javawock7618/comfy-wan2.2-workflows

Wan2.2 GGUF & INT8 Workflows Collection (with lightx2v LoRA)

This repository provides a collection of highly optimized ComfyUI workflows tailored for the Wan-AI/Wan2.2 model family using GGUF and INT8 quantized variants. Built completely around ComfyUI's native core nodes and quantized loaders, these workflows leverage the lightx2v LoRA for accelerated high-speed generation, combined with advanced features such as specific LoRA combinations, motion reference transfer (v2v), and localized detail upscaling while keeping VRAM usage exceptionally low.

---

Workflow Files & Overview

The repository contains the following production-ready and experimental workflow files.
(Note: YYMM.x in the file names represents the release year, month, and version number, e.g., 2606.1 or 2607.1)

#	File Name	Workflow Target	Key Features & Quick Notes
1	Wan2.2_SVI-javanoYYMM.x.json	Image-to-Video (SVI Long Gen)	Long-form video generation using Wan2.2 GGUF sped up by lightx2v LoRA, and stacked with SVI2 Pro LoRA for ultra high-fidelity i2v outputs.
2	Wan2.2_int8_SVI-javanoYYMM.x.json	Image-to-Video (SVI INT8 / Long Gen)	Highly optimized INT8 variant for SVI long video workflows. Delivers a significant speed boost over GGUF versions when running on Ampere-generation GPUs and newer.
3	Wan2.2_Anime-javanoYYMM.x.json	Anime / Motion Transfer v2v	Video-to-Video workflow running on GGUF with lightx2v LoRA acceleration that extracts and transfers motion data from reference videos.
4	Wan2.x_Inpaint-DetailerYYMM.x.json	Localized Inpaint & Detailer	An accelerated GGUF-driven v2v utility designed for partial correction and targeted regional enhancement beyond just the face.

---

Detailed Workflow Breakdown

1 & 2. Image-to-Video / SVI Long Gen (SVI / SVI INT8)

Designed for robust, accelerated long-form video production utilizing Wan2.2's native quantized Image-to-Video (i2v) architecture.

• INT8 Computational Acceleration: The Wan2.2_int8_SVI-javanoYYMM.x.json workflow introduces native INT8 support. For users running Ampere architecture cards (e.g., RTX 30-series, A100) or newer, this option minimizes processing latency and enhances inference throughput compared to standard GGUF pipelines.
• Dual LoRA Stacking (lightx2v + SVI2 Pro): Both SVI configurations are pre-tuned to stack the lightx2v LoRA (for fast, few-step sampling) together with the SVI2 Pro LoRA (for structural stability and high-fidelity texturing). This provides a perfect balance between rapid processing and cinematic realism.
• Extended Sampling Configuration: Features customized temporal scheduling nodes that prevent visual degradation or color bleeding when generating videos beyond standard lengths.

3. Anime / Motion Reference (Anime)

Specialized in traditional Video-to-Video (v2v) conversion utilizing fast GGUF inference and lightx2v acceleration.

• Motion Transfer Engine: Extracts structural motion vectors from a source reference video and projects them onto your new text prompts or initial subject image setup using native core nodes.
• High-Speed Stylization Continuity: Leverages the few-step sampling capability of lightx2v to transform live-action motion into precise anime configurations quickly, while maintaining fluid character tracking across the entire timeline.

4. Localized Inpaint & Detailer (Inpaint-Detailer)

A versatile, fast-rendering Video-to-Video utility built to break the limitations of standard facial-only upscalers.

• Universal Regional Upscaling: Easily isolate and upscale specific, arbitrary regions inside the frame (such as background elements, complex costumes, armor texturing, or hand geometry).
• Targeted Partials & Accelerated Corrections: Driven by lightx2v LoRA for rapid iterations, this advanced native masking sub-graph allows you to partially fix or rewrite prompt items on specified areas ([Inpaint]) while preserving the surrounding environments without full-frame re-rendering.

---

Prerequisites & Installation

To ensure all workflows load correctly without errors, please ensure your environment is fully updated:

1. ComfyUI Core: Update ComfyUI to v0.24.0 / Frontend 1.45.15 or newer. These workflows run entirely on ComfyUI's native GGUF and quantized weights implementation—no extra model wrapper extensions are needed.
2. Custom Nodes:
   • Ensure ComfyUI-GGUF (or the core node equivalent for loading GGUF/quantized model weights) and standard utility nodes (such as ComfyUI-KJNodes) are fully updated via the ComfyUI Manager.
   • If any supporting node appears red upon loading, simply use ComfyUI Manager -> Install Missing Custom Nodes.
3. Required Assets:
   • Base Model (GGUF / INT8): Download the correct Wan2.2 GGUF checkpoints or the newly supported INT8 quantized weights and place them into your models/unet/ or designated model directory.
   • LoRAs:
     • lightx2v LoRA is required across all workflows for fast sampling. Place it in your models/loras/ directory.
     • SVI2 Pro LoRA is additionally required for the SVI and SVI INT8 workflows. Place it in your models/loras/ directory.