Wan2.2 T2V A14B text-to-video on the RunPod video endpoint, full-precision fp16 experts. Two-expert schedule: the high-noise expert samples steps 0..k and hands its leftover noise to the low-noise expert for k..N, where k = round(steps x boundary_ratio). Graph shape and every sampler default are taken from ComfyUI 0.34.0's own video_wan2_2_14B_t2v template, not invented: euler/simple, shift 5.0, 20 steps split at 10, CFG 3.5, EmptyHunyuanLatentVideo, 16 fps. speed_mode=lightx2v_4step splices the distilled LoRA onto both experts and forces the template's lightning schedule, through the V.06.04 post-render transform.
16Graph nodes
12Node classes
16Inputs
8Models
Tags: motionwan2.2text-to-videorunpod-serverlessvolume-backed
Inputs (16)#
The typed parameter surface callers bind when they request this workflow. Enum options and numeric bounds are the values the workflow document declares.
promptstringrequireddefault a paper boat drifting down a rain gutter, water beading on its folds, late afternoon lightWhat the clip should show, in natural prose. Wan responds to descriptive scene language and to explicit statements of what moves.
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 default negative prompt (Chinese, as shipped). Replacing it with an English translation measurably weakens it — the model was trained with this text.
seedintegerdefault -1Random seed (-1 = random). Both experts share it.
resolutionenumdefault 480p_landscape_832x480Frame size. 720p costs roughly 2.4x a 480p clip of the same length.
options:
480p_landscape_832x480480p_portrait_480x832720p_landscape_1280x720720p_portrait_720x1280square_624x624lengthintegerdefault 81min 5max 121Frames, and it must be 4n+1 — Wan's temporal compression plus the first frame. At 16 fps the default 81 is about 5 seconds. The executor refuses anything else before billing (motion-caps.ts).
fpsintegerdefault 16min 8max 30Frame rate written into the container. Wan 2.2 was trained at 16 fps; a higher number here plays the same frames faster, it does not generate more.
stepsintegerdefault 20min 4max 40Total sampling steps across BOTH experts. Overridden by speed_mode=lightx2v_4step, which forces 4.
boundary_ratiofloatdefault 0.5min 0.1max 0.9Where the high-noise expert hands over to the low-noise one, as a fraction of steps. The template splits 20 at 10 and 4 at 2 — both 0.5.
cfgfloatdefault 3.5min 1.0max 10.0Classifier-free guidance, applied to both experts. Overridden by speed_mode=lightx2v_4step, which forces 1.0.
shiftfloatdefault 5.0min 1.0max 12.0ModelSamplingSD3 flow shift on both experts. 5.0 is the template value.
speed_modeenumdefault qualitylightx2v_4step splices the 4-step distilled LoRA onto BOTH experts and forces the template's lightning schedule (4 steps, split at 2, CFG 1.0), overriding steps/cfg/boundary_ratio. A LoRA is not a quantization, so this stays inside the full-precision rule (D-V5).
options:
qualitylightx2v_4stepsamplerenumdefault eulerSampler for both experts
options:
euleruni_pcdpmpp_2mschedulerenumdefault simpleSigma schedule for both experts
options:
simplebetamotion_presetenumdefault noneA documented phrase appended to the prompt (motion_map); `none` appends nothing at all. This is prompt guidance and nothing more — there is no camera control in this graph, and the model may ignore it.
options:
nonestatic cameraslow push inorbit lefthandhelddolly outloraenumdefault noneOptional Wan 2.2 LoRA pair (A.03.03). Wan 2.2 A14B is two experts, so every entry is two Civitai files: its high-noise file loads on the high-noise expert and its low-noise file on the low-noise expert, both at lora_strength, and an entry missing either file is refused at render rather than loaded by halves. Only pairs whose Civitai versions name Wan Video 2.2 T2V-A14B as their base are offered here, so the I2V-trained motion LoRAs are on wan22-i2v only. breast-slider-v2 = [Wan2.2] Breast slider v2.0, driven by strength in either direction; its versions list no trained words. It rendered one live clip on 2026-09-14 (A.05.02; see proof_note). none adds no loader.
options:
nonebreast-slider-v2lora_strengthfloatdefault 1min -2max 2LoRA strength on BOTH experts (LoraLoaderModelOnly.strength_model on the high-noise and the low-noise half alike). A slider LoRA is meant to be driven negative as well as positive; the motion LoRAs sit at 1.0.
