Unlock Liquid Magic: Advanced I2V Workflow for Stunning Visual Effects
1. Workflow Overview

This workflow is designed for Image-to-Video (I2V) generation, specializing in simulating liquid collision and splash effects (e.g., purple TV with transparent liquid). It combines WanVideo's custom models with GIMM-VFI frame interpolation, outputting MP4 videos or GIFs.
2. Core Models
Wan2.1-I2V-14B:
Function: Generates video latent from images (720P support).
Source: Download
Wan2_1-I2V-14B-720P_fp8_e4m3fn.safetensors
(BF16 precision).
UMT5-XXL Text Encoder:
Encodes Chinese prompts (e.g., "liquid splash") into embeddings.
Source:
umt5-xxl-enc-bf16.safetensors
.
GIMM-VFI Model:
Enhances video smoothness via frame interpolation. Requires
gimmvfi_r_arb_lpips_fp32.safetensors
.
3. Key Nodes
Node Name | Function | Installation |
---|---|---|
WanVideoBlockSwap | Optimizes VRAM via block swapping | Install |
WanVideoLoraSelect | Loads LoRA (e.g., liquid effects) | Same as above |
GIMMVFI_interpolate | Frame interpolation for videos | Requires |
VHS_VideoCombine | Renders frames to MP4/GIF | Install |
4. Workflow Groups
Group 1: Model & Encoder Loading
Input: Model files, LoRA, text encoder.
Output: Initialized models.
Group 2: Text & Image Encoding
Input: Chinese prompts + image (1024x1440 PNG).
Output: Text/image embeddings.
Group 3: Video Generation & Interpolation
Input: Embeddings + seed (e.g.,
1032809127571661
).Output: Interpolated frames (272x272).
5. Inputs & Outputs
Required Inputs:
Image: Clear subject (e.g., TV) with simple background.
Prompts: Positive (e.g., "transparent liquid splash"), negative (e.g., "color change").
Seed: Fixed (e.g.,
1032809127571661
) or randomized.
Output:
Default: MP4 (NVENC H.264, 30FPS), optional GIF.
6. Notes
Hardware:
≥12GB VRAM (RTX 3090+ recommended) due to UMT5-XXL and Wan2.1 model size.
Troubleshooting:
VRAM errors: Reduce resolution or disable
teacache_args
optimization.Missing nodes: Install
ComfyUI-WanVideoWrapper
andComfyUI-GIMM-VFI
.
LoRA:
Download
WAN2.1 ZOEY Liquid Splash_Alpha.safetensors
, set weight=1.0.