WAN 2.1 vs 2.2: ComfyUI T2V, I2V, CausVid, VACE Guide

WAN 2.1 and 2.2 are the two open-weight video generation families that run locally in ComfyUI without commercial APIs. Both come from Alibaba under Apache 2.0, but their architectures diverge enough that the same prompt, LoRA, or sampler can produce opposite results. This guide maps every workflow variant — native text-to-video, image-to-video, first-last-frame, CausVid distillation, and VACE editing — to the hardware you have and the clip you need.

TL;DR

• Start with WAN 2.1 T2V 1.3B if you have ≤8 GB VRAM; it runs at 8 GB and needs only one diffusion model file.
• Switch to WAN 2.2 T2V/I2V 14B when you have enough VRAM for the MoE two-expert pipeline (high-noise + low-noise) for sharper motion.
• Use CausVid LoRA at CFG 1.0 only on the first step, then drop it for normal sampling — otherwise motion collapses.
• Pick VACE for any edit that needs a reference video plus mask (inpainting, subject swap, motion transfer); native I2V only conditions on a single start frame.

Architecture Difference: Single Model vs MoE Experts

WAN 2.1 packages each task as one diffusion model: wan2.1_t2v_1.3B_fp16.safetensors for text-to-video, wan2.1_i2v_480p_14B_fp16.safetensors or wan2.1_i2v_720p_14B_fp16.safetensors for image-to-video. The text encoder (umt5_xxl_fp8_e4m3fn_scaled.safetensors), VAE (wan_2.1_vae.safetensors), and CLIP Vision (clip_vision_h.safetensors) are shared across tasks.

WAN 2.2 introduces a Mixture-of-Experts split. For 14B T2V and I2V you must load two diffusion models — a high-noise expert for early timesteps and a low-noise expert for late timesteps:

The 5B TI2V (text+image-to-video) remains a single model. S2V (speech-to-video) uses wan2.2_s2v_14B_fp8_scaled.safetensors plus an audio encoder wav2vec2_large_english_fp16.safetensors.

Native T2V and I2V Graphs in ComfyUI

ComfyUI ships native WanVideo nodes — no wrapper required. A minimal T2V graph chains:

1. Load Diffusion Model → 2.1: wan2.1_t2v_1.3B_fp16.safetensors / 2.2: both high/low noise models
2. Load CLIP → umt5_xxl_fp8_e4m3fn_scaled.safetensors
3. Load VAE → wan_2.1_vae.safetensors (2.1) or wan_2.2_vae.safetensors (2.2 5B) / wan_2.1_vae.safetensors (2.2 14B)
4. EmptyHunyuanLatentVideo → set frames, width, height
5. WanVideoSampler (2.1) or dual-expert sampler (2.2)
6. VAE Decode → Save Video

For I2V, insert Load CLIP Vision → clip_vision_h.safetensors and Load Image before the sampler; the WanImageToVideo node handles conditioning.

WAN 2.2 Two-Stage Sampling: High-Noise + Low-Noise

Because 2.2 splits experts by timestep, the sampler must hand off between them. Community workflows use a high-noise stage followed by a low-noise stage:

Community testing confirms WAN 2.1 LoRAs only work in the low-noise stage — the high-noise expert's weight distribution diverges from the single-model trajectory the LoRAs were trained on. A Civitai workflow author notes: "Only Wan 2.2 High Noise LoRAs go here. Do not use any Wan 2.1 LoRA in the High Noise section" while "Use any Wan 2.1 LoRAs here. Also any Wan 2.2 Low Noise LoRAs."

CausVid: Step-Distilled, Not a Drop-In Replacement

CausVid LoRAs (Wan2.1-T2V-14B-CausVid, Wan2.1-T2V-14B-StepDistill-CfgDistill) compress multi-step diffusion into fewer steps. They are CFG-distilled — trained at CFG 1.0 — so running them at CFG > 1.0 double-applies guidance and creates artifacts. The Kijai repo lists them as "CFG and Step distill 14B" and ComfyUI's S2V guide confirms 4-step Lightning LoRA at CFG 1.0.

The working pattern from Reddit experimentation:

1. Enable CausVid LoRA at strength 1.0 for step 1 only (or first N < total steps)
2. Disable it for remaining steps via a "CausVid Remaining Steps Only" node or equivalent logic
3. Sample normally with your preferred sampler

Leaving CausVid active for all steps weakens motion; the first-step-only trick preserves dynamics while keeping most of the speed gain.

VACE: Unified Video Editing, Not Generation-from-Scratch

VACE models (Wan2.1-VACE-1.3B, Wan2.1-VACE-14B, and 2.2 Fun VACE high/low noise pairs) accept three inputs simultaneously:

• Reference video
• Mask / control signal (depth, pose, canny, custom)
• Text prompt

This enables inpainting, outpainting, subject replacement, motion transfer, and style transfer in one graph. Native I2V only conditions on the first frame; native T2V only on text. VACE is the only WAN path for "change this object in frames 10–30 while keeping the rest."

ComfyUI workflows for VACE are still emerging — model files are identified but full node graphs are not yet documented in official guides.

Decision Matrix: Which Workflow for Your Clip

Common Errors and Fixes

FAQ

Why does WAN 2.2 require two diffusion models for T2V and I2V? It uses a Mixture-of-Experts architecture that splits denoising into a high-noise expert (early timesteps) and a low-noise expert (late timesteps). The 5B TI2V model is the exception — it remains a single model.

Can I use my WAN 2.1 LoRAs in WAN 2.2? Yes, but only in the low-noise stage. The high-noise expert's weight distribution diverges from the single-model trajectory the 2.1 LoRAs were trained on.

What CFG should I use with Lightx2v / CausVid LoRAs? CFG = 1.0 exactly. These LoRAs are CFG-distilled; higher values double-apply guidance and create artifacts.

How do I get better motion with CausVid? Apply the CausVid LoRA only on the first sampling step (or first N < total steps), then disable it for the remaining steps. This preserves the motion dynamics that full CausVid distillation weakens.

What is the difference between I2V and FLF2V in WAN 2.2? I2V conditions on a single start frame. FLF2V (WanFirstLastFrameToVideo node) conditions on both first and last frames, giving temporal boundary control.

Does VACE replace I2V for image-to-video? No. VACE is for editing an existing video with masks and control signals. I2V generates fresh video from one reference image. Use VACE when you have a source video to modify.

Why do Chinese prompts work better for WAN? The training data skew toward Chinese captions is documented by users but not explained in primary sources. Mechanism unknown.

Sources

1. ComfyUI WAN 2.1 Tutorial — https://docs.comfy.org/tutorials/video/wan/wan-video
2. ComfyUI WAN 2.2 Tutorial — https://docs.comfy.org/tutorials/video/wan/wan2_2
3. ComfyUI WAN 2.2 S2V Tutorial — https://docs.comfy.org/tutorials/video/wan/wan2-2-s2v
4. Comfy-Org WAN 2.1 Repo — https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged
5. Comfy-Org WAN 2.2 Repo — https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged
6. Kijai WanVideo_comfy Repo — https://huggingface.co/Kijai/WanVideo_comfy
7. Reddit: CausVid motion improvement — https://www.reddit.com/r/StableDiffusion/comments/1ksxy6m/causvid_wan_img2vid_improved_motion_with_two/
8. Civitai: WAN 2.2 workflow favorites — https://civitai.com/articles/23316

Internal Links

• WAN 2.1 ComfyUI install guide
• Optimizing VRAM for video diffusion models
• ComfyUI custom node manager walkthrough
• LoRA training basics for video models

Image Placeholders

• ``
• ``