SCAIL-2 (Wan 2.1 14B): the character or characters in a reference image replace the people in a driving clip, or perform its motion on a new background. SAM 3.1 tracks every person in both, and each identity is painted one colour in both masks, so several characters bind to several performers in left-to-right order. One segment of the ComfyUI SCAIL-2 character-replacement template (Base subgraph) on its non-turbo schedule, with the DPO and relight LoRAs as toggles.
29Graph nodes
24Node classes
20Inputs
7Models
Tags: motionscail-2wan2.1character-replacementperformance-transfermulti-characterrunpod-serverlessvolume-backed
Inputs (20)#
The typed parameter surface callers bind when they request this workflow. Enum options and numeric bounds are the values the workflow document declares.
reference_imageimagerequiredThe character, or several characters side by side in one image. SAM 3.1 finds each person named by `reference_subject`; in `replacement` the reference is cut out on black by its mask, so its background does not carry over.
driving_videovideorequiredThe performance. Its frames drive the pose from the first frame (downscaled to half size inside WanSCAILToVideo), SAM 3.1 tracks its people for the colour mask, and its frame rate and audio are carried into the output. A clip shorter than `length` gives a shorter output.
promptstringrequireddefault The character from the reference image dances in the scene of the driving video, natural lighting, cinematic.The output video described in full: who, wearing what, doing what, where. The template's own prompt is written that way. SAM 3.1's words are `driving_subject` and `reference_subject`, not this.
negative_promptstringdefault The template leaves it empty; it is used at the default CFG 5.
modeenumdefault replacementreplacement (the template's default): the reference characters take the place of the tracked people in the driving clip, whose scene is kept. animation: the reference characters perform the driving motion and the scene comes from the prompt. Sets replacement_mode on both SCAIL2ColoredMask and WanSCAILToVideo, which flips the masks' background colours.
options:
replacementanimationdriving_subjectstringdefault humanSAM 3.1 text for the people to track in the driving clip (the template's `sam3_video_object`, default `human`). In `replacement` these are who gets replaced; be more specific in a crowded scene.
reference_subjectstringdefault humanSAM 3.1 text for the characters in the reference image (the template's `sam3_image_object`, default `human`). Use the same word as `driving_subject` when the subjects match.
character_indicesstringdefault Which tracked identities to include, comma-separated in `identity_order` (e.g. `0` for only the first, `0,2`); empty is all. Applied to both the reference and the driving masks (SCAIL2ColoredMask object_indices), so it is how one character is chosen for one performer.
identity_orderenumdefault left_to_rightHow identities are numbered and coloured, the same in the reference and the driving clip (SCAIL2ColoredMask sort_by): left_to_right (template) pairs the leftmost reference character with the leftmost performer; area pairs the largest with the largest; none keeps SAM 3.1's order.
options:
noneleft_to_rightareamax_charactersintegerdefault 4min 1max 64Most people SAM 3.1 tracks in the reference and in the driving clip (SAM3_VideoTrack max_objects; the template's 4).
dpo_lorabooleandefault trueLoad wan2.1_SCAIL_2_DPO_lora_bf16 at strength 1.0, as the template always does (its quality LoRA). Off samples the bare SCAIL-2 transformer.
relight_lorabooleandefault falseLoad wan2.1_SCAIL_2_relight_lora_bf16, meant to match the characters' lighting to the scene. No ComfyUI SCAIL-2 template loads it, so it is chained after the DPO LoRA the way the template chains that one, at LoraLoaderModelOnly's default strength 1.0; its effect is unproved until a live render.
resolutionenumdefault 896x512Output size, multiples of 32 as the template rounds them. The driving frames are scaled to it with a centre crop (area filter).
options:
896x512512x896832x480480x832lengthintegerdefault 33min 5max 81Driving frames taken from the start of the clip, 4n+1. 81 is the template's segment (76 new frames plus a 5-frame overlap when extending); 33 is the default so the non-turbo schedule stays inside the 900 s execution timeout (cost_note). Longer clips chain Extend segments, which this graph does not build.
stepsintegerdefault 40min 1max 40Sampling steps; the template's non-turbo value is 40.
cfgfloatdefault 5.0min 1.0max 20.0Classifier-free guidance; the template's non-turbo value is 5.
pose_strengthfloatdefault 1.0min 0.0max 10.0Scales the pose latent (WanSCAILToVideo).
pose_startfloatdefault 0.0min 0.0max 1.0Sampling percent at which pose conditioning starts.
pose_endfloatdefault 1.0min 0.0max 1.0Sampling percent at which pose conditioning ends.
seedintegerdefault 0min 0Sampler noise seed.
