Wan 2.2 14B i2v & t2v Enhanced Motion + 5B Latent Upscaler - Ultimate 6 Steps HD Pipeline
The ONLY workflow you need.
Fixes slow motion, boosts detail with Pusa LoRAs, and features a revolutionary 2-stage upscaler with WanNAG for breathtaking HD videos.
Just load your image and go!
🚀 The Ultimate Wan2.2 Workflow is HERE! Tired of these problems?
Slow, sluggish motion from your Wan2.2 generations?
Low-quality, blurry results when you try to generate faster?
VRAM errors when trying to upscale to HD?
Complex, messy workflows that are hard to manage?
This all-in-one solution fixes it ALL. We've cracked the code on high-speed, high-motion, high-detail generation.
This isn't just another workflow; it's a complete, optimized production pipeline that takes you from a single image to a stunning, smooth, high-definition video with unparalleled ease and efficiency. Everything is automated and packaged in a clean, intuitive interface using subgraphs for a clutter-free experience.
✨ Revolutionary Features & "Magic Sauce" Ingredients:
1. 🎯 AUTOMATED & USER-FRIENDLY
Fully Automatic Scaling: Just plug in your image! The workflow intelligently analyzes and scales it to the perfect resolution (~0.23 Megapixels) for the Wan 14B model, ensuring optimal stability and quality without any manual input.
Clean, Subgraph Architecture: The complex tech is hidden away in organized, collapsible groups ("Settings", "Prompts", "Upscaler"). What you see is a simple, linear flow: Image -> Prompts -> SD Output -> HD Output. It’s powerful, but not complicated.
2. ⚡ ENHANCED MOTION ENGINE (The 14B Core)
This is the heart of the solution. We solve the slow-motion problem with a sophisticated dual-sampler system:
Dual Model Power: Uses both the Wan2.2-I2V-A14B-HighNoise and -LowNoise models in tandem.
Pusa LoRA Quality Anchor: The breakthrough! We inject Pusa V1 LoRAs (
HIGH_resized@ 1.5,LOW_resized@ 1.4) into both models. This allows us to run at an incredibly low 6 steps while preserving the sharp details, contrast, and texture of a high-step generation. No more quality loss for speed!Lightx2v Motion Catalyst: To supercharge motion at low steps, we apply the powerful lightx2v 14B LoRA at different strengths: a massive 5.6 strength on the High-Noise model to establish strong, coherent motion, and a refined 2.0 strength on the Low-Noise model to clean it up. Result: Dynamic motion without the slowness.
3. 🎨 LOW-RAM HD UPsCALING CHAIN (The 5B Power-Up)
This is where your video becomes a masterpiece. A genius 2-stage process that is shockingly light on VRAM:
Stage 1 - RealESRGAN x2: The initial video is first upscaled 2x for a solid foundation.
Stage 2 - Latent Detail Injection: This is the secret weapon. The upscaled frames are refined in the latent space by the Wan2.2-TI2V-5B model.
FastWan LoRA: We use the FastWanFullAttn LoRA to make the 5B model efficient, requiring only 6 steps at a denoise of 0.2.
WanVideoNAG Node: Critically, this stage uses the WanVideoNAG (Nested Adaptive Gradient) technique. This allows us to use a very low CFG (1.0) for natural, non-burned images while maintaining the power of your negative prompt to eliminate artifacts and guide the upscale. It’s the best of both worlds.
Result: You get the incredible detail and coherence of a 5B model pass without the typical massive VRAM cost.
4. 🍿 CINEMATIC FINISHING TOUCHES
RIFE Frame Interpolation: The final step. The upscaled video is interpolated to a silky-smooth 32 FPS, eliminating any minor stutter and delivering a professional, cinematic motion quality.
📊 Technical Summary & Requirements:
Core Tech: Advanced dual KSamplerAdvanced setup, Latent Upscaling, WanNAG, RIFE VFI.
Steps: Only 6 steps for both 14B generation and 5B upscaling.
Output: Two auto-saved videos: Initial SD (640x352@16fps) and Final HD (1280x704@32fps).
Optimization: Includes Patch Sage Attention, Torch FP16 patches, and automatic GPU RAM cleanup for maximum stability.
