ComfyUI Node
WAN Tiled Sampler
A ComfyUI node in Mago Nodes/Sampling with 21 inputs and 1 output.
WAN Tiled Sampler
- model
- positive
- negative
- latent_image
- LATENT
◄add_noise▾►
◄noise_seed0►
◄steps20►
◄cfg1.0►
◄sampler_name▾►
◄scheduler▾►
◄start_at_step0►
◄end_at_step10000►
◄return_with_leftover_noise▾►
◄tiles_h2►
◄tiles_w2►
◄overlap_h25►
◄overlap_w25►
◄scale_schedule►
◄bypass_tilingfalse►
◄reference_latent_fulltrue►
◄debugfalse►
CategoryMago Nodes/Sampling
Inputs (21)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| add_noise | COMBO | 2 options: enable, disable | |
| noise_seed | INT | 00–18446744073709550000 | — |
| steps | INT | 201–10000 | — |
| cfg | FLOAT | 1.00–100 | Classifier-Free Guidance scale. WAN distilled models typically use cfg=1. |
| sampler_name | COMBO | 44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38 | |
| scheduler | COMBO | 9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3 | |
| positive | CONDITIONING | — | |
| negative | CONDITIONING | — | |
| latent_image | LATENT | — | |
| start_at_step | INT | 00–10000 | — |
| end_at_step | INT | 100000–10000 | — |
| return_with_leftover_noise | COMBO | 2 options: disable, enable | |
| tiles_h | INT | 21–8 | Number of spatial tiles along the height axis. Set 1 to disable height tiling. |
| tiles_w | INT | 21–8 | Number of spatial tiles along the width axis. Set 1 to disable width tiling. |
| overlap_h | INT | 250–50 | Overlap between height tiles as a percentage of tile height (0–50%). 25% = one quarter of each tile overlaps its neighbour. Larger values reduce seam artifacts from cross-tile motion at the cost of more compute per step. |
| overlap_w | INT | 250–50 | Overlap between width tiles as a percentage of tile width (0–50%). 25% = one quarter of each tile overlaps its neighbour. |
| scale_scheduleopt | STRING | Multiscale (coarse-to-fine) schedule mapping step → resolution %. e.g. {0:25, 10:50, 20:100} runs the WHOLE frame downscaled to 25% for steps 0–9, 50% for 10–19, then full-res from 20. Values are held until the next key. While scale < 100% the frame is evaluated in a single pass (tiling is skipped) so global motion stays coherent — the key fix for I2V drift. Empty = always 100%. | |
| bypass_tilingopt | BOOLEAN | false | Skip tiling entirely — equivalent to plain KSamplerAdvanced (unless a scale_schedule is set). |
| reference_latent_fullopt | BOOLEAN | true | When enabled, the full reference frame is passed to every tile instead of cropping it to the tile region. Useful for I2V workflows: the reference tokens are appended to the sequence so there is no size constraint, and the full image gives each tile global scene context. |
| debugopt | BOOLEAN | false | Print tile layout, blend weight sanity check, per-step scale/tile decisions, and conditioning slice info. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| LATENT | LATENT | — |