WAN Tiled Sampler

If you've generated WAN at 1080p and wondered why the colors drifted, why everything looks washed out, or why the conditioning seems to have evaporated - that's not a bad seed. WAN models are trained at specific spatial resolutions, and sampling way outside that range pushes token counts and positional encodings off the rails. The community's standard answer is "generate at 480p, upscale later." WanTiledSampler is the "no, generate at the target resolution directly" answer: a drop-in replacement for KSamplerAdvanced that adds per-step MultiDiffusion spatial tiling and a coarse-to-fine multiscale schedule, built specifically for WAN 2.1 / 2.2 video.

It's the flagship node of the MagoNodes pack, and it's genuinely two features stacked. Tiling solves the spatial problem; the multiscale schedule solves the temporal one. You want both, so it's worth understanding what each actually does.

How it works

Tiling. Each denoising step is split into overlapping spatial tiles, the model runs each tile at a training-distribution token count, and the predictions are blended back into one full-resolution estimate before the sampler takes its step. Because the blend happens at every sigma step, all tiles ride a single denoising trajectory - content never diverges between tiles. Tiling is spatial only; temporal frames are never split, so causal temporal attention stays intact. This works great for V2V, where the control video pins motion per tile.

Multiscale (the I2V fix). Tiling alone can't fix I2V, because each tile decides its own motion from local context and the tiles drift apart over time. The scale_schedule input fixes that: during early, high-noise steps - when global motion and layout get decided - the whole frame is run downscaled in a single pass; then resolution ramps up and tiling takes over for detail. WAN derives its positional encodings from the latent's spatial size, so a smaller latent automatically brings tokens and positions back in-distribution, and one shared low-res pass keeps motion globally coherent.

One subtle bit the author got right: the noisy latent is downscaled by subsampling, not averaging. Averaging collapses the variance of i.i.d. noise, which would hand the model a latent whose noise level no longer matches the timestep and produce garbage. Subsampling keeps unit-variance noise. The conditioning tensors are clean signal, so those downscale with antialiased area.

The inputs that matter

It's a KSamplerAdvanced clone, so you get all the familiar stuff: model, positive, negative, latent_image, steps, cfg, sampler_name, scheduler, noise_seed, start_at_step / end_at_step, add_noise, return_with_leftover_noise. The rest are the new controls:

• tiles_h / tiles_w - spatial tiles, default 2×2. 1 disables that axis.
• overlap_h / overlap_w - overlap as a percentage of tile size (0–50%), default 25. The guidance: 12% for static scenes, 25% recommended, 50% for maximum continuity at ~2× compute.
• scale_schedule - the multiscale plan, e.g. {0:25, 10:50, 20:100}. Empty = always full-res. The pack's recommended I2V recipe is a gentle ramp: {0:50, 15:100}.
• bypass_tiling - skip tiling entirely (unless a schedule is set).
• reference_latent_full - keep it on for I2V; it passes the full reference frame to every tile as global context instead of cropping it.

Output is LATENT, straight into a VAE decode. Note the cfg default of 1 - WAN distilled models run at cfg 1, so this node starts where you'd end up anyway.

Installing

In the MagoStudio/ComfyUI-MagoNodes pack: ComfyUI Manager → search "ComfyUI-MagoNodes", or

cd ComfyUI/custom_nodes
git clone https://github.com/MagoStudio/ComfyUI-MagoNodes

Restart ComfyUI. No extra dependencies.

Where people get burned

• More tiles = lower per-tile memory but more model evals per step. Start 2×2 and only go up if you still see drift or OOM.
• The schedule's step indices are relative to the sampled range, so they respect start_at_step / end_at_step.
• In a WAN 2.2 high/low-noise two-sampler setup, each node has its own schedule and step indices restart at 0 per pass. The latent handed between passes is always full-res - the downscaling lives inside each model eval and never persists.
• Turn on debug for your first run - it prints the tile layout and per-step decisions, which is the fastest way to confirm tiling is actually firing.