Nodes/comfyUI-TiledWan/TiledWan Video VACE Pipeline
ComfyUI Node

TiledWan Video VACE Pipeline

Wan2.1 VACE video processing node that applies WanVideo VACE pipeline to large videos through tiling. Enables processing of arbitrarily large videos that would otherwise exceed memory or models limits. Uses dimension-wise tiling with temporal and spatial overlap to maintain quality and consistency. The complete WanVideo VACE pipeline is applied to each tile with overwriting on overlapping regions for consistency. Processing Algorithm: 1. Temporal tiling: Split video into chunks with overlap (default: 81 frames, 10-frame overlap) 2. Spatial tiling: Split each chunk into tiles with overlap (default: 832×480, 20-pixel overlap) 3. Temporal consistency: Previous chunks provide reference frames for upcoming chunks 4. Spatial consistency: Already-processed neighboring tiles overwrite overlapping regions 5. WanVideo VACE processing: Each tile processed through complete pipeline 6. Dimension-wise stitching: Column-wise → Line-wise → Temporal stitching 7. Final cropping: Output matches exact input dimensions Key Features: - Handles any video size through tiling - Temporal consistency across chunks via frame reference chaining - Spatial consistency through neighbor tile overwriting - Memory-efficient with model offloading between tiles - Complete WanVideo VACE pipeline integration - Fade blending for seamless stitching - Tensor safety with hard copies to prevent data contamination Consistency Mechanisms: - Temporal: Last frames from previous chunks overwrite first frames of current chunks - Spatial: Left/top neighbors overwrite overlapping edges in current tiles - All overwritten regions have masks zeroed for proper processing - Reference frames are used to maintain temporal coherence across temporal chunks

By Baverne·Created about a year ago·Updated 6 months ago· 8
TiledWan Video VACE Pipeline
  • video
  • mask
  • model
  • vae
  • text_embeds
  • samples
  • feta_args
  • context_options
  • loop_args
  • sigmas
  • unianimate_poses
  • fantasytalking_embeds
  • uni3c_embeds
  • multitalk_embeds
  • freeinit_args
  • cache_args
  • slg_args
  • experimental_args
  • vace_ref_images
  • processed_video
  • processing_info
target_frames81
target_width832
target_height480
frame_overlap10
spatial_overlap20
steps30
cfg6.00
shift5.00
seed0
schedulerunipc
vace_strength1.000
vace_start_percent0.00
vace_end_percent1.00
decode_enable_vae_tilingfalse
decode_tile_x272
decode_tile_y272
decode_tile_stride_x144
decode_tile_stride_y128
debug_modetrue
debug_color_shiftfalse
force_offload_between_tilestrue
riflex_freq_index0
denoise_strength1.00
force_offloadtrue
batched_cfgfalse
rope_functioncomfy
vace_tiled_vaefalse
decode_normalizationdefault
CategoryTiledWan

Inputs (47)

NameTypeDefaultDescription
videoIMAGEInput video frames to process. Can be very large videos that will be automatically tiled for memory efficiency.
maskMASKMask defining areas to be processed. Should match the video dimensions and frame count.
modelWANVIDEOMODELWanVideo model for video generation/processing.
vaeWANVAEVideo VAE for encoding/decoding frames.
target_framesINT8116–200Target number of frames per temporal chunk. The model works best with 81 frames. If different, might generate artefact or frame drops
target_widthINT83264–2048Target width for spatial tiles. Works best around 832.
target_heightINT48064–2048Target height for spatial tiles. Works best around 480.
frame_overlapINT100–40Number of overlapping frames between temporal chunks.
spatial_overlapINT200–100Pixel overlap between spatial tiles.
stepsINT30Number of diffusion sampling steps.
cfgFLOAT6.000–30Classifier-free guidance strength. Higher values follow the conditioning more closely.
shiftFLOAT5.000–1000Shift parameter for the sampling schedule.
seedINT00–18446744073709550000Random seed for reproducible results. Use the same seed to get identical outputs.
schedulerCOMBOunipcSampling scheduler algorithm.
vace_strengthFLOAT1.0000–10VACE encoding strength. Controls how much the video is modified during processing. 1.0 mean total reconstruction, 0.0 means no change.
vace_start_percentFLOAT0.000–1
vace_end_percentFLOAT1.000–1
decode_enable_vae_tilingBOOLEANfalse
decode_tile_xINT27240–2048
decode_tile_yINT27240–2048
decode_tile_stride_xINT14432–2040
decode_tile_stride_yINT12832–2040
debug_modeBOOLEANtrue
debug_color_shiftBOOLEANfalseEnable color shift debugging to visualize tile boundaries. Useful for debugging tiling artifacts.
force_offload_between_tilesBOOLEANtrue
text_embedsoptWANVIDEOTEXTEMBEDS
samplesoptLATENT
riflex_freq_indexoptINT00–1000
denoise_strengthoptFLOAT1.000–1
force_offloadoptBOOLEANtrue
batched_cfgoptBOOLEANfalse
rope_functionoptCOMBOcomfy2 options: default, comfy
feta_argsoptFETAARGS
context_optionsoptCONTEXTOPTIONS
loop_argsoptLOOPARGS
sigmasoptSIGMAS
unianimate_posesoptUNIANIMATE_POSES
fantasytalking_embedsoptFANTASYTALKING_EMBEDS
uni3c_embedsoptUNI3C_EMBEDS
multitalk_embedsoptMULTITALK_EMBEDS
freeinit_argsoptFREEINIT_ARGS
cache_argsoptCACHEARGS
slg_argsoptSLGARGS
experimental_argsoptEXPERIMENTALARGS
vace_ref_imagesoptIMAGE
vace_tiled_vaeoptBOOLEANfalse
decode_normalizationoptCOMBOdefault2 options: default, minmax

Outputs (2)

NameTypeDescription
processed_videoIMAGEProcessed video with WanVideo VACE pipeline applied through tiled processing. Maintains original dimensions and quality.
processing_infoSTRINGDetailed information about the processing including tile counts, timing, and processing summary.