Wan 2.2 & FLUX Krea Full Tutorial - Automated Install - Ready Perfect Presets - SwarmUI with ComfyUI - Install Wan 2.2 and FLUX Krea with literally 1-click and use our pre-made most amazing quality presets.
Tutorial URL
Info
Install Wan 2.2 and FLUX Krea with literally 1-click and use our pre-made most amazing quality presets. I did literally 100s of parameter testing you to bring you best Wan 2.2 and FLUX Krea configuration so that you can generate the highest quality videos from images or from text. Moreover, with our FLUX preset, you will be able to generate much better quality images from FLUX Dev with FLUX Krea DEV model. This tutorial will show you everything step by step easiest possible way.
Links
🔗Follow below link to download the zip file that contains SwarmUI installer and AI models downloader Gradio App - the one used in the tutorial for downloading models, presets, prompt generator guide txt ⤵️
▶️ How to install SwarmUI main tutorial : https://youtu.be/fTzlQ0tjxj0
🔗Follow below link to download the zip file that contains ComfyUI 1-click installer that has all the Flash Attention, Sage Attention, xFormers, Triton, DeepSpeed, RTX 5000 series support ⤵️
▶️ RunPod SwarmUI & ComfyUI Install Tutorial : https://youtu.be/R02kPf9Y3_w
▶️ Massed Compute SwarmUI & ComfyUI Install Tutorial : https://youtu.be/8cMIwS9qo4M
🔗 Python, Git, CUDA, C++, FFMPEG, MSVC installation tutorial - needed for ComfyUI ⤵️
🔗 SECourses Official Discord 10500+ Members ⤵️
🔗 Stable Diffusion, FLUX, Generative AI Tutorials and Resources GitHub ⤵️
🔗 SECourses Official Reddit - Stay Subscribed To Learn All The News and More ⤵️
Presets
Video Chapters
0:00 Introduction: The Ultimate Wan 2.2 Tutorial with Optimized Presets
1:03 Free Prompt Generation Tool & Introducing the New FLUX Krea Dev Model
2:01 How SwarmUI & ComfyUI Enable Video Generation on Low-End Hardware
2:46 Quick Start Guide: Downloading the Latest SwarmUI & ComfyUI Installers
3:10 Step-by-Step: How to Update or Perform a Fresh Installation of ComfyUI
3:51 Step-by-Step: How to Update or Perform a Fresh Installation of SwarmUI
4:18 Essential Setup: Configuring the SwarmUI Backend for ComfyUI
4:53 One-Click Setup: Downloading All Required Wan 2.2 Models Automatically
5:46 Importing the Ultimate SwarmUI Presets Pack for Best Results
6:22 Wan 2.2 Image-to-Video Generation: A Complete Step-by-Step Guide
7:33 How to Generate Amazing, Detailed Prompts for Free with Google Studio AI
8:12 Starting Your First Generation & How to Monitor Logs for Errors
8:53 Pro Tip: How to Fix Low GPU Utilization and VRAM Issues for Max Speed
10:32 Wan 2.2 Text-to-Video: Choosing the Right Preset & Workflow
11:22 Generating a Detailed Dinosaur Animation Scene from a Simple Text Prompt
12:15 In-Depth Analysis: 8 Steps vs 20 Steps & The Impact of LoRA on Quality
13:11 Finding the Best Parameters: A Deep Dive into CFG Scale & Step Counts
13:42 Advanced Optimization: Using TeaCache for Text-to-Video Generation
15:31 FLUX Krea Dev vs FLUX Dev: A Detailed Side-by-Side Image Comparison
16:26 How to Easily Train Your Own LoRAs on the New FLUX Krea Dev Model
17:02 Complete Workflow for Generating High-Quality Images with FLUX Krea Dev
18:20 The Final Verdict: Side-by-Side Result of FLUX Krea Dev vs FLUX Dev
19:20 An Experiment: Attempting to Generate Still Images with Wan 2.2
21:18 Final Thoughts, Summary, and What's Coming Next in Future Tutorials
Wan Development Team Announces Wan2.2 Video Generation Model
The Wan development team has announced the release of Wan2.2, positioning it as a significant advancement in their video generation model series. The team highlights four key technical improvements in this iteration.
The model introduces a Mixture-of-Experts (MoE) architecture specifically designed for video diffusion models. According to the developers, this approach separates the denoising process across different timesteps using specialized expert models, which they claim expands overall model capacity while maintaining computational efficiency.
Wan2.2 incorporates what the team describes as "cinematic-level aesthetics" through curated training data that includes detailed annotations for visual elements such as lighting, composition, contrast, and color tone. The developers state this enables more precise control over cinematic style generation and customizable aesthetic outputs.
The training dataset for Wan2.2 represents a substantial expansion compared to its predecessor, with the team reporting increases of 65.6% more images and 83.2% more videos in the training corpus.
