WAN 2.2 – I2V Workflow (Optimized for 12GB GPUs)
A fast, clean, and VRAM-efficient Image-to-Video workflow built around WAN 2.2. Fast render times on mid-range GPUs. I tried to keep this simple and easy to use, while maintaining good results. Utilizing well known nodes, and minimizing node bloat. The workflow also has comments everywhere and clear flow.
Ver 1.0 - Base workflow, can do 5 second clips in one iteration. (very fast for 12gb)
Ver 1.1 - More stability, can run 100 times consecutively in 8hrs
Ver 1.2 - Renders 20 second videos. Cleanup of wires.
Ver 1.3 - MMAudio added.
Ver 1.4 - 2x Upscaling, color correction, & sharpening in between passes for quality consistency.
Ver 1.5 - Fixed MMAudio, Updated controls & ability to do 5, 10, 15, & 20 second videos easy. Split RIFE between phases. Fixed prompts. Cleaned up workflow.
WAN 2.2 Model
Use either lightweight GGUF models or full.safetensorscheckpoints.
Lightning LoRAs are baked into the .safetensors model, so LoRA use is optional.WAN 2.1 VAE
Reduces VRAM load while maintaining strong color, detail structure, and temporal consistency.SageAttention + FP16 Accumulation Patch
Automatically applied for speed and throughput.Dual KSampler passes (HIGH/ LOW)
Uses 6 steps, and 101-frame length for smooth animations and solid adherence.LoRA Loader + Model Shift Controls
Supports stylistic LoRAs with synchronized shift values across both samplers.RIFE Frame Interpolation (60 FPS)
Creates ultra-fluid motion.Upscale + Adaptive Sharpen Pass
In between each phase ultrasharp upscaler is used for consistency.
Performance
GPU: 12GB VRAM
Render Speed: ~6 minutes (v1.0)
Output FPS: ~60–64FPS
Resolution: ~1072 × 1616 (post-upscale)
Requirements
KJNodes
MMAudio
Rgthree Nodes
ComfyUI-Frame-Interpolation
Cache-Cleaner
CG-Use-Everywhere
ComfyUI-Easy-Use
ComfyUI-Essentials
ComfyUI-Videohelpersuite
ComfyUI-Custom_Scripts (https://github.com/pythongosssss/ComfyUI-Custom-Scripts)
GGUF Nodes
AIO Wan2.2 Model: https://civarchive.com/models/2053259/wan-22-enhanced-nsfw-or-camera-prompt-adherence-lightning-edition-i2v-and-t2v-fp8-gguf
GGUF Model (if not using AIO): https://civarchive.com/models/1820829/wan22-i2v-a14b-gguf
Lightning Lora (If not using AIO): https://civarchive.com/models/1585622/lightning-lora-massive-speed-up-for-wan21-wan22-made-by-lightx2v-kijai
CLIP: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/text_encoders
VAE: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/vae
Ultrasharp: https://huggingface.co/Kim2091/UltraSharpV2/tree/main
Description
Added color correction, 2x ultrasharp upscaling between passes, & sharpening nodes for consistency between phases.
FAQ
Comments (26)
Hello, Génial ! et géniale la correction de couleur. belle qualité. merci
@sdktertiaire2 Bonjour, de rien ! Je suis ravi de cette mise à jour, car elle améliore considérablement la qualité des coutures ! Thank you!
Im getting an error from Load upscale model
You will need ultrasharp https://huggingface.co/Kim2091/UltraSharpV2/tree/main
Yeah I figured it out and Ill say thankyou. Of all the long vid workflows I've tried yours is the best. Way less degradation compared to others.
@Yoorak_Hunt appreciate it! I worked on it for a long time to figure out a solution to the degradation problems
how to install ultrasharp, and what version, anyone?
@rsamd123923 grab 4x-UltraSharp.pth from the link above and put it in comfyui/models/upscale_models
@EchoSingularity got it running thx
Thanks for this. if you had a 5090 with 32gb vram, how would you modify your workflow to take advantage of it?
I would up the resolution past the 812. Try like 896 or higher. Send me a 5090 I can get you some killer results lol
@EchoSingularity did you already receive the new GPU, and test? =)
@efv7792 I wish! Lol still on a 5070
@EchoSingularity shame... @fancypantzzz whatup, send GPU already... pfff
@efv7792 Yeah @fancypantzzz its for science!
bruv whazaaaaa send nu... send GPUS!
Great workflow but i cant get more that one segment to work. you state in the workflow (make sure to check video combine save option) i cant find this option anywhere. am i blind?
What is the behavior that only 1 segment works? Is it looping one segment? Crashing? Look at each segment are there preview images for the first image of the segment being shown?
@EchoSingularity I was getting the same result stopping on Sampling 1/4 - I didn't have the Ultrasharp installed during those iterations, I am attempting again as we speak.
It seems to continue to only spit out the first generated PNG and call the workflow done.
investigating with AI now.
@EchoSingularity Turns out it was a me issue, not following directions. I installed the:
open the three UpscaleModelLoader nodes (433, 437, 435) in ComfyUI and change their dropdown from 4x-UltraSharp.pth → 4x-UltraSharpV2.pth
You have the reg ultrasharp in this workflow. I just needed to update these in (Load Upscale Model) nodes.
For the record, this is my 2nd day doing this stuff! lol
@angrydocks you are correct. i made the change and all working now. i'm new to this too. took me a while to workout i could expand the node lol. again thx
This is a great workflow! However, running into a problem with MMAudio. Generating a 20 second vid and audio seems to not be synced. Like I'll hear audio in the first step of the video, that should be playing in the second. I think with MMA it is unnecessary to have 4 positive prompt boxes. I think it would be a bit better if the full video finished processing and then the video was ran through mmaudio. Otherwise, great workflow, thank you.
Edit: Also, it sometimes with make separate parts of a 20 second video sound completely different. Again, I think it would be cool if mmaudio was ran after the full video was processed.
I can try getting that to work but last time I tried it kept crashing sending all the frames at 60fps through MMAudio in one run. It's likely not syncing right because of how the audio combines. I'll look into it thanks!
The problem is that MMAUDIO works fine with 24 frames
I use ComfyUiportable - I have been able to get 12! YES 12 sucessful videos across the finishline using this workflow! Thank you
I am now trying to get Load Random Image working, but having issues locating the folder in the input folder. I have tried different folders, directly to input - unsure where to go from here. until resolved, I will continue to run my Single Load image which is working FABULOUSLY now!
ProTip - Read Directions!
when using the random image node I use the pathing starting at /Desktop/..../.../folder