ComfyUI node graph (16)#
The executable ComfyUI prompt graph: 16 nodes across 12 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: UNETLoader"]
n2["3: CLIPLoader"]
n3["4: VAELoader"]
n4["5: ModelSamplingSD3"]
n5["6: ModelSamplingSD3"]
n6["7: CLIPTextEncode"]
n7["8: CLIPTextEncode"]
n8["9: EmptyHunyuanLatentVideo"]
n9["10: KSamplerAdvanced"]
n10["11: KSamplerAdvanced"]
n11["12: VAEDecode"]
n12["13: CreateVideo"]
n13["14: SaveVideo"]
n14["15: ImageFromBatch"]
n15["16: SaveImage"]
n0 -->|model| n4
n1 -->|model| n5
n2 -->|clip| n6
n2 -->|clip| n7
n4 -->|model| n9
n6 -->|positive| n9
n7 -->|negative| n9
n8 -->|latent_image| n9
n5 -->|model| n10
n6 -->|positive| n10
n7 -->|negative| n10
n9 -->|latent_image| n10
n10 -->|samples| n11
n3 -->|vae| n11
n11 -->|images| n12
n12 -->|video| n13
n11 -->|image| n14
n14 -->|images| n15
Nodes (16)#
1UNETLoadercoreunet_name = wan2.2_t2v_high_noise_14B_fp16.safetensorsweight_dtype = defaultoutputs:
MODEL2UNETLoadercoreunet_name = wan2.2_t2v_low_noise_14B_fp16.safetensorsweight_dtype = defaultoutputs:
MODEL3CLIPLoadercoreclip_name = umt5_xxl_fp16.safetensorstype = wandevice = defaultoutputs:
CLIP4VAELoadercorevae_name = wan_2.1_vae.safetensorsoutputs:
VAE5ModelSamplingSD3coremodel = ◂ node 1 · out[0]shift = {{shift}} tmploutputs:
MODEL6ModelSamplingSD3coremodel = ◂ node 2 · out[0]shift = {{shift}} tmploutputs:
MODEL7CLIPTextEncodecoretext = {{constructed_prompt}} tmplclip = ◂ node 3 · out[0]outputs:
CONDITIONING8CLIPTextEncodecoretext = {{negative_prompt}} tmplclip = ◂ node 3 · out[0]outputs:
CONDITIONING9EmptyHunyuanLatentVideocorewidth = {{resolution_map[resolution].width}} tmplheight = {{resolution_map[resolution].height}} tmpllength = {{length}} tmplbatch_size = 1outputs:
LATENT10KSamplerAdvancedcoremodel = ◂ node 5 · out[0]positive = ◂ node 7 · out[0]negative = ◂ node 8 · out[0]latent_image = ◂ node 9 · out[0]add_noise = enablenoise_seed = {{seed}} tmplsteps = {{steps}} tmplcfg = {{cfg}} tmplsampler_name = {{sampler}} tmplscheduler = {{scheduler}} tmplstart_at_step = 0end_at_step = {{int(round(steps * boundary_ratio))}} tmplreturn_with_leftover_noise = enableoutputs:
LATENT11KSamplerAdvancedcoremodel = ◂ node 6 · out[0]positive = ◂ node 7 · out[0]negative = ◂ node 8 · out[0]latent_image = ◂ node 10 · out[0]add_noise = disablenoise_seed = {{seed}} tmplsteps = {{steps}} tmplcfg = {{cfg}} tmplsampler_name = {{sampler}} tmplscheduler = {{scheduler}} tmplstart_at_step = {{int(round(steps * boundary_ratio))}} tmplend_at_step = {{steps}} tmplreturn_with_leftover_noise = disableoutputs:
LATENT12VAEDecodecoresamples = ◂ node 11 · out[0]vae = ◂ node 4 · out[0]outputs:
IMAGE13CreateVideocoreimages = ◂ node 12 · out[0]fps = {{fps}} tmploutputs:
VIDEO14SaveVideocorevideo = ◂ node 13 · out[0]filename_prefix = wan22-t2vformat = mp4codec = h26415ImageFromBatchcoreimage = ◂ node 12 · out[0]batch_index = {{length - 1}} tmpllength = 1outputs:
IMAGE16SaveImagecoreimages = ◂ node 15 · out[0]filename_prefix = wan22-t2v-lastframePrompt construction#
template{base_prompt}{motion}Variables (2)#
base_prompt{{prompt}} tmplmotion{{motion_map[motion_preset]}} tmplParameter banks (6)#
The prompt / configuration lookup tables this workflow keys into from its inputs — the vocabulary that turns a style / palette / preset selection into graph parameters.
option_policy (1)#
lora{"breast-slider-v2": {"distribution": "service", "content": "racy"}}resolution_map (5)#
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}square_624x624{"width": 624, "height": 624}requires_families (2)#
wan22-t2vwan-sharedmotion_map (6)#
nonestatic camera, static cameraslow push in, slow push inorbit left, camera orbits lefthandheld, handheld cameradolly out, camera dollies outlora_pair_map (1)#
breast-slider-v2{"high": "wan22-breast-slider-v2-high.safetensors", "low": "wan22-breast-slider-v2-low.safetensors"}post_render (2)#
wan_lora_pair{"input": "lora", "strength_input": "lora_strength", "pair_map": "lora_pair_map", "none_value": "none"}wan_speed_mode{"input": "speed_mode", "fast_value": "lightx2v_4step", "lora_high": "wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors", "lora_low": "wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise.safetensors", "steps": 4, "split": 2, "cfg": 1.0, "strength": 1.0}Models & dependencies#
Models required (8)#
wan2.2_t2v_high_noise_14B_fp16.safetensorswan2.2_t2v_low_noise_14B_fp16.safetensorsumt5_xxl_fp16.safetensorswan_2.1_vae.safetensorswan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensorswan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise.safetensorswan22-breast-slider-v2-high.safetensorswan22-breast-slider-v2-low.safetensorsOutput contract#
What a successful run of this workflow returns.
typevideoformatmp4codech264fps_sourcefixedaudiofalsealphafalseexpected_dimensions{"width": "{{resolution_map[resolution].width}}", "height": "{{resolution_map[resolution].height}}", "frames": "{{length}}"}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.
assetFamilykeyframe-anchored-motionoutputPackageProfilevideo-master-profilecontrolModalitiesmodel-locksampler-scheduler-lockseed-lockprompt-template-lockmotion-vector-locktemporal-lockconsistencyDimensionsmotionidentitylightingnotesWan2.2 T2V A14B, full-precision two-expert schedule on the RunPod video endpoint. controlModalities lists only what the graph really exposes: there is no camera control here, and motion_preset is prompt text, not a camera path.