I luckily made a workflow which can generate beautiful high quality videos with wan2-2 with only 4 steps so decided to share with community
Description
FAQ
Comments (61)
What's the vram requirements?
To fit everything in the VRAM it's 16 GB.
Cool. Thanks. I will be checking this out!
i think even with 8gb you can make it work if you download GGUF version.
after high noise step the first model clears out from VRAM.
so if you download a quantized version of the high and low noise less than 8gb you should be fine.
so even if High noise model is 8gb and low noise 8gb you still must be fine.
Even bigger than 8 GB still you are fine cause in GGUF version you can even offload to your ram or hard drive and use them as virtual ram.
in my tests using virtual ram did not have huge difference. It was like 3 or 4 second extra which who cares?
bpcmaster366 Thanks for that. I've got 16GB, but that info is good to have. Probably going to give this one a go this weekend.
Nice man, so far this is the best wan 2.2 i2v workflow i've got. Using 14B fp16, 121 frame, 720p, only took around 4mins.
Happy that you liked it :D
Do you have 24 GB VRAM?
ai839 yea, 24vram, 64ram
bpcmaster366 the only problem is the result does'nt follow the prompt like a basic workflow, did i miss something or is there anything i should do?
WildCentaur This workflow was a mock up just to show how it can work with as little as 4 steps.
i might start a youtube channel since i see so much need in community and i will share there my ultimate workflow which i generate uptil 20 seconds video.
This is incredible. SO fast and great results. Has anyone had any luck applying LORAs to this WF? Seems to not take.
I had no issue adding loras to the workflow.
What settings are you using, because I keep getting blurry results ? :/ Can you post a screenshot ?
It's great, but I think it needs a bit more low-noise step. Even with higher resolution, there's a lack of detail and blurriness in the eyes, head, and other fast-moving objects.
Also, may I ask why you use light2 t2v lora when there are light2 lora i2v?
Agreed.
I do 10 step 5 high 5 low for production
but for testing i do 4 step
Sorry for the nub question, where can i get the models?
my previous comment didn't passed,
here: https://huggingface.co/bullerwins/Wan2.2-I2V-A14B-GGUF/tree/main
any way to use WAN 2.2+ controlnet? VACE model hasn't been released yet unfortunately
Not yet unfortunately
Thanks to the author! Thanks to Kijai. Amazing things! 4 min for HQ video.
You are welcome man. I love Kijai too
and i love Wan team and all people which make these possible
wow amazing and simple , thanks a lot for sharing !
It's not working here on a 4080 Super. It gets to 55% and then says "reconnecting" forever.
If it says reconnecting it means your comfyui server stopped due to low vram.
you can change the model with gguf file format and use this node
UnetLoaderGGUFDisTorchMultiGPU
search for this node in comfyui manager > custom node manager > ComfyUI-MultiGPU
then add this node to your workflow > UnetLoaderGGUFDisTorchMultiGPU
change the model to gguf format and add some virtual vram to your workflow
bpcmaster366 I installed the ComfyUI-MultiGPU node, but the UnetLoaderGGUFDisTorchMultiGPU isn't listed in the multiGPU submenu for adding nodes.
Ok, I fixed that problem by installing Comfy-GGUF, but now I get the following:
KSamplerAdvanced
AttributeError: 'UserDefinedObjectVariable' object has no attribute 'proxy' from user code: File "D:\pinokio\api\comfy.git\app\comfy\ops.py", line 58, in torch_dynamo_resume_in_cast_bias_weight_at_51 has_function = len(s.weight_function) > 0 Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information You can suppress this exception and fall back to eager by setting: import torch._dynamo torch._dynamo.config.suppress_errors = True
Silly me, I saw a workflow that "runs on a 4060", which has 12gb VRAM, and thought, "I have a 4080 Super, and that has 16gb VRAM, so it should run even better." LOL, joke's on me.
phexitol well, it runs on my 4070TiS 16 gb pretty good. Try ask some AI about your errors.
Or, download https://www.comfy.org/download desktop comfy, install it, take this workflow, put inside new confy, download missing nodes trough already built in manager, correct paths to models and lora - and here you go.
It's always hard to tell something about different custom comfy builds...
forfreelsd368 Thanks. I've been doing some research, and beginning to think Pinokio might be the problem. I'll install ComfyUI standalone and see if I can get better results.
When i try to download i get: suspicious download blocked
There is nothing inside this file is suspicious. Zip file and json file and mp4 file can not contain virus you can ask this from chatgpt.
there is no hidden files also inside the zip.
So it must be false alarm.
bpcmaster366 Is there a way you can provide/upload the json seperate?
Contents flagged as suspicious by Firefox, this may just mean that it is a newly added file that has not been downloaded frequently.
Would just love the JSON file only..
Just grab one of the videos - save it - then drag it into comfyui. It will load the workflow.
hdean wish every other video worked that way.
forfreelsd368 9x outta 10 usually
Hi Master, how to add lora in this workflow,where do i link it, many thanks,
Hi buddy, my question is that the lora I added seems don't work, I try to add individual lora, but dont know where to link the node,
i guess link to the load lora node behind the Power Lora Loader (rgthree) node
"hi master" is crazy
It does work really fast, but also very blurry.
Actually I didn't encounter this problem yesterday. I've got no idea why the frames are all messed up. No error pops up during the generation process.
I'm now using the rtx3080ti 12G
Did you manage to fix the blur issue?
The video gets blurred after a couple of frames. No idea how to fix it
Power LORA Loader (rgthree)
lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16.safetensors
FastWan_T2V_14B_480p_lora_rank_128_bf16.safetensors
Do you have these two LoRAs installed?
i adjusted workflow for i2v and used GGUF and it works ok on 8GB RTX 3070. About 6 minutes for 480x832 4 steps
Please share your workflow Sir and what about the character consistency and generation quality is it good?
Can you share yo workflow?
can you please share your workflow
The colors change, everything is blurry, It works a little faster than the classic build with this lore. It's all somehow strange.
I2V is blurry as hell and not at all as the sample
T2V Error : AttributeError: module 'triton' has no attribute '__version__'
:(
你需要triton环境安装，也许你可以在youtube或者deepseek了解triton的安装，这个问题我也遇到过，解决起来比较繁琐。
Remove Triton from the workflow and it should work
Or instal Triton, its a pain on Windows and I couldn't make it work... sadly
@Setian91 请问你是否懂中文，在中文视频网站bilibili有Triton的安装教学，如果你不懂中文的话，那的确十分遗憾。
@2193738084205 sadly no, I used a translator
I did manage to make Triton work outside ComfyUI sadly not on my ComfyUI install (I use the desktop installation)
@Setian91 桌面安装的COMFYUI或者说官方版本推出的COMFYUI我并不太清楚，CN地区绝大部分都是使用的视频博主免费公开的COMFYUI整合包，所以CN地区的COMFYUI用起来很方便。也许是这个原因导致你的triton无法在官方版本的COMFYUI中运行。
@2193738084205 its basically running as a all in one app, maybe similar to the solution you are using, however in this version I can't use another Cuda, Phyton or Pip version since the app wants to return to the stable version.
I might have to wait for a update, one which will consider using Triton also for windows. Currently using other optimisation nodes which makes the result good and the speed a lot better. Just not as good as the Triton examples I saw. Might have to wait then or use Linux.