This is a simple workflow for the WAN2.2_T2V_14b and WAN2.2_I2V_14b models, perfect for beginners. Essentially, it’s just a basic workflow that I’ve only slightly modified.
Key features: it works with SageAttention (which roughly doubles generation speed) and is friendly to systems with limited VRAM — it can run even on 8 GB graphics cards. Generating a 5‑second clip at 480×720 resolution takes around 5 minutes, while a 2‑second clip takes about 2 minutes. Hopefully, someone will find this helpful.
When setting up your project, locate the EmptyHunyuanLatentVideo or WanImageToVideo node and adjust the length parameter: - Set it to 33 for a 2‑second video - Set it to 81 for a 5‑second video.
A great step-by-step guide on installing Triton and SageAttention for ComfyUI Portable can be found here. It’s a straightforward set of clear, easy-to-follow instructions.
There’s also an excellent beginner-friendly prompting guide from the Civitai team here.
If you’re looking for the LoRA Lightx2v, you can find it here.
Kijai’s models are available here.
Wishing you creativity, inspiration, and happy generating!
Description
FAQ
Comments (45)
Great workflow! 3080 12GB not TI. Generation times <5 minutes, 270-280 sec.
Nice! Thanks for the feedback, and good luck with your generations!
Power Lora Loader (rgthree) 65:
- Required input is missing: clip
Why is not the second Power Lora Loader module connected to clip?
Make sure you're using the correct text encoder model in the “Load CLIP” node — it should be: umt5_xxl_fp8_e4m3fn_scaled.safetensors
@Akalabeth I have that one, but its not the problem. In your workflow, there are 2 "Power Lora Loader" modules. One is connected to "Load Clip" but the other not. The message says that both require to be connected right??
P.S. I saw that both diffusion models are 15GB. I use 2 of 6GB: "Wan2.2-I2V-A14B-High/LowNoise-Q3_K_S". Do you think this is a better option? (I dont have enought disk space for your models lol)
@Mewris
No, the second LoRA doesn’t need to be connected. In fact, if you connect that LoRA to the CLIP, it will automatically disconnect the first one. To be honest, I haven’t fully figured out yet why the second power LoRA doesn’t connect to the CLIP, but I’ve seen the same trick used in other people’s workflows.
You can use any version of the WAN High or WAN Low models: the original ones, GGUF, or Kijai versions - you won’t notice much of a difference.
@Mewris I took another careful look and realized I didn’t explain it to you quite correctly. In my workflow, both Power Lora Loaders (rgthree) are connected to the Load CLIP node - it’s just that the first one is further linked to both the positive and negative prompts (via CLIP Text Encode), while the second one isn’t. That’s why I said the second one isn’t connected to CLIP. Try unpacking my workflow again and running it, just replacing the models with the ones you need.
Sorry for the misunderstanding.
@Akalabeth thanks for the help dude!! I'll try it 😁
Great WF
Thank you so much for making this workflow, this is the first Wan 2.2 related workflow that's worked perfectly for me. I've went through a lot of them this week and they just didn't work. This one actually works and includes a space for LORAs. It's pretty fast too and gives great quality for images and videos.
I know the feeling — that’s why I decided to share this workflow. I’m really glad it’s working for you! I’d love to see a couple of your generations if you feel like sharing.
Hello I have a stupid question. It runs great until the FILM VFI box and errors out saying it could not detect film_net_32.pt. is this a model I have to download?
Thanks to the Film VFI node, animations become noticeably smoother and more natural. You can play around with the FPS settings — I usually go with 32, but some people prefer 24. I recommend sticking with 32.
The model file itself should be placed in: custom_nodes\comfyui-frame-interpolation\ckpts\film You can download it from HuggingFace, although in my case it downloaded automatically when I installed Film VFI through Comfy Manager.
All in all, it’s a must-have for smoothing transitions in animations.
@Akalabeth thank you for your response. For some reason it didn't install it with the comfy manager. Also thank you for guiding me where to put it. This runs flawlessly now. Thank you so much and great work!
@tenhunter500 Glad everything worked out for you! I’d really appreciate it if you shared a couple of your works under the post — only if you feel like it, of course. It would mean a lot to me.
Thanks for these workflows, really. I'd nearly given up on 2.2 until I found these. I am having an issue with t2i, though: No matter what I prompt, I cannot control the lighting in my image. I can specify the type of light, direction, time of day, etc. and the generations all look like flash photography. Very strange. Models and environments all look correct, but the lighting is wrong. Adjusted schedulers, samplers, prompts, steps, lora weight, and nothing fixes it.
A lot of it is the WAN model’s fault, but there are still a few things you can try—for example, add everything related to the type of lighting you don’t want into the negative prompt.
For example, for a flash photo you could use the following in the negative prompt: overexposed lighting, harsh shadows, blown-out highlights, unnatural skin tone, shiny forehead, oily skin glare, eye glare, lens flare, light reflection on skin.
I’m not sure how much it’ll actually help, but it’s definitely worth a try.
By the way—thanks for the feedback!
THANKS A LOT for this workflow, with SAGE + FP16 ACCUM. + LIGHTING 4-STEP + LIGHTX2V STEP DISTILL, wan22 has a great quality-speed ratio. I'm able to generate 1024*768 @ 97 frame length on a 5070 12Gb with 5-6 stacked loras on both HIGH and LOW passes, and if i need additional crispness I add Ultimate SD upscaler with 4Xultrasharp or 2 passes of 2XRealESRGAN for a smoother refining. Think I'm going to be sticking with this setup for 2025 2026. Cheers.
huh its been 8 months and I was gonna ask a question but will do anyways;
what was your total run time?
and did you find anything else better than this workflow for your 12gb vram?
