Wan T2v Video

WAN 2.2 Text-to-Video (T2V) Workflows

Overview

WAN 2.2 T2V generates videos from text prompts using a 14B parameter MoE (Mixture of Experts) architecture split across two specialized models:

• HighNoise model: Handles early denoising — establishes structure, motion, composition
• LowNoise model: Handles late denoising — refines details, sharpens output

This dual-model technique is the same as FLF/I2V (see wan-flf-video skill) but without image conditioning nodes.

Key difference from I2V/FLF: T2V does NOT use CLIPVisionEncode, WanFirstLastFrameToVideo, or any image input. It uses EmptyHunyuanLatentVideo for latent initialization and text-only conditioning.

Models

UNET (Installed)

Text Encoder

VAE

VACE Modules (Installed — For Advanced Control)

VACE modules add reference image / pose / depth conditioning to T2V. See WanVideoWrapper section below.

Lightning LoRAs (Installed)

T2V Lightning v1.1 (Paired Hi/Lo)

T2V Lightning Seko V2.0 (Alternative Paired)

T2V CFG-Step Distill (Higher Quality)

Sampler Settings

Lightning (4-Step, Recommended for Speed)

Standard (20-Step, Full Quality)

ModelSamplingSD3

Required for WAN 2.2 flow matching. Apply to BOTH models:

{
  "class_type": "ModelSamplingSD3",
  "inputs": { "model": ["<unet>", 0], "shift": 8 }
}

T2V shift values:

• Standard: shift=8 (good balance of motion and detail)
• Lightning: shift=5 (lower shift for distilled models)
• Range 6–9: Higher shift = more detail, lower shift = stronger motion

EmptyHunyuanLatentVideo

Creates the initial video latent for T2V (no image input):

{
  "class_type": "EmptyHunyuanLatentVideo",
  "inputs": {
    "width": 832,
    "height": 480,
    "length": 81,
    "batch_size": 1
  }
}

This replaces WanFirstLastFrameToVideo (which is for FLF/I2V only). The latent goes directly to KSamplerAdvanced Pass 1.

Negative Prompt

The tones are vibrant, overexposed, static, details are unclear, subtitles, style, work, painting, image, still, overall grayish, worst quality, low quality, JPEG compression artifacts, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, distorted limbs, merged fingers, motionless image, cluttered background, three legs, many people in the background, walking backwards

Pipeline Flow

UNETLoader (HIGH T2V) → ModelSamplingSD3 (shift) → LoraLoaderModelOnly (Hi Lightning) → MODEL_HI
UNETLoader (LOW T2V) → ModelSamplingSD3 (shift) → LoraLoaderModelOnly (Lo Lightning) → MODEL_LO
CLIPLoader (wan) → CLIP
  ├─ CLIPTextEncode (positive) → CONDITIONING
  └─ CLIPTextEncode (negative) → CONDITIONING
VAELoader → VAE

EmptyHunyuanLatentVideo (832x480, 81 frames) → LATENT

KSamplerAdvanced (Hi: MODEL_HI, steps 0-2, add_noise=enable, return_leftover=enable)
  → noisy LATENT
KSamplerAdvanced (Lo: MODEL_LO, steps 2-4, add_noise=disable, return_leftover=disable)
  → final LATENT

VAEDecode → IMAGE → VHS_VideoCombine → MP4

Complete Workflow: T2V Lightning (4-Step)

{
  "1": { "class_type": "UNETLoader", "inputs": { "unet_name": "Wan2_2-T2V-A14B_HIGH_fp8_e4m3fn_scaled_KJ.safetensors", "weight_dtype": "default" }, "_meta": { "title": "UNET HighNoise T2V" }},
  "2": { "class_type": "UNETLoader", "inputs": { "unet_name": "Wan2_2-T2V-A14B-LOW_fp8_e4m3fn_scaled_KJ.safetensors", "weight_dtype": "default" }, "_meta": { "title": "UNET LowNoise T2V" }},
  "3": { "class_type": "ModelSamplingSD3", "inputs": { "model": ["1", 0], "shift": 5 }, "_meta": { "title": "Hi Shift" }},
  "4": { "class_type": "ModelSamplingSD3", "inputs": { "model": ["2", 0], "shift": 5 }, "_meta": { "title": "Lo Shift" }},
  "5": { "class_type": "LoraLoaderModelOnly", "inputs": {
    "model": ["3", 0],
    "lora_name": "Unknown\\no tags\\wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors",
    "strength_model": 1.0
  }, "_meta": { "title": "Hi Lightning" }},
  "6": { "class_type": "LoraLoaderModelOnly", "inputs": {
    "model": ["4", 0],
    "lora_name": "Unknown\\no tags\\wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise.safetensors",
    "strength_model": 1.0
  }, "_meta": { "title": "Lo Lightning" }},
  "7": { "class_type": "CLIPLoader", "inputs": { "clip_name": "umt5_xxl_fp8_e4m3fn_scaled.safetensors", "type": "wan" }},
  "8": { "class_type": "VAELoader", "inputs": { "vae_name": "wan_2.1_vae.safetensors" }},
  "9": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["7", 0], "text": "<positive prompt describing the video scene and motion>" }, "_meta": { "title": "Positive" }},
  "10": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["7", 0], "text": "The tones are vibrant, overexposed, static, details are unclear, subtitles, worst quality, low quality, motionless image" }, "_meta": { "title": "Negative" }},
  "11": { "class_type": "EmptyHunyuanLatentVideo", "inputs": {
    "width": 832, "height": 480, "length": 81, "batch_size": 1
  }},
  "12": { "class_type": "KSamplerAdvanced", "inputs": {
    "model": ["5", 0],
    "positive": ["9", 0],
    "negative": ["10", 0],
    "latent_image": ["11", 0],
    "add_noise": "enable", "noise_seed": 0, "steps": 4, "cfg": 1,
    "sampler_name": "euler", "scheduler": "simple",
    "start_at_step": 0, "end_at_step": 2, "return_with_leftover_noise": "enable"
  }, "_meta": { "title": "Hi Pass" }},
  "13": { "class_type": "KSamplerAdvanced", "inputs": {
    "model": ["6", 0],
    "positive": ["9", 0],
    "negative": ["10", 0],
    "latent_image": ["12", 0],
    "add_noise": "disable", "noise_seed": 0, "steps": 4, "cfg": 1,
    "sampler_name": "euler", "scheduler": "simple",
    "start_at_step": 2, "end_at_step": 4, "return_with_leftover_noise": "disable"
  }, "_meta": { "title": "Lo Pass" }},
  "14": { "class_type": "VAEDecode", "inputs": { "samples": ["13", 0], "vae": ["8", 0] }},
  "15": { "class_type": "VHS_VideoCombine", "inputs": {
    "images": ["14", 0], "frame_rate": 16, "loop_count": 0,
    "filename_prefix": "wan_t2v", "format": "video/h264-mp4",
    "pingpong": false, "save_output": true,
    "pix_fmt": "yuv420p", "crf": 19, "save_metadata": true, "trim_to_audio": false
  }}
}