ComfyUI node graph (29)#
The executable ComfyUI prompt graph: 29 nodes across 24 distinct node classes, wired by 43 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: LoraLoaderModelOnly"]
n2["3: LoraLoaderModelOnly"]
n3["4: ModelSamplingSD3"]
n4["5: CLIPLoader"]
n5["6: CLIPTextEncode"]
n6["7: CLIPTextEncode"]
n7["8: VAELoader"]
n8["9: CLIPVisionLoader"]
n9["10: LoadImage"]
n10["11: CLIPVisionEncode"]
n11["12: LoadVideo"]
n12["13: GetVideoComponents"]
n13["14: ImageFromBatch"]
n14["15: ResizeImageMaskNode"]
n15["16: GetImageSize"]
n16["17: CheckpointLoaderSimple"]
n17["18: CLIPTextEncode"]
n18["19: SAM3_VideoTrack"]
n19["20: CLIPTextEncode"]
n20["21: SAM3_VideoTrack"]
n21["22: SCAIL2ColoredMask"]
n22["23: WanSCAILToVideo"]
n23["24: KSamplerSelect"]
n24["25: BasicScheduler"]
n25["26: SamplerCustom"]
n26["27: VAEDecode"]
n27["28: CreateVideo"]
n28["29: SaveVideo"]
n0 -->|model| n1
n1 -->|model| n2
n2 -->|model| n3
n4 -->|clip| n5
n4 -->|clip| n6
n8 -->|clip_vision| n10
n9 -->|image| n10
n11 -->|video| n12
n12 -->|image| n13
n13 -->|input| n14
n14 -->|image| n15
n16 -->|clip| n17
n14 -->|images| n18
n16 -->|model| n18
n17 -->|conditioning| n18
n16 -->|clip| n19
n9 -->|images| n20
n16 -->|model| n20
n19 -->|conditioning| n20
n18 -->|driving_track_data| n21
n20 -->|ref_track_data| n21
n5 -->|positive| n22
n6 -->|negative| n22
n7 -->|vae| n22
n14 -->|pose_video| n22
n21 -->|pose_video_mask| n22
n9 -->|reference_image| n22
n21 -->|reference_image_mask| n22
n10 -->|clip_vision_output| n22
n15 -->|length| n22
n2 -->|model| n24
n3 -->|model| n25
n22 -->|positive| n25
n22 -->|negative| n25
n23 -->|sampler| n25
n24 -->|sigmas| n25
n22 -->|latent_image| n25
n25 -->|samples| n26
n7 -->|vae| n26
n26 -->|images| n27
n12 -->|audio| n27
n12 -->|fps| n27
n27 -->|video| n28
Nodes (29)#
1UNETLoadercoreunet_name = wan2.1_14B_SCAIL_2_fp16.safetensorsweight_dtype = defaultoutputs:
MODEL2LoraLoaderModelOnlycoremodel = ◂ node 1 · out[0]lora_name = wan2.1_SCAIL_2_DPO_lora_bf16.safetensorsstrength_model = 1.0outputs:
MODEL3LoraLoaderModelOnlycoremodel = ◂ node 2 · out[0]lora_name = wan2.1_SCAIL_2_relight_lora_bf16.safetensorsstrength_model = 1.0outputs:
MODEL4ModelSamplingSD3coremodel = ◂ node 3 · out[0]shift = 5outputs:
MODEL5CLIPLoadercoreclip_name = umt5_xxl_fp16.safetensorstype = wandevice = defaultoutputs:
CLIP6CLIPTextEncodecoreclip = ◂ node 5 · out[0]text = {{prompt}} tmploutputs:
CONDITIONING7CLIPTextEncodecoreclip = ◂ node 5 · out[0]text = {{negative_prompt}} tmploutputs:
CONDITIONING8VAELoadercorevae_name = wan_2.1_vae.safetensorsoutputs:
VAE9CLIPVisionLoadercoreclip_name = clip_vision_h.safetensorsoutputs:
CLIP_VISION10LoadImagecoreimage = {{reference_image}} tmploutputs:
IMAGEMASK11CLIPVisionEncodecoreclip_vision = ◂ node 9 · out[0]image = ◂ node 10 · out[0]crop = noneoutputs:
CLIP_VISION_OUTPUT12LoadVideocorefile = {{driving_video}} tmploutputs:
VIDEO13GetVideoComponentscorevideo = ◂ node 12 · out[0]outputs:
IMAGEAUDIOFLOATCOMBOCOMBO14ImageFromBatchcoreimage = ◂ node 13 · out[0]batch_index = 0length = {{length}} tmploutputs:
IMAGE15ResizeImageMaskNodecoreinput = ◂ node 14 · out[0]resize_type = scale dimensionsresize_type.width = {{resolution_map[resolution].width}} tmplresize_type.height = {{resolution_map[resolution].height}} tmplresize_type.crop = centerscale_method = areaoutputs:
IMAGE16GetImageSizecoreimage = ◂ node 15 · out[0]outputs:
INTINTINT17CheckpointLoaderSimplecoreckpt_name = sam3.1_multiplex_fp16.safetensorsoutputs:
MODELCLIPVAE18CLIPTextEncodecoreclip = ◂ node 17 · out[1]text = {{driving_subject}} tmploutputs:
