance 2.0 and its Spicy edition handled the strongest levels differently; see our Seedance versions test.
Is Seedance 2.0 open source, and is it on Hugging Face or GitHub?
No. ByteDance's official Hugging Face and GitHub accounts host many other models but no Seedance repository. Repositories with "Seedance" in the name belong to individual accounts and contain a README, not a model.
Do the Seedance ComfyUI nodes run on my GPU?
No. The nodes we read send your prompt to a remote API with a key and download the finished video. Your GPU is not used for generation, and the content rules are the service's.
What is the closest open model to Seedance 2.0 that I can run?
Wan 2.2 is the nearest in prompt-following and licence (Apache 2.0), and HunyuanVideo 1.5 has the lowest VRAM requirement. Neither matched Seedance 2.0 on face consistency in our single test, though both were far cheaper and faster.
How much VRAM do I need?
From 8.19 GB for Wan 2.1 1.3B and 14 GB for HunyuanVideo 1.5 with offloading, to 24 GB for Wan 2.2 5B. Wan 2.2 A14B officially needs 80 GB, and LTX-2.3 needs 32 GB or more according to the maker's docs.
