Wan Segment Chain — ComfyUI workflow

Long-form chunk chaining on Wan 2.2 TI2V-5B: renders the NEXT segment of a longer film conditioned on the previous segment's tail frame, color-locks the result against a canonical reference frame (cross-segment drift is the documented FLF failure mode), and emits both the segment video and its own tail frame for the next link in the chain

17Graph nodes

15Node classes

12Inputs

4Models

Tags: flagshipwan2.2videosegment-chainlong-formcolor-lock2026-sota

Inputs (12)#

The typed parameter surface callers bind when they request this workflow. Enum options and numeric bounds are the values the workflow document declares.

promptstringrequired

Motion-and-scene description for THIS segment

previous_tail_frameimagerequired

Last frame of the previous segment (the chain anchor)

reference_frameimage

Canonical color reference (usually the first segment's first frame); when provided, every rendered frame is color-matched to it

color_lockbooleandefault true

Reinhard Lab-space color transfer of the whole segment onto the reference frame (needs reference_frame)

resolutionenumdefault 720p_landscape

TI2V-5B native resolution (multiples of 32)

options: 720p_landscape720p_portrait

framesintegerdefault 121min 25max 121

Segment length in frames (24fps; must be 4n+1 - the latent quantizes in steps of 4)

fpsintegerdefault 24min 24max 24

TI2V-5B is trained at 24fps

stepsintegerdefault 20min 10max 30

Sampling steps (official reference: 20; ignored in fast_mode)

guidancefloatdefault 5.0min 1.0max 10.0

CFG (official reference: 5.0; ignored in fast_mode)

shiftfloatdefault 8.0min 1.0max 12.0

ModelSamplingSD3 shift (official reference: 8.0)

fast_modebooleandefault true

Turbo distill (Self-Forcing rank-64 LoRA): 4 steps without CFG - the segment-economics default for chains, where many segments amortize small per-segment quality deltas

seedintegerdefault -1

Random seed (-1 for random)

ComfyUI node graph (17)#

The executable ComfyUI prompt graph: 17 nodes across 15 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["turbo_lora: LoraLoaderModelOnly"] n4["4: ModelSamplingSD3"] n5["5: CLIPTextEncode"] n6["6: CLIPTextEncode"] n7["7: LoadImage"] n8["8: Wan22ImageToVideoLatent"] n9["9: KSampler"] n10["10: VAEDecode"] n11["ref_frame: LoadImage"] n12["color_lock_node: ColorTransfer"] n13["11: CreateVideo"] n14["12: SaveVideo"] n15["tail_pick: ImageFromBatch"] n16["13: SaveImage"] n0 -->|model| n3 n3 -->|model| n4 n1 -->|clip| n5 n1 -->|clip| n6 n2 -->|vae| n8 n7 -->|start_image| n8 n4 -->|model| n9 n5 -->|positive| n9 n6 -->|negative| n9 n8 -->|latent_image| n9 n9 -->|samples| n10 n2 -->|vae| n10 n10 -->|content_image| n12 n11 -->|style_image| n12 n12 -->|images| n13 n13 -->|video| n14 n12 -->|image| n15 n15 -->|images| n16 classDef custom fill:#12b886,stroke:#0b6b4f,color:#04110c; class n12 custom;

Data-flow DAG — scroll to zoom, drag to pan.

Nodes (17)#

1UNETLoadercore

unet_name = wan2.2_ti2v_5B_fp16.safetensorsweight_dtype = default

outputs: MODEL

2CLIPLoadercore

clip_name = umt5_xxl_fp8_e4m3fn_scaled.safetensorstype = wan

outputs: CLIP

3VAELoadercore

vae_name = wan2.2_vae.safetensors

outputs: VAE

turbo_loraLoraLoaderModelOnlycore

model = ◂ node 1 · out[0]lora_name = Wan22_TI2V_5B_Turbo_lora_rank_64_fp16.safetensorsstrength_model = 1.0

