Local AI video workflow research Verified · Aug 21, 2026

EasyWan22

A one-click Wan 2.2 production environment for Windows

EasyWan22 combines Wan2.2 I2V-A14B, ComfyUI, GGUF quantization, WanVideoWrapper, SageAttention, LoRA presets and a full post-production chain into one large workflow built for repeatable local video output.

Wan2.2 is the engine. ComfyUI is the workbench. EasyWan22 is the assembled vehicle.
Explore the workflow Open GitHub repository
Original illustration of a local AI video workstation turning draft frames into refined video through a node workflow
Image → Motion → Finish
Windows localWan2.2 I2VLow VRAMBatch workflow
WindowsPrimary platformShort install paths are recommended
8GBRecommended VRAM floorRTX 30-series or newer in the README
1.1MBMain workflow JSONOne large all-in-one I2V graph
82/100Overall assessmentWorkflow value, not a speed benchmark
01 · Category definition

A curated ComfyUI distribution—not a foundation model

The project productizes a difficult open-source stack for Windows users. Its value comes from installation orchestration, pinned dependencies, sensible defaults and a video-production method that extends well beyond generation.

Not a new video modelEasyWan22 ≠ Wan2.2
Engine

Wan2.2 I2V-A14B

High-noise and low-noise model stages provide the underlying image-to-video generation.

Workbench

ComfyUI + WanVideoWrapper

Nodes expose loading, sampling, memory swapping, conditioning and video I/O.

Optimization

GGUF + FastMix + SageAttention

Quantization, faster draft models and attention optimization make 14B-class local use more practical.

Production layer

EasyWan22 workflow

Presets, variations, refinement, local repair, upscale, interpolation, color, looping and batch output become one process.

It is designed to reduce clicks while producing many videos—not to teach the internal logic of ComfyUI.

02 · Production method

Draft cheaply, pick a winner, then spend compute on finishing

The central idea is Seed Gacha: generate inexpensive variations first, select the promising motion, and only then run costly post-processing. This changes local AI video from one-shot generation into a selection-and-finishing loop.

01 · Draft

Fast variations

Use FastMix or lower resolution, disable expensive finishing, and test several seeds.

→
02 · Select

Pick the useful motion

Compare composition, motion and identity before committing more GPU time.

→
03 · Finish

Refine and deliver

Run low-noise refinement, local detail repair, upscale, interpolation and color correction.

01Input image
02Motion preset
03Prompt enhance
04Fast draft
05Seed variations
06Pick winner
07Refiner
08Detailer
09Upscale
10Frame interpolation
11Color match
12Trim · loop · export
03 · Production toolset

More like an AI video post workstation than a Generate button

The large workflow packages the decisions a production user repeats most often, including repair and delivery stages that many hosted generators leave outside the product.

Enhance

Low-noise Refiner

Redraw a selected draft at higher resolution and use an additional seed to explore refined variants without restarting from zero.

Repair

Region Detailer

Detect faces, heads, eyes, hair or other regions; crop, enlarge, regenerate and composite only the weak area back into the video.

Deliver

Upscale + interpolation

Use an anime-oriented 2× upscaler and frame interpolation options for smoother 30/60 fps delivery.

Control

Start, end and loop

Condition on start and end images, swap their direction, or build A→B→A loops for wallpapers, GIFs and idle animation.

Prompt

Translate + rewrite

Translate Japanese prompts and expand short instructions with a bundled Qwen2.5 3B prompt extender before generation.

Preset

LoRA + FastMix modes

Bundle a LoRA, trigger phrase and strength into a selectable motion preset; switch between fast drafting and base-model quality.

04 · Low-VRAM engineering

How a 14B workflow reaches consumer GPUs

EasyWan22 combines Q4_K_M GGUF weights, transformer block swapping between GPU and system RAM, and SageAttention. It lowers the entry barrier; it does not make a 6GB card perform like a 24GB card.

GGUF Q4_K_M

Quantizes both high-noise and low-noise weights to reduce memory pressure.

BlocksToSwap

Moves more transformer blocks through system RAM when VRAM is scarce, trading speed for feasibility.

SageAttention

The default launcher enables an optimized attention path to reduce compute and memory pressure.

Use levelGPUSystem RAMPractical verdict
Barely runsGTX 1660 Ti · 6GB16GBPossible in a reported test; heavy offload and long waits
Recommended floorRTX 30-series+ · 8GB32GBREADME target; quantization and swapping still matter
Comfort zoneRTX 3060 · 12GB+64GBAuthor sample environment; better for repeated generation

Less VRAM means more RAM traffic and PCIe transfers. ‘Runs’ describes compatibility, not comfortable generation speed.