Complete Workflow: T2V Standard (20-Step)

Same structure as above but replace the LoRA and sampler settings:

• Remove LoRA nodes (5 and 6) — connect ModelSamplingSD3 outputs directly to KSamplerAdvanced
• Change shift to 8 in ModelSamplingSD3 nodes
• Change KSamplerAdvanced settings:
  • steps: 20, cfg: 3.5
  • Pass 1: start_at_step: 0, end_at_step: 10
  • Pass 2: start_at_step: 10, end_at_step: 20

WanVideoWrapper Approach (Advanced)

For more control, use the WanVideoWrapper custom node pack. Key differences from native:

• Uses WanVideoModelLoader → WANVIDEOMODEL type
• Uses WanVideoSampler with built-in shift parameter
• Supports TeaCache, context windows, block swap for VRAM management
• LoRAs load via WanVideoLoraSelect → WanVideoModelLoader lora input

⚠️ CRITICAL: merge_loras=false with fp8-scaled models

When a LoRA (e.g. the lightx2v 4-step lightning hi/lo LoRAs) is loaded onto an fp8-quantized model (quantization=fp8_e4m3fn_scaled) via WanVideoLoraSelect, set the node's merge_loras widget to false. The default merge_loras=true tries to bake the LoRA into the already-quantized fp8 weights and hard-crashes ComfyUI during LoRA loading with no Python traceback (process dies — looks like an unexplained restart/OOM). merge_loras=false applies the LoRA as a runtime patch instead, which is fp8-safe. Use merge_loras=true only on non-quantized bf16/fp16 models. Pairs cleanly with WanVideoBlockSwap for fp8 14B on 24GB cards.

WanVideoWrapper T2V Pipeline

WanVideoModelLoader (T2V model) → WANVIDEOMODEL
WanVideoVAELoader → WANVAE
WanVideoTextEncode (positive + negative prompts) → WANVIDEOTEXTEMBEDS
WanVideoImageToVideoEncode (no images — creates empty embeds for T2V)
  → WANVIDIMAGE_EMBEDS

WanVideoSampler (model, image_embeds, text_embeds, steps, cfg, shift, scheduler)
  → LATENT

WanVideoDecode → IMAGE → VHS_VideoCombine → MP4

WanVideoSampler T2V Settings

Concept LoRAs (Installed)

Located in loras/Wan Video 2.2 T2V-A14B/:

• concept/PussyLoRA_HighNoise_Wan2.2_HearmemanAI.safetensors + LowNoise pair

Apply concept LoRAs the same way as lightning LoRAs — match hi/lo to the correct model pass. Use LoraLoaderModelOnly with strength 0.5–1.0.

Resolution & Frame Count

Resolutions

Width and height must be divisible by 16.

Frame Count (4n + 1)

• 81 frames at 16fps = ~5 seconds (default, recommended)
• 49 frames at 16fps = ~3 seconds (faster)
• 121 frames at 16fps = ~7.5 seconds (longer, more VRAM)

Frame Rate

Standard: 16 fps for WAN 2.2 output.

VRAM Considerations

• Always clear_vram before switching to WAN T2V from another model family
• Lightning (4 steps) dramatically reduces generation time: ~70s vs ~5-10 min for 20 steps
• Only one UNET is active during each pass — they swap in/out

Prompt Tips

Describe motion and temporal progression, not just a scene:

Good: "A beautiful young woman slowly walks through a blooming cherry blossom garden, petals drifting in the breeze, soft sunlight filtering through branches, cinematic slow motion, 4K quality"
Bad: "woman in garden"

Include motion cues: "slowly walks", "camera pans", "wind blowing", "gradually reveals"

T2V vs I2V/FLF Comparison