TurboWan I2V Sampler

Don't confuse this with TurboDiffusionI2VSampler. That node is the pack's all-in-one: models, prompt, image, done. This one - TurboWanSampler - is the other path, the "do it the ComfyUI way" route. It doesn't sample anything. It prepares the conditioning and the initial latent for a standard Wan I2V workflow driven by KSamplerAdvanced, exactly the way ComfyUI's own built-in WanImageToVideo node does.

The pack's shipped workflow (turbowan_workflow.json) uses the all-in-one node, so you don't need this one to get results. But if you want KSamplerAdvanced's scheduler control - say, to pair a distilled Wan model with your own sampler/schedule, or because you're already running a standard Wan 2.2 UNet - this is the setup node that makes it possible.

How it works

Give it positive and negative conditioning from a CLIPTextEncode, a VAE, and your target width/height/length, and it builds a blank video latent in Wan's format: [batch, 16, ((length - 1) // 4) + 1, height / 8, width / 8]. The temporal compression (÷4, plus the extra frame) is Wan's, not yours to fiddle with.

Feed it the optional start_image and it gets more interesting: it encodes the image into latent space, then bakes it into both the positive and negative conditioning as a concat_latent_image plus a concat_mask, marking which temporal positions are fixed. That's what makes I2V I2V - the first frames are known and the sampler animates forward from them. There's also an optional clip_vision_output input if you want to add CLIP-vision conditioning on top.

Outputs are positive, negative, and latent - all three go straight into KSamplerAdvanced. It's a genuinely simple node to wire once you see the pattern: it converts "here's my image and dimensions" into "here's a latent and conditioning the sampler understands."

The inputs that matter

• width / height / length - resolution and frame count. length defaults to 121, steps by 8, and the temporal math handles the +1 for you.
• batch_size - default 1; only bump it if you know what you're doing, it multiplies VRAM.
• start_image - the whole point of I2V. Leave it empty and you've built a T2V setup instead.

A word on which path to pick

There are two ways to use this pack, and the README isn't loud about it. The all-in-one TurboDiffusionI2VSampler is the fast, distilled route - 4-step rCM sampling with the int8 TurboDiffusion checkpoints, and it's what the "100x speedup" chatter is about. The TurboWanSampler + KSamplerAdvanced route is the conventional one: you bring your own Wan model and sampler settings, which means no distilled speed unless you add the distilled weights yourself. If you're here for TurboDiffusion's speed, go all-in-one. If you want to stay in familiar ComfyUI territory with full sampler control, use this one.

Troubleshooting

• "Nothing happens / no video out" - this node produces no video; it produces conditioning and a latent. If you haven't wired positive, negative, latent into a KSamplerAdvanced and decoded the result, you'll see nothing.
• Shape errors downstream - the latent is built for a Wan 16-channel video model. Running it through a non-Wan sampler or an SD-class VAE will fail at the shapes, not gracefully.
• start_image silently ignored - the node resizes it to your width/height; if your source is wildly off-aspect it'll letterbox rather than distort, and the effect shows up as a weird first frame.