I2V is image-to-video: you supply a first frame and the model animates from it. Adding GGUF means that model has been quantised into a single weight file so it can run on a smaller card. wan2-6.net does not distribute these files — if your interest is the hosted version of the same task, that is image to video. This page is about the trade-off.
Text-to-video asks the model to build a scene from words, so the model has freedom about what it draws. Image-to-video has no such freedom: it must preserve the frame you supplied while inventing plausible motion inside it. That splits the workload in an awkward way — the weights still have to encode the motion, but the VAE now has to reconstruct a specific image faithfully, and reconstruction errors show up as drift or mush on exactly the details you chose the image for.
The practical consequence: a quantisation that looks acceptable for text-to-video can look noticeably worse for image-to-video, because there is a ground truth to compare against. Faces, fine texture and sharp edges are where you will notice it first.
And keep the source image honest: a soft or noisy first frame animated by a quantised model gives you a soft result with motion in it, which is rarely what anyone wanted. Sharpen or re-shoot before you animate.
Run it locally when you are doing many image-to-video generations with your own pipeline and a card that comfortably fits the quantisation. Use the hosted route when your card is marginal, when the install is not finished, or when you want the cost of a test to be visible before you commit to it — the credit cost appears beside the form before submission, and credit prices are on pricing.
Either way, treat the first render as a test rather than an asset: output quality varies by route and by how large a movement you ask for.
Wan 2.2 GGUF covers the format itself, image to video is the hosted route, and Wan 2.2 Animate goes into getting a still that animates well.
Does this site offer the I2V GGUF build? No — this site runs image-to-video for you; it does not provide weight files.
Is quantised image-to-video worse? Usually slightly, and most visibly on faces and fine detail where you have a fixed first frame to compare against.
What is the minimum I need to run it locally? Enough VRAM to hold the quantised weights plus the attention workspace and the VAE. Short low-resolution clips fit in far less than the numbers quoted for full-size renders.
Where do I start on the hosted route? Open the Wan workspace and choose the image-to-video tab.