Exploring Wan 2.2: The Next Frontier in AI-Driven Video and Image Generation
Introduction
In the rapidly evolving landscape of artificial intelligence, generative models are pushing the boundaries of creative content creation. Among the latest advancements is Wan 2.2, a sophisticated AI model designed specifically for high-quality video and image generation. Developed as an upgrade to its predecessor, Wan 2.1, Wan 2.2 introduces enhanced capabilities in multimodal content creation, making it a powerful tool for artists, filmmakers, advertisers, and developers. Released in recent months, this model stands out for its focus on cinematic-quality outputs, efficient performance on consumer hardware, and innovative features like advanced camera control and temporal consistency.
This article delves into the intricacies of Wan 2.2, exploring its features, models, capabilities, improvements, system requirements, and practical applications. Whether you're a creative professional or an AI enthusiast, understanding Wan 2.2 can unlock new possibilities in digital media production.
What is Wan 2.2?
Wan 2.2 is a generative AI model built by Wan AI, optimized for the "language of film." It excels in text-to-video (T2V) and image-to-video (I2V) generation, producing realistic, high-resolution videos and images from descriptive prompts. Unlike traditional video editing tools, Wan 2.2 leverages deep learning to create content from scratch, incorporating elements like motion, lighting, and composition intelligently.
At its core, Wan 2.2 is designed for cinematic storytelling, allowing users to generate videos that mimic professional film techniques. It supports resolutions up to 720p at 24 frames per second (fps) and can produce clips up to 5 seconds long in its base configuration. The model is multimodal, meaning it handles both static images and dynamic videos seamlessly, with enhancements in special effects, style adjustments, and cross-modal transformations (e.g., converting an image into an animated video).
Wan 2.2 was introduced as an open model, accessible through platforms like Hugging Face and integrated into tools such as ComfyUI, making it available for both hobbyists and professionals.
Key Features
Wan 2.2 boasts a suite of advanced features that set it apart in the AI video generation space:
Enhanced Image and Video Generation: Improved detail in textures (e.g., skin, fabrics, landscapes), higher-resolution rendering, and precise style controls like brushstroke intensity and color saturation.
Special Effects and Motion: Realistic lighting (global illumination, reflections), particle systems for effects like smoke or water, and smoother animations with reduced flickering.
LoRA Training Upgrades: Faster training (over 50% reduction in time), few-shot learning with minimal samples, and multi-model fusion for personalized styles.
Cross-Modal Capabilities: Seamless image-to-video conversions (e.g., adding motion to static scenes) and style consistency across formats.
Camera Control System: Mathematical precision for movements such as dolly, pan, crane, and handheld shots, including motion blur, speed ramping, and easing curves.
Composition Intelligence: Automatic adherence to rules like the golden ratio, rule-of-thirds, and platform-specific optimizations for aspect ratios.
SNR Switching Mechanism: A signal-to-noise ratio-based system that switches between high-noise (structure-building) and low-noise (detail-refining) experts during generation, mimicking professional workflows.
Temporal Consistency: Ensures coherent frame-to-frame transitions, maintaining character persistence and smooth motion.
These features make Wan 2.2 particularly suited for creating lifelike, professional-grade content with minimal user intervention.
Models Available
Wan 2.2 is released in multiple variants to cater to different use cases and hardware setups:
Model VariantParametersCapabilitiesKey FilesTI2V-5B5 BillionText-to-Video and Image-to-Videowan2.2_ti2v_5B_fp16.safetensors, wan2.2_vae.safetensorsT2V-A14B14 Billion (MoE: 27B total, 14B active)Text-to-Videowan2.2_t2v_high_noise_14B_fp8_scaled.safetensors, wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors, umt5_xxl_fp8_e4m3fn_scaled.safetensors, wan_2.1_vae.safetensorsI2V-A14B14 Billion (MoE: 27B total, 14B active)Image-to-Videowan2.2_i2v_high_noise_14B_fp8_scaled.safetensors, wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors, umt5_xxl_fp8_e4m3fn_scaled.safetensors, wan_2.1_vae.safetensors
The Mixture-of-Experts (MoE) architecture in the 14B models uses dual experts (High-Noise and Low-Noise) for efficient processing.
Capabilities and Performance
Wan 2.2 generates 5-second videos at 720p resolution and 24fps, with options for text or image inputs. It outperforms competitors like VEO3 and Kling 2.1 in realism and lifelike quality, particularly in human image and video generation.
Performance highlights include:
Generation time: 9 minutes for TI2V-5B on an RTX 4090; 2-3 minutes for A14B on 8x GPUs.
Higher quality scores: Rated "Good to Excellent" compared to baselines.
Support for natural language prompts, eliminating the need for technical parameters.
Demos and examples showcase applications like turning static images into dynamic scenes (e.g., a character smiling or leaves blowing in the wind) and creating cinematic sequences from descriptive text.
Improvements Over Previous Versions
Compared to Wan 2.1, Wan 2.2 offers significant upgrades:
Training Data: 65.6% more images and 83.2% more videos, with advanced aesthetic labels for better scene and motion understanding.
Efficiency: MoE architecture reduces computational load; faster generation times (e.g., from 45-60 seconds to minutes).
Quality: Superior temporal consistency, camera controls, and composition tools.
Prompting: Shift from rigid technical prompts to intuitive natural language.