Required Models (GGUF Format):
14B Models:
Wan2.2-I2V-A14B-HighNoise-Q8_0.gguf&LowNoise-Q8_0.gguf5B Model:
Wan2.2-TI2V-5B-Q8_0.ggufLoRAs: Pusa (High/Low), Lightx2v 14B, FastWan 5B
Text Encoder:
umt5-xxl-encoder-Q8_0.ggufVAE:
Wan2_1_VAE_fp32.safetensorsorpig_wan2_vae_fp32-f16.ggufUpscaler:
RealESRGAN_x2plus.pth
Required Custom Nodes: VideoHelperSuite, Frame-Interpolation, KJNodes, easy-use, gguf, memory_cleanup.
🎬 How to Use (It's Simple!):
DOWNLOAD the workflow and all models (links below).
DRAG & DROP the
.jsonfile into ComfyUI.CLICK on the "Load Image" node to choose your input picture.
EDIT the prompts in the "CLIP Text Encode" nodes. The positive prompt includes detailed motion instructions – make it your own!
QUEUE PROMPT and watch the magic unfold.
That's it! The workflow handles everything else automatically.
🔗 Download Links
Place these in your ComfyUI/models/ directory in the corresponding folders.
Model / LoRA / Component
Wan2.2 14B - High NoiseWan2.2-I2V-A14B-HighNoise-Q8_0.gguf
Wan2.2 14B - Low NoiseWan2.2-I2V-A14B-LowNoise-Q8_0.gguf
Wan2.2 5B - TI2VWan2.2-TI2V-5B-Q8_0.gguf
Pusa V1 LoRA (High)Wan22_PusaV1_lora_HIGH_resized_dynamic_avg_rank_98_bf16.safetensors
Pusa V1 LoRA (Low)Wan22_PusaV1_lora_LOW_resized_dynamic_avg_rank_98_bf16.safetensors
Lightx2v 14B LoRAlightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors
FastWan 5B LoRAWan2_2_5B_FastWanFullAttn_lora_rank_128_bf16.safetensors
UMT5-XXL Text Encoderumt5-xxl-encoder-Q8_0.gguf
Wan VAEWan2_1_VAE_fp32.safetensors / pig_wan2_vae_fp32-f16.gguf
RealESRGAN UpscalerRealESRGAN_x2plus.pth
RIFE Modelrife49.pth (Usually bundled with ComfyUI-Frame-Interpolation)
Required Custom Nodes:
ComfyUI-VideoHelperSuite
ComfyUI-Frame-Interpolation
ComfyUI-KJNodes
ComfyUI-easy-use
ComfyUI-gguf (or ComfyUI-GGUF-Integrated)
ComfyUI_memory_cleanup
Transform your ideas into fluid, high-definition reality. Download now and experience the future of Wan2.2 video generation!
Description
FAQ
Comments (31)
Nice workflow organization. Pusa lora however is placebo and does not load with i2v currently. You can check the logs in ComfyUI and it's not natively supported yet. Were you able to get PUSA working?
To be honest, PUSA, it's an addition that has very little upside. In the details I've seen, there are fewer artifacts in the final result.
A lot of workflows have 'black magic' nodes that either don't work, or worse make things much worse (teacache anyone)- this is simply because of the astonishing amateur interest in local AI, where many people lacking CS skills are suddenly immersed in the most complex arena in computer science.
Take the completely wrong idea that video models must run in VRAM to be 'fast'. 95% of people believe this utter fallacy. Which is sad, since RAM is much cheaper than a GPU with lots of VRAM, and usually a darn sight easier for a normie to upgrade.
Now where this workflow gets things right is focusing on split samplers. There is a ton of useful improvements that can be achieved by splitting samplers to focus on motion OR detail, allowing speed up LoRAs to improve render speed without catastrophic reduction in general motion or prompt adherence, although ultimately there will be trade-offs!
Pusa is not supported yet https://huggingface.co/RaphaelLiu/Pusa-Wan2.2-V1/discussions/2
oo neat -- loved your last one, I'll give this a try
Thanks, buddy! Let me know if you need anything. For this one, you can just add "Loras" to the Lora section in part 14B.
@zardozai so q8 model huh? I'm scared! lol think I can run it with a 4090 rtx?