CONDITIONING19SAM3_VideoTrackcoreimages = ◂ node 15 · out[0]model = ◂ node 17 · out[0]conditioning = ◂ node 18 · out[0]detection_threshold = 0.5max_objects = {{max_characters}} tmpldetect_interval = 1outputs:
SAM3_TRACK_DATA20CLIPTextEncodecoreclip = ◂ node 17 · out[1]text = {{reference_subject}} tmploutputs:
CONDITIONING21SAM3_VideoTrackcoreimages = ◂ node 10 · out[0]model = ◂ node 17 · out[0]conditioning = ◂ node 20 · out[0]detection_threshold = 0.5max_objects = {{max_characters}} tmpldetect_interval = 1outputs:
SAM3_TRACK_DATA22SCAIL2ColoredMaskcoredriving_track_data = ◂ node 19 · out[0]ref_track_data = ◂ node 21 · out[0]object_indices = {{character_indices}} tmplsort_by = {{identity_order}} tmplreplacement_mode = {{mode == 'replacement'}} tmploutputs:
IMAGEIMAGE23WanSCAILToVideocorepositive = ◂ node 6 · out[0]negative = ◂ node 7 · out[0]vae = ◂ node 8 · out[0]pose_video = ◂ node 15 · out[0]pose_video_mask = ◂ node 22 · out[0]reference_image = ◂ node 10 · out[0]reference_image_mask = ◂ node 22 · out[1]clip_vision_output = ◂ node 11 · out[0]width = {{resolution_map[resolution].width}} tmplheight = {{resolution_map[resolution].height}} tmpllength = ◂ node 16 · out[2]batch_size = 1pose_strength = {{pose_strength}} tmplpose_start = {{pose_start}} tmplpose_end = {{pose_end}} tmplvideo_frame_offset = 0previous_frame_count = 5replacement_mode = {{mode == 'replacement'}} tmploutputs:
CONDITIONINGCONDITIONINGLATENTINT24KSamplerSelectcoresampler_name = euleroutputs:
SAMPLER25BasicSchedulercoremodel = ◂ node 3 · out[0]scheduler = simplesteps = {{steps}} tmpldenoise = 1.0outputs:
SIGMAS26SamplerCustomcoremodel = ◂ node 4 · out[0]add_noise = truenoise_seed = {{seed}} tmplcfg = {{cfg}} tmplpositive = ◂ node 23 · out[0]negative = ◂ node 23 · out[1]sampler = ◂ node 24 · out[0]sigmas = ◂ node 25 · out[0]latent_image = ◂ node 23 · out[2]outputs:
LATENTLATENT27VAEDecodecoresamples = ◂ node 26 · out[1]vae = ◂ node 8 · out[0]outputs:
IMAGE28CreateVideocoreimages = ◂ node 27 · out[0]audio = ◂ node 13 · out[1]fps = ◂ node 13 · out[2]outputs:
VIDEO29SaveVideocorevideo = ◂ node 28 · out[0]filename_prefix = scail2-animateformat = mp4Parameter banks (2)#
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)#
896x512{"width": 896, "height": 512}512x896{"width": 512, "height": 896}832x480{"width": 832, "height": 480}480x832{"width": 480, "height": 832}requires_families (1)#
scail2Models & dependencies#
Models required (7)#
wan2.1_14B_SCAIL_2_fp16.safetensorswan2.1_SCAIL_2_DPO_lora_bf16.safetensorswan2.1_SCAIL_2_relight_lora_bf16.safetensorsumt5_xxl_fp16.safetensorswan_2.1_vae.safetensorsclip_vision_h.safetensorssam3.1_multiplex_fp16.safetensorsOutput contract#
What a successful run of this workflow returns.
primary{"type": "video", "format": "mp4", "codec": "h264", "fps_source": "source", "audio": true, "alpha": false, "description": "The reference characters in the driving performance, at the driving clip's frame rate with its audio track when it has one."}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.
assetFamilyanim-character-turnaround-packoutputPackageProfilevideo-master-profilecontrolModalitiesmodel-lockprompt-template-locksampler-scheduler-lockseed-lockidentity-lockpose-constraintcontrolnet-segmentationtemporal-lockreference-ensembleconsistencyDimensionsidentitymotionlightingwardrobenotesSCAIL-2 (A.02.06). controlnet-segmentation is literal: SAM 3.1 tracks every person in the driving clip and the reference, and SCAIL2ColoredMask paints each identity one colour in both, which binds reference characters to driving performers (reference-ensemble when there are several). lighting is a real dimension only with the relight LoRA on.