Wan 2.2 FP8

FP8 is a numerical precision: eight bits per value instead of the sixteen or thirty-two a model is often trained and stored in. Running Wan 2.2 in FP8 is a way to fit it into less memory and move more per second, at some cost in fidelity. wan2-6.net does not publish weight files or precision variants — it runs routes for you. Here is what the choice actually involves.

The three-way trade

Precision buys you three things at once, and you cannot have all of them:

• Memory. Lower precision means smaller tensors, so a card that could not hold the model can now run it.
• Speed. Less data to move per step usually means fewer seconds per clip.
• Fidelity. Fewer bits per value means more rounding error, and the error does not land evenly — it shows first in fine texture, small text and faces.

FP8 sits in the middle: meaningfully lighter than full precision, generally closer to it than an aggressive quantisation. It is the pragmatic choice when you are a little short of memory rather than far short.

Where the loss shows up

Quantisation error is not uniform across a frame. It concentrates where there is high-frequency detail, which is precisely the detail people notice:

• Skin and hair, where small errors read as plastic or wiry.
• Small on-screen text, which smears before anything else does.
• Fast motion, because motion blur and quantisation noise compound in the same regions.

For a wide landscape at a slow camera move, the difference is often negligible. For a close-up portrait, it is usually visible side by side.

Deciding for yourself

The only reliable test is a controlled one: same prompt, same seed, two precisions, compare both. Changing the precision and the prompt together tells you nothing about either. And be careful about judging at a small output band — artefacts that are invisible there can become obvious once the same shot is rendered at the size you intend to publish. See Wan 2.2 Resolution for why the band is part of the route here.

If you would rather not manage precision at all, that is a legitimate reason to host: the components are chosen for you, and the only decisions left are the prompt, the route and the band. Credit costs are shown beside the form before you submit, and prices are on pricing.

Where to go next

Wan 2.2 GGUF for the single-file format, Wan 2.2 VAE Safetensors for the component that decodes latents into frames, and Wan 2.2 Prompts for the part of the workflow that matters most regardless of precision.

Questions people ask

Is FP8 worse than full precision? Slightly, and mostly in fine detail and faces. On wide shots at low motion the difference is often hard to see.

Can this site return FP8 output? No — the precision used on the hosted routes is not a setting you pick, and this site serves no weight files.

Does lower precision make generation cheaper? Locally it can make it faster and let it fit; on a hosted route cost depends on the model, resolution and duration shown beside the form.

Where do I start hosted? Open the Wan workspace and choose the route you need.