I had to bypass Sageattention. Dont know why Triton and Sageattetion always make my workflow fail. I even tryed to go back to python 3.12 instead of 3.13 and still failed
For Python 3.13, you can try installing Triton and SageAttention following the updated guide.
Don't ever use sage attention if you have 16VRAM card and above. Below 16VRAM is peasant land. I never use it. A bitch to install and barely no gain lol
@ProvenFlawless I manage to install and I though something was wrong because the video is taking longer using sagetention ! Thanks man !
@AkalabethI manage to install it but the video takes longer to create, will keep it on bypass ! thank you for the workflow!
I get this error.. Failed to set fp16 accumulation, this requires pytorch 2.7.0 nightly currently
You just have to do exactly what it tells you, update your pytorch to version 2.7.0 nightly, and also you should launch your comfy with (run_nvidia_gpu_fast_fp16_accumulation) and if you dont want to update, you can the bypass the (Model Patch Torch settings) node or simply turn the switch off in the node itself
I Hope it helps 😊
@VFX__SENSEI i have tried but is not working
@Manugt Do you have sage attention and triton installed?
If not, then you should also bypass sage attention and run comfy without fp16 accumulation, just run using comfy.bat
Let me know if this helps and also which gpu are you using ? If its compatible i can help you install sage attention and triton
Getting the same error. Looked around in Comfy, but I'm not seeing where to change the option.
@ptraskaski134 hey if you are using this workflow you should have these nodes in your interface, (Power Lora Loader ➡️ Patch Sage Attention ➡️Model Patch Torch settings) and if you dont have sage attention and triton installed, these settings will not work, so you can just bypass these two nodes ( Patch Sage Attention ➡️Model Patch Torch settings)
for some reason i keep outputting an mix match color video. i tried bypassing sageattention, use different lora, lightx2v on and off other thing but problem still persist
How can we have better visual quality?
You can upscale the video and improve clarity using the Image Smart Sharpen node, but if you're aiming for even better quality, it's worth switching to a different workflow. This one is designed to deliver good results on lower-end systems — it's more about accessibility for users than pushing for top-tier quality.
@Akalabeth Thank you very much for your quick response! I understand completely but you either need to have a big card or go through runpod? Because I have a 4070 super it's good but still limiting! Did you have a workflow to have better quality? I'm still going to try the node for upscale
hi! im downloading the 2 HIGH and LOW wan models, since i am really new to comfyUI where do i place the downloaded models? also which version of Lightx2v i should choose? i have an RTX 3080
The models should be placed in the models/diffusion_models folder. As for LoRAs to speed up generation, I'm still using lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16, but you can also try: - Wan2.2-Lightning-T2V-0928-high and Wan2.2-Lightning-T2V-0928-low (for T2V) - Wan2.2-Lightning_I2V-A14B-4steps-lora_HIGH_fp16 and Wan2.2-Lightning_I2V-A14B-4steps-lora_LOW_fp16 (for I2V)
The LoRAs should be placed in the models/loras folder.
@Akalabeth thanks! i added all the files but noticed something the KSampler(advanced) are highlighted in red after hitting run and nothing would happen, any advice? https://imgur.com/a/j2QBXMu this is how it looks for me
@jaimd You need to install ComfyUI Manager and use it to install the missing custom nodes.
@Akalabeth strange that your response is not here but it went to my notifications XD it actually downlaoded the 4 custom nodes before i even did anything this is strange, why it doesnt do anything as i have the full aplication not the portable one, it already came with the comfyUI manager
@Akalabeth hi! so in the end i donwloaded th incorrect workflow entirely lol i needed the T2V and i went the other way around, now for some reason on the ksampler i get this error PY_SSIZE_T_CLEAN macro must be defined for '#' formats what do you recommend to fix this?
@Akalabeth well after hours of trying my hardest, installing wrong versions and what not i finally did it lol manged to generate a tiger in a subway https://imgur.com/a/c3ZyDZL it took 16 minutes for the very first time since it had to load the models, now that they are loadded im gonna try again and see how much it takes, one thing i will heavily need is to get myself with more ram, 32GB of ram is way too low for this, i had to offload 50GB into pagefile on my SSD and obviously that makes it so slow, this mean that the model alone would take like a minimum of 32GB or i will need to go straight into 128GB ram just to run this smaller model i dont wanna know how much will take to run the actual model, one thing i noticed is that it doesnt have any audio generation, how hard hardware wise is to add audio generation? like sora have for example
@jaimd 8GB of VRAM and 62GB of RAM are quite sufficient to work with this workflow.
I even think that 32GB of RAM might be enough, as long as you’re not running other heavy tasks in parallel.
Generating 5 seconds of video takes approximately 5 minutes in real time - I mentioned this in the description.
As for audio, WAN2.2 doesn’t support it.
Including sound would make the model even larger and the generation time would skyrocket.
If you need video generation with audio, a new model just came out - you can give it a try, though I believe it has even higher system requirements: Ovi Tutorial
good workflow rtx 3080 10gb finish in 300s for 5sec vid (but need 64 system ram) please add WanVaceTovideo so thanks
good workflow good quality with gguf 4Q_K_S
but takes around 8 minutes to get 960x544 with 129 frames(8 sec). which is a little bit too much for me :(
Can I make it faster? I also tried sage attention.
4070 super 12 gb vram 32 gb ram