outputs: MODEL

4ModelSamplingSD3core

model = ◂ node turbo_lora · out[0]shift = {{shift}} tmpl

outputs: MODEL

5CLIPTextEncodecore

text = {{constructed_prompt}} tmplclip = ◂ node 2 · out[0]

outputs: CONDITIONING

6CLIPTextEncodecore

text =

overexposed, static motion, blurry details, subtitles, watermark, jpeg artifacts, deformed limbs, fused fingers, color shift, brightness flicker, scene cut

clip = ◂ node 2 · out[0]

outputs: CONDITIONING

7LoadImagecore

image = {{previous_tail_frame}} tmpl

outputs: IMAGEMASK

8Wan22ImageToVideoLatentcore

vae = ◂ node 3 · out[0]width = {{resolution_map[resolution].width}} tmplheight = {{resolution_map[resolution].height}} tmpllength = {{frames}} tmplbatch_size = 1start_image = ◂ node 7 · out[0]

outputs: LATENT

9KSamplercore

model = ◂ node 4 · out[0]positive = ◂ node 5 · out[0]negative = ◂ node 6 · out[0]latent_image = ◂ node 8 · out[0]seed = {{seed}} tmplsteps = {{4 if fast_mode else steps}} tmplcfg = {{1.0 if fast_mode else guidance}} tmplsampler_name = uni_pcscheduler = simpledenoise = 1.0

outputs: LATENT

10VAEDecodecore

samples = ◂ node 9 · out[0]vae = ◂ node 3 · out[0]

outputs: IMAGE

ref_frameLoadImagecore

image = {{reference_frame}} tmpl

outputs: IMAGEMASK

color_lock_nodeColorTransfercustom pack

content_image = ◂ node 10 · out[0]style_image = ◂ node ref_frame · out[0]mode = reinhard

outputs: IMAGE

11CreateVideocore

images = ◂ node color_lock_node · out[0]fps = {{fps}} tmpl

outputs: VIDEO

12SaveVideocore

video = ◂ node 11 · out[0]filename_prefix = wan_segmentformat = autocodec = auto

tail_pickImageFromBatchcore

image = ◂ node color_lock_node · out[0]batch_index = -1length = 1

outputs: IMAGE

13SaveImagecore

images = ◂ node tail_pick · out[0]filename_prefix = wan_segment_tail

Prompt construction#

template

{base_prompt}. Continuous take, seamless continuation of the previous shot, consistent lighting and color grade, smooth cinematic camera movement.

Variables (1)#

base_prompt

{{prompt}} tmpl

Parameter banks (1)#

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` (2)#

720p_landscape

{"width": 1280, "height": 704}

720p_portrait

{"width": 704, "height": 1280}

Models & 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.

ColorTransfer

Models required (4)#

Wan22_TI2V_5B_Turbo_lora_rank_64_fp16.safetensors

umt5_xxl_fp8_e4m3fn_scaled.safetensors

wan2.2_ti2v_5B_fp16.safetensors

wan2.2_vae.safetensors

Output contract#

What a successful run of this workflow returns.

type

video

format

mp4

description

Segment MP4 plus the tail-frame PNG used to condition the next chain link

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.

assetFamily

long-form-segment

outputPackageProfile

video-master-profile

controlModalities

keyframe-anchor-lockcolor-script-lockseed-lockmodel-locksampler-scheduler-locktemporal-lock

consistencyDimensions

motioncolor-scriptenvironmentlighting

notes

L4 chunk chaining: previous tail frame anchors the next Wan 2.2 segment; Reinhard color-lock against the canonical reference frame; emits its own tail frame for the next link. fast_mode = TI2V-5B Turbo distill (4 steps, no CFG).

Inputs (12)#

ComfyUI node graph (17)#

Nodes (17)#

Prompt construction#

Variables (1)#

Parameter banks (1)#

resolution_map (2)#

Models & dependencies#

Custom node packs (1)#

Models required (4)#

Output contract#

Taxonomy & routing#

`resolution_map` (2)#