Directions

Here’s version 2 of my V2V WAN 2.2 + VACE workflow. Here’s how it works at a high level

1. Upload subject reference (Node Name: ⭐️ Load Image of Reference Subject)

2. Upload video reference (Node Name: ⭐️ Media Selection (Reference Video)

3. Enter Prompt

   1. Enter Manually in WanVideo TextEncode (Default)

   2. Use SwissArmyKnife LLM nodes

      1. Requires additional setup if using Qwen3-VL via LLM Studio (locally hosted model)

      2. Requires API key if using Gemini API

4. Run Workflow

Notes

Subject/Character

• Use a high quality subject reference image

  • Closeups work best from my testing

  • The background of your subject reference will influence your video output slightly

    • I haven’t figured exactly how to properly mask just the subject in a way that works with WAN VACE

  • Your subject’s identity will not be preserved perfectly due to the nature of VACE and other variables like seed

    • Best best is to use a subject/character lora is you need it to be consistent

• You can turn down the fun reward lora strength if your generated video is too “shiny”

Lora Additions

• You can add more loras to fine tune the generated video but try not to add too many because you end up having loras fighting each other and you will get a burned out looking generation

Prompting

1. Subject: Describe your main subject with clarity — who or what it is, what they’re doing, and how they appear.

2. Clothing: Focus on what the subject is wearing or how the outfit contributes to mood, texture, colour or story. Consider description of fabrics, accessories, era, and fit.

3. Movement: Elaborate on how the subject moves, how the camera moves, or any dynamic elements in the scene. Use cinematic language when helpful.

4. Scene: Define the environment: time of day, location, background/foreground elements, mood, composition and lighting.

5. Visual Style: Establish the look and feel: lighting, colour-grading, lens effects, film stock, level of realism vs stylised, any elements you don’t want (negative prompt awareness).

I have added prompt examples to the markdown file in the zip

NSFW

This works with NSFW assuming you have a NSFW lora + good prompt. I haven’t found a great high quality nsfw lora so that’s why it’s not included

Major changes from V1

• Added image upload for subject reference

• Fixed node mismatches with SwissArmyKnife custom nodes

• Switched from Gemini to Qwen3 VL (running locally and exposes via Swiss Army Knife nodes)

• Added a path to input prompt instead of relying on SwissArmyKnife LLM nodes

• Overhauled & Simplified VACE Encoding nodes, now it just uses the subject reference and depth map

Roadmap

• Figure how to mask character so subject’s identity is preserved better and the background of the reference image doesn’t influence the generated video too much

• Need a better solution for upscaling and interpolation

• Explore VACE’s First Frame Last Frame capabilities to generate longer videos

• Dial in settings for NSFW loras

Model links

You can find all the models on Huggingface. I am running a Nvidia 3090TI w 24GB VRAM & 128 GB DDR4 RAM. The FP8_e5m2 work best for the 3000 series generation. Generations take about 300-500 seconds on my system

Diffusion model

• WAN 2.2 T2V

  • High Noise

    • Wan2_2-T2V-A14B-HIGH_fp8_e5m2_scaled_KJ.safetensors

    • Wan2_2_Fun_VACE_module_A14B_HIGH_fp8_e5m2_scaled_KJ.safetensors

  • Low Noise

    • Wan2_2-T2V-A14B-LOW_fp8_e5m2_scaled_KJ.safetensors

    • Wan2_2_Fun_VACE_module_A14B_LOW_fp8_e5m2_scaled_KJ.safetensors

Text encoder

• umt5_xxl_fp16.safetensors

VAE

• Wan2.1_VAE.safetensors

LoRAs

• High Noise

  • Wan22_A14B_T2V_LOW_Lightning_4steps_lora_250928_rank64_fp16.safetensors

  • Wan2.2-Fun-A14B-InP-HIGH-MPS_resized_dynamic_avg_rank_21_bf16.safetensors

  • Instagirlv2.5-HIGH.safetensors

• Low Noise

  • Wan22_A14B_T2V_LOW_Lightning_4steps_lora_250928_rank64_fp16.safetensors

  • Wan2.2-Fun-A14B-InP-LOW-MPS_resized_dynamic_avg_rank_22_bf16.safetensors

  • Instagirlv2.5-LOW.safetensors

Model Storage Location

📂 ComfyUI/
├── 📂 models/
│   ├── 📂 diffusion_models/
│   │   ├── Wan2_2-T2V-A14B-HIGH_fp8_e5m2_scaled_KJ.safetensors
│   │   ├── Wan2_2-T2V-A14B-LOW_fp8_e5m2_scaled_KJ.safetensors
│   │   ├── Wan2_2_Fun_VACE_module_A14B_HIGH_fp8_e5m2_scaled_KJ.safetensors
│   │   └── Wan2_2_Fun_VACE_module_A14B_LOW_fp8_e5m2_scaled_KJ.safetensors
│   ├── 📂 vae/
│   │   └── Wan2.1_VAE.safetensors
│   ├── 📂 text_encoders/
│   │   └── umt5_xxl_fp16.safetensors
│   └── 📂 loras/
│       ├── Wan22_A14B_T2V_LOW_Lightning_4steps_lora_250928_rank64_fp16.safetensors
│       ├── Wan2.2-Fun-A14B-InP-HIGH-MPS_resized_dynamic_avg_rank_21_bf16.safetensors
│       ├── Wan2.2-Fun-A14B-InP-LOW-MPS_resized_dynamic_avg_rank_22_bf16.safetensors
│       ├── Instagirlv2.5-HIGH.safetensors
│       └── Instagirlv2.5-LOW.safetensors

Custom Nodes

ComfyUI-WanVideoWrapper - nightly

comfyui_controlnet_aux - v1.1.2

ComfyUI-Easy-Use - v1.3.4

ComfyUI-KJNodes - v1.1.7

ComfyUI-VideoHelperSuite - v1.7.7

ComfyUI-Frame-Interpolation - v.1.0.7

ComfyUI Video Depth Anything - nightly

CRT-Nodes - v1.8.2

Swiss Army Knife - v2.9.1

ComfyUI

ComfyUI - v0.3.65

ComfyUI_frontend - v1.27.10

Python - v3.12.3

Pytorch - 2.9.0+cu128