These enhancements make Wan 2.2 more practical for professional use, running on accessible hardware while delivering broadcast-quality results.
System Requirements and Integration
Hardware: TI2V-5B runs on consumer GPUs like RTX 4090 (24GB VRAM recommended). A14B models require 8x GPUs or cloud services for optimal performance.
Costs: Local setup around $4,000; cloud via FAL.ai at $0.08-0.12 per video.
Integration options include:
ComfyUI: Visual workflows for T2V and I2V, with JSON examples available.
Diffusers and APIs: Standardized pipelines for easy migration.
Upgrade Process: A phased approach involving assessment, prompt migration, integration, and validation (typically 4+ weeks).
Use Cases
Wan 2.2's versatility shines in various domains:
Digital Art and Illustration: Generating concept art or animated illustrations.
Advertising: Creating promotional videos and product visuals.
Film and Pre-Visualization: Storyboard animations and scene prototypes.
Game Development: Designing characters, environments, and skill effects.
Social Media and Content Creation: Quick, high-quality short videos.
Conclusion
Wan 2.2 represents a leap forward in AI-generated media, blending creativity with technical prowess to democratize professional video production. With its open-source nature, efficient models, and cinematic focus, it empowers users to bring imaginative concepts to life effortlessly. As AI continues to advance, tools like Wan 2.2 will undoubtedly shape the future of content creation. For those interested in exploring it, resources on platforms like Hugging Face and ComfyUI provide a great starting point.
Comments (24)
When I try to generate any images to video with the Wan 2.2 lightx2v, all of my results turn out blurry or "ghosty" if that makes sense. Any suggestions? No LoRAs or anything, just the preset as is. Fresh install.
i noticed that too. try to shorten your prompt and make it simpler it helped me
With the new preset my I2V at 20 steps now takes 4 hours instead of 2. I notice that sigma shift is now 5 instead of 8. Can that have anything to do with it being slower?
it is the new sampler. it is slower but way higher quality. perhaps you may do lesser steps try 10 still should yield excellent results
Any way we can make longer videos than 5 seconds? I'm able to squeeze out almost 8 seconds with 12 fps and then using Flowframe with 5x Rife for 60 fps. But 20-30 seconds would be great for YouTube shorts. I tried to take the last frame from the video I generated and then generate a new one based on that but the transition wasn't good.
did you try SwarmUI Video Extend Frame Overlap ? it might help give it a try
Hej, I've choose Wan 2.2 lora and it says thats it's set lora from wan 2.1 and give me an error: Invalid value for parameter LoRAs: Invalid value for param LoRAs - 'Wan21_I2V_14B_lightx2v_cfg_step_distill_lora_rank64_fixed' - must be one of: `(None)`" Please help if I do smth wrong or fix if it's not my fault. thx :)
hello. that means either your swarmui or comfyui backend is not up to date. our presets are all working i just tested yesterday
Hi Secourse, did you will make, like you did for flux, a wan 2.2 lora trainning, and finetuning, parameters for text to image ? I searched on your patreon but there is nothing I seen about wan lora trainning. And thank you for all the thing you do !
yes i am working on both. text to image parameters hopefully soon as a preset
Can you make "Best_Working_Params" for videos? For example, you mention that 40 steps is excellent quality but there is no preset for that.
well for image to video i tested and it really not worth it. for text to video we have 20 step and you can just increase it to 40. but sure i will add those too
I have a 3090, python 3.12.7, pytorch 2.51+cu124. Can I use the auto install? If so do I need to update anything first?
you can use we have python 3.12 installer.
T2V is working fine but for I2V I'm getting "Sizes of tensors must match except dimension 2. Expected size 36 but got size 16 for tensor number 1 in the list." Have you run into this before? I made sure to submit a 960x960 pixel image but still no dice.
did you update comfyui and swarmui fully? it may cause
Is there a way to extend the video (max for me is 96 fps), I think we used something called Rife in FramePack? Like generate 96 fps but in 12 fps speed and then use Rife x2 to make it smooth?
it may work in some cases. so you can set fps to like 12 and set max frame 121 and see what result you can get. it will be 10 second video. then you can rife 4x to make it 48 fps
I2V WAN 2.2 without Lora and 20 steps yields better results as with T2V WAN 2.2?
Yes t2v Wan is not great yet with loras. I tested all
Running out of memory following the video for Wan 2.2 generation of image to video having applied the preset. Windows 11, RTX 3080 16 Gb VRAM laptop, 64 Gb RAM. I have tried to lower the resolution to 256x256 but either it stops due to out of memory or it uses shared VRAM at about 20 Gb. I don't have lots of software on and VRAM before generation is 0,1-0,2 Gb of 16,0 Gb. SwarmUI: "ComfyUI execution error: Allocation on device This error means you ran out of memory on your GPU. Update: Okay so I had set it to 121 frames and lowered it to 96 frames and then it worked.
Great. That is why I set 73 frames so you can initially test
Generating an 80 frame long, 1280x720 I2V was around 38 min. on a 4090. Lowered the resolution to 848 reduced the time to 3.5 min. Looking forward to your jsons. Thanks!
awesome