@Robopsycho You can use any quantization method that works best for your setup. It performs well even with q4.
'Burned' frames happen when the AI is fighting to meet two opposing goals. One can see this in the new is2v model, when one forgets to add a few frames of SILENCE at the beginning of the audio sample. The AI then tries to respect the original image AND give the first frame correct lip sync detail- both of which cannot be achieved- hence the burned frame effect!
Pusa is sadly not supported yet for Wan2.2 native comfyui. Kijais Wrapper has an optional scheduler "flowedit-pusa" but I prefer native so didnt test. Link to discussion https://huggingface.co/RaphaelLiu/Pusa-Wan2.2-V1/discussions/2
Yeah this is the real deal, many wfs say that they have great motion or whatever, but this one delivers. Fantastic motion, which by itself would be reason enough to check it out. However by some witchcraft it created a 5 sec 720p 32 fps vid in 14 min on my 4080 16gb VRAM. So you're not getting only great motion, but also high fps and high res.
The downside is that it's a bit of a pita to config, it has several models like a new vae, 5b model, new lora that are not your usual stuff, and also it uses subgraphs which i mostly hate, but this is one of the few wfs that makes a good point about using them, since it's very well distributed.
TLDR: So imho, this is the best wan 2.2 out there as far as i can tell, it can take some time to configure but it's well worth it.
Thank you for your feedback; I appreciate it! At first, the subgraph may seem overwhelming, but once you get used to it, it becomes easier to use and much cleaner. Think of it like organizing folders on your machine.
I've been troubleshooting this for an hour, my WF hangs at the final upscaling step, I always get to 91% VAE Decode before it seems to stop? Using the pig2_vae (in 5B Upscale settings) and all of the correct loras and GGUF models.
Use the VAE tile Decode then instead of regular VAE Decode it should work.
why lightx2v fom wan2.1 if version for 2.2 is avaiable?
I get Error: "Loop (146,7,) - not submitting workflow"
Problem with ComfyUI? all nodes+comfyi is updated. Almost all other workflows works fine.
It seems that you may be bypassing and reactivating some subgraphs, and ComfyUI is not currently handling that properly. You will need to check each subgraph, or alternatively, the best approach is to re-download the workflow in its original state and attempt to run it again.
I am running out of RAM. My system hasunfortunately only 32GB RAM.
Any suggestions? How much RAM does it usually need?
I've put settings to minimum in the image resizer, still no luck. comfy crashes at the last step, I see RAM filling to 100% before the crash
You might want to take a closer look at your ComfyUI configuration, as I've been able to run this WF smoothly on my 32GB RAM machine without any issues.
32GB of RAM is a bad joke for video generation- full stop. RAM is cheap- and for video work is essential. 128GB is ideal, 64GB is an acceptable minimum. Too many people mistakenly think models should live in VRAM- which is only true for LLMs.
With a modern NVMe drive (5GB/s) it would be possible to bypass RAM and stream from flash to the GPU, but Comfy is simply not coded to do this. In fact the memory management of Comfy is a literal disaster, since it is built without any awareness of proper data flow for image and video generation.
Instead we have incorrect 'urban myths' about how non-LLM AI should function. Since we all have to use Comfy, theory is irrelevant tho, so invest in more RAM if you can.
@blobby99 it's 💯 true that memory management in ComfyUI is a joke...
@blobby99 thanks. It is not my machine, therefore I have to stick to 32gb for now
@zardozai not really sure what could cause the high memory consumption or how to check my configuration
@zardozai I have --cache-none and --normalvram set in cmd line. Do you use any such params when starting comfy?
@mcanti In fact, I am using exactly those settings. Perhaps you could try a lower quantization GGUF?
@zardozai I removed --normalvram and increased my swap file from 8gb to 16gb and I don't have the OOM crash anymore. Not sure which did the trick but I successfully run your Workflow twice already :)
Thanks!
You guys are forgetting Swap file in Windows/Linux! Try to use up to 30gb (yes, all of that) as swap file (if you have 2 ssds, then 15gb each), even if you have only 16gb of RAM. It does wonders! I have 32gb with a 3060 (12gb VRAM) and OOM are rarer now.