Plan storage before installation

120GBPractical minimum
150GBRecommended reserve
200GB+Long-term SSD workspace
05 · Pinned environment

A complex distribution made stable by freezing versions

The installer coordinates Python, ComfyUI, PyTorch, FFmpeg, models and a broad custom-node ecosystem. Version locks made the 2025 setup repeatable, while creating dependency-drift risk as drivers and upstream projects continue to evolve.

Versions pinned by Setup.bat

ComfyUI
v0.3.55
ComfyUI Manager
3.35
PyTorch
2.7.1 + cu128
TorchVision / Audio
0.22.1 / 2.7.1
Triton Windows
3.3.1.post19
SageAttention
2.2.0
Transformers
4.49.0
Python
3.10

Installation shape

  1. 1

    Create a short empty path such as C:\EasyWan22

  2. 2

    Save and run EasyWan22Installer.bat

  3. 3

    Provide a Civitai API key when requested

  4. 4

    Let the installer fetch Python, ComfyUI, nodes, models and FFmpeg

  5. 5

    Launch ComfyUi.bat

  6. 6

    Open Easy/00_I2V(ImageToVideo) and queue the workflow

A whole ComfyUI ecosystem underneath

WanVideoWrapper, ComfyUI-GGUF, KJNodes, Florence2, Frame Interpolation, Impact Pack, DepthAnythingV2, Segment Anything 2, RMBG, VideoHelperSuite, ControlNet Aux and more are orchestrated as one distribution.

WanVideoWrapperComfyUI-GGUFKJNodesFlorence2Frame InterpolationImpact PackDepthAnythingV2SAM 2RMBGVideoHelperSuiteControlNet AuxComfyUI Essentials
06 · 2026 reality check

The same version locks that created stability now create aging risk

The repository is still public and used, but code activity has slowed. Evaluate it as a maintained snapshot that may need manual repair—not as a guaranteed current installer.

Repository snapshot244 ★12 Forks10 Open IssuesLast push · 2025-11-30
01

Slower maintenance

The last code push shown by GitHub is November 30, 2025, although users continued opening issues in 2026.

244 stars · 10 open issues
02

One-click downloads can fail

An open April 2026 issue reports that many preset LoRAs and models no longer download automatically after external service changes.

Issue #50 · open
03

Dependency drift

Old ComfyUI, Triton and SageAttention combinations can conflict with newer drivers and hardware; a DLL-import failure is already documented.

Pinned stack vs changing system
04

Model scope is moving on

A June 2026 request asks for an LTX 2.3 equivalent, reflecting demand for a model-agnostic video workflow rather than a Wan2.2-only bundle.

Issue #51 · no response
07 · Product lesson

Copy the user journey, not the giant node graph

The enduring insight is a model-independent finishing pipeline. A modern product should expose user intent and hide samplers, block swapping, noise stages and loader details unless advanced control is requested.

EasyWan22 conceptModern product control
StartImage→Upload image
Prompt + rewrite→Describe motion · Enhance prompt
LoRA preset→Motion / effect
Seed Gacha→Variations ×4
FastMix / Base→Fast / Quality mode
Refiner + Detailer→Enhance · Fix face
Upscale + Interpolation→HD · Smooth motion
EndImage + RepeatFade→End frame · Loop video
Input → variations → select → repair → enhance → deliver is a stronger product loop than model → prompt → generate → download.
08 · Who it fits

Still worth studying for local Wan2.2 I2V—less ideal as a future-proof platform

Its best audience understands that a one-click distribution can still require manual troubleshooting in 2026.

Strong fit

  • Windows users with RTX 3060 12GB or similar
  • Batch I2V and seed exploration
  • Anime, illustration and LoRA video workflows
  • Creators who value local files and repeatable presets

Weak fit

  • macOS or AMD GPU users
  • People learning ComfyUI fundamentals
  • Users expecting the newest multi-model stack
  • Teams requiring a fully supported commercial production system

Research scorecard

Scores separate its original workflow achievement from its current maintenance and technical freshness.

2025 innovation9/10
One-click deployment value9/10
Low-VRAM optimization9/10
Production workflow9/10
ComfyUI learning value4/10
2026 technical freshness5/10
Current maintenance4/10
Long-term product insight8/10

The larger opportunity

Turn Seed Gacha → Pick Winner → Refine → Detail → Upscale → Interpolate → Extend → Loop into a model-agnostic Easy AI Video workflow that can switch between Wan, LTX, Hunyuan, Seedance and future engines.

Primary resources EasyWan22 repository Post-process Wiki Wan2.2 official project

© 2026 wan2.video