One VACE graph with four jobs: replace a masked region, restyle a whole clip under a control signal, extend a clip past its last frame, or outpaint its canvas. Which job it does is decided by what the control video and mask track are fed, which is why the node set changes after rendering.
14Graph nodes
13Node classes
22Inputs
3Models
Tags: motionwan2.1vacevideo-to-videoinpaintoutpaintrunpod-serverlessvolume-backed
Inputs (22)#
The typed parameter surface callers bind when they request this workflow. Enum options and numeric bounds are the values the workflow document declares.
source_videovideorequiredThe clip being edited. Every mode reads it; what is done with it is the mode's business.
modeenumdefault restyleinpaint replaces what the mask covers; restyle regenerates the whole frame under the control signal; extend keeps the first keep_frames and generates the rest; outpaint grows the canvas by the pads and fills the new area.
options:
inpaintrestyleextendoutpaintcontrolenumdefault rawWhat VACE is shown as structure. raw passes the frames through; depth/pose/canny preprocess them; none supplies no control video at all, which is a pure prompt-driven generation inside the mask.
options:
nonedepthposecannyrawmask_videovideoGrayscale mask clip, one frame per source frame — sam3-video-track produces exactly this. Required by `inpaint` and ignored by every other mode.
reference_imageimageOne still VACE conditions on, in any mode. The node takes ONE — the tracker's "0-3" is not what WanVaceToVideo offers.
keep_framesintegerdefault 16min 1max 236extend only: how many frames of the source are kept as known context. The rest are generated.
pad_leftintegerdefault 0min 0max 512outpaint only: pixels added on the left. Pads are in multiples of 8 in the node; values between are rounded by ComfyUI.
pad_rightintegerdefault 0min 0max 512outpaint only: pixels added on the right.
pad_topintegerdefault 0min 0max 512outpaint only: pixels added on top.
pad_bottomintegerdefault 0min 0max 512outpaint only: pixels added below.
promptstringrequireddefault the same scene, cinematic lightingWhat the generated region should contain.
negative_promptstringdefault 色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走Wan's own published negative prompt.
strengthfloatdefault 1.0min 0.0max 2.0How hard the control track binds. 1.0 follows it closely.
resolutionenumdefault 480p_landscape_832x480Frame size before any outpaint pads are added.
options:
480p_landscape_832x480480p_portrait_480x832720p_landscape_1280x720720p_portrait_720x1280lengthintegerdefault 81min 5max 241Frames to produce. Wan wants 4n+1.
fpsfloatdefault 16.0min 1.0max 30.0Frame rate written into the output.
stepsintegerdefault 20min 1max 60Sampling steps.
cfgfloatdefault 5.0min 1.0max 20.0Classifier-free guidance.
shiftfloatdefault 8.0min 0.0max 20.0ModelSamplingSD3 shift, from Wan 2.1's own template.
seedintegerdefault 0min 0Sampler seed.
samplerenumdefault uni_pcSampler. uni_pc is Wan 2.1's template default.
options:
uni_pceulerdpmpp_2mschedulerenumdefault simpleSchedule.
options:
simplenormalbetakarrasComfyUI node graph (14)#
The executable ComfyUI prompt graph: 14 nodes across 13 distinct node classes, wired by 18 data dependencies. Nodes tinted green come from a custom node pack this workflow declares; the rest are ComfyUI core / baked-community classes.
flowchart TD
n0["1: UNETLoader"]
n1["2: CLIPLoader"]
n2["3: VAELoader"]
n3["4: ModelSamplingSD3"]
n4["5: CLIPTextEncode"]
n5["6: CLIPTextEncode"]
n6["7: LoadVideo"]
n7["8: GetVideoComponents"]
n8["9: WanVaceToVideo"]
n9["10: KSampler"]
n10["11: TrimVideoLatent"]
n11["12: VAEDecode"]
n12["13: CreateVideo"]
n13["14: SaveVideo"]
n0 -->|model| n3
n1 -->|clip| n4
n1 -->|clip| n5
n6 -->|video| n7
n4 -->|positive| n8
n5 -->|negative| n8
n2 -->|vae| n8
n7 -->|control_video| n8
n3 -->|model| n9
n8 -->|positive| n9
n8 -->|negative| n9
n8 -->|latent_image| n9
n9 -->|samples| n10
n8 -->|trim_amount| n10
n10 -->|samples| n11
n2 -->|vae| n11
n11 -->|images| n12
n12 -->|video| n13
Nodes (14)#
1UNETLoadercoreunet_name = wan2.1_vace_14B_fp16.safetensorsweight_dtype = defaultoutputs:
MODEL2CLIPLoadercoreclip_name = umt5_xxl_fp16.safetensorstype = wandevice = defaultoutputs:
CLIP3VAELoadercorevae_name = wan_2.1_vae.safetensorsoutputs:
VAE4ModelSamplingSD3coremodel = ◂ node 1 · out[0]shift = {{shift}} tmploutputs:
MODEL5CLIPTextEncodecoreclip = ◂ node 2 · out[0]text = {{prompt}} tmploutputs:
CONDITIONING6CLIPTextEncodecoreclip = ◂ node 2 · out[0]text = {{negative_prompt}} tmploutputs:
CONDITIONING7LoadVideocorefile = {{source_video}} tmploutputs:
VIDEO8GetVideoComponentscorevideo = ◂ node 7 · out[0]outputs:
IMAGEAUDIOFLOATCOMBOCOMBO9WanVaceToVideocorepositive = ◂ node 5 · out[0]negative = ◂ node 6 · out[0]vae = ◂ node 3 · out[0]width = {{resolution_map[resolution].width}} tmplheight = {{resolution_map[resolution].height}} tmpllength = {{length}} tmplbatch_size = 1strength = {{strength}} tmplcontrol_video = ◂ node 8 · out[0]outputs:
CONDITIONINGCONDITIONINGLATENTINT10KSamplercoremodel = ◂ node 4 · out[0]positive = ◂ node 9 · out[0]negative = ◂ node 9 · out[1]latent_image = ◂ node 9 · out[2]seed = {{seed}} tmplsteps = {{steps}} tmplcfg = {{cfg}} tmplsampler_name = {{sampler}} tmplscheduler = {{scheduler}} tmpldenoise = 1.0outputs:
LATENT11TrimVideoLatentcoresamples = ◂ node 10 · out[0]trim_amount = ◂ node 9 · out[3]outputs:
LATENT12VAEDecodecoresamples = ◂ node 11 · out[0]vae = ◂ node 3 · out[0]outputs:
IMAGE13CreateVideocoreimages = ◂ node 12 · out[0]fps = {{fps}} tmploutputs:
VIDEO14SaveVideocorevideo = ◂ node 13 · out[0]filename_prefix = wan21-vaceformat = mp4Parameter banks (3)#
The prompt / configuration lookup tables this workflow keys into from its inputs — the vocabulary that turns a style / palette / preset selection into graph parameters.
resolution_map (4)#
480p_landscape_832x480{"width": 832, "height": 480}480p_portrait_480x832{"width": 480, "height": 832}720p_landscape_1280x720{"width": 1280, "height": 720}720p_portrait_720x1280{"width": 720, "height": 1280}post_render (1)#
vace_modes{"mode_input": "mode", "control_input": "control", "mask_video_input": "mask_video", "reference_image_input": "reference_image", "keep_frames_input": "keep_frames", "pad_inputs": {"left": "pad_left", "right": "pad_right", "top": "pad_top", "bottom": "pad_bottom"}, "target_class": "WanVaceToVideo", "frames_class": "GetVideoComponents", "frames_output": 0}requires_families (3)#
wan21-vacewan-sharedcontrolnet-auxModels & dependencies#
Custom node packs (1)#
The non-core ComfyUI node classes this graph requires; the RunPod worker image the workflow runs on must bake or install a pack that provides every one of them.
comfyui_controlnet_auxModels required (3)#
wan2.1_vace_14B_fp16.safetensorsumt5_xxl_fp16.safetensorswan_2.1_vae.safetensorsOutput contract#
What a successful run of this workflow returns.
primary{"type": "video", "format": "mp4", "codec": "h264", "fps_source": "declared", "audio": false, "alpha": false, "description": "The edited clip."}Taxonomy & routing#
How the control plane classifies this workflow — from the committed workflow-taxonomy-registry.json. It drives the consistency / control surface the agentic director can exercise over the workflow.
assetFamilyregion-edited-videooutputPackageProfilevideo-master-profilecontrolModalitiesmodel-lockprompt-template-locksampler-scheduler-lockseed-locktemporal-lockcontrolnet-depthcontrolnet-posecontrolnet-cannycontrolnet-inpaintreference-ensembleconsistencyDimensionsidentitymotionlightingenvironmentnotesWan 2.1 VACE, four modes in one graph. Every controlnet-* modality is listed because the `control` input really does select between depth, pose and canny preprocessors, and `inpaint` mode really does take a mask track; reference-ensemble covers the optional reference still. The mode is a node-set change applied after rendering, not a value substitution.