Video Generation on a Laptop
Hello!
This workflow utilizes a few custom nodes from Kijai and other sources to ensure smooth performance on an RTX 3050 Laptop Edition with just 4GB of VRAM. It's optimized to improve generation length, visual quality, and overall functionality.
🧠 Workflow Info
This is several ComfyUI workflow capable of running:
2.0-ALL -- Includes all workflows:
Wan2.1 T2V
Wan2.1 I2V
Wan2.1 Vace
Wan2.1 First Frame Last Frame
Funcontrol (experimental)
Funcameraimage (experimental)
Coming soon: Inpainting experimentals get updated
🚀 Results (Performance)
*to be updated
🎥 Video Explainer (Vace edition):
🎥 Installation Guide (V1.8):
📦 DOWNLOAD SECTION
⚙️ Nodes Used (Install via ComfyUI Manager or links below)
Note: rgthree Only needed for Stack Lora Loader
📦 Model Downloads
*these are conversions from the original models to run on less VRAM.
most versions
Faster/Better quants for i2v
fun,inpainting,T2V,Vace
fun-control
🔗 WAN2.1 Fun-Camera-control 14B GGUF
fun-Camera-Control
All these GGUF conversions are done by:
https://huggingface.co/calcuis
https://huggingface.co/QuantStack
*If you cant find the model you are looking for check out there profiles!
🧩 Additional Required Files (Do not downlaod from Model Downloads)
📥 What to Download & How to Use It
✅ Quantization Tips:
Q_5 – 🔥 Best balance of speed and quality
Q_3_K_M – Fast and fairly accurate
Q_2_K – Usable, but with some quality loss
1.3B models – ⚡ Super fast, lower detail (good for testing)
14B models – 🎯 High quality, slower and VRAM-heavy
Reminder: Lower "Q" = faster and less VRAM, but lower quality
Higher "Q" = better quality, but more VRAM and slower speed
🧩 Model Types & What They Do
Wan Video – Generates video from a text prompt (Text-to-Video)
Wan VACE – Generates video from a single image (Image-to-Video)
Wan2.1 Fun Control – Adds control inputs like depth, pose, or edges for guided video generation
Wan2.1 Fun Camera – Simulates camera movements (zoom, pan, etc.) for dynamic video from static input
Wan2.1 Fun InP – Allows video inpainting (fix or edit specific regions in video frames)
First–Last Frame – Generates a video by interpolating between a start and end image
📂 File Placement Guide
All WAN model
.gguffiles →
Place them in yourComfyUI/models/diffusion_models/folder⚠️ Always check the model's download page for instructions —
Converted models often list exact folder structure or dependencies
🔗 Helpful Sources:
Installing Triton: https://www.patreon.com/posts/easy-guide-sage-124253103
Common Errors: https://civarchive.com/articles/17240
Reddit Threads:
https://www.reddit.com/r/StableDiffusion/comments/1j1r791/wan_21_comfyui_prompting_tips https://civarchive.com/articles/17240
https://www.reddit.com/r/comfyui/comments/1j1ieqd/going_to_do_a_detailed_wan_guide_post_including
🚀 Performance Tips
To improve speed further, use:
✅ Xformer
✅ Sage Attention
✅ Triton
✅ Adjust internal settings for optimization
If you have any questions or need help, feel free to reach out!
Hope this helps you generate realistic AI video with just a laptop 🙌
Description
-New Teacache feature
-Compile Torch feature
-Different GGUF Loader
-Notes
-Organised
FAQ
Comments (121)
My
ApplyRifleXRoPE_WanVideo
WanVideoTeaCacheKJ
TorchCompileModelWanVideo
are not working.I did the -requirements thing at the custom nodes and it fixed the vae loader only
hm do you have the comfy ui manager installed?
You should have ComfyUi-KJNodes updated to nightly 1.0.6
I had to uninstall the KJNodes and reinstall for those nodes to be detected properly.
boys what do u think
a laptop rtx 2050 with 4gb of vram and 12gb of ram
Yes defenetly works i would chose the q2_k version from city96 and only like 25frames 30 steps
@The_frizzy1 i wish i knew what stable diffusion in general meant really
im really bad with code but i like generating cool images
especially with fnaf animatronics and ben 10 aliens
@xmattar hahah thats nice im also still more noob than not but the community really helps with learning and understanding.
@The_frizzy1 the thing is the community will use these models for porn
i want to use it to create an online series
it will be bad and everything but the only purpose is to pitch a pilot since animation budgets cost a whopping 1.2 million for a pilot episode and for something that might not get green-lit by a studio
a concept pilot will do the job, especially since im just a college student with a uni laptop
glad u were able to help with my 4gb vram potato!
Tiled KSampler
backend='inductor' raised: RuntimeError: Cannot find a working triton installation. Either the package is not installed or it is too old. More information on installing Triton can be found at https://github.com/openai/triton 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
Set everything in TorchCompileModelWanVideo to False. This helped me
@kalamnuzed291 Yeah I have had some issues with those settings aswell.
For me atleast the most basic settings when you put the node in the workflow always works hope that helps.
I get error: mat1 and mat2 shapes cannot be multiplied (77x768 and 4096x5120)
e: okay, it really only wants the specific files listed in this workflow. It won't work with the other models found elsewhere for wan.. at least some of them.
@kalamnuzed291 i dont see , i only have this in torchcompile ......https://prnt.sc/l8g5wuA5ui20
I got this error instead...
BNK_TiledKSampler
The expanded size of the tensor (104) must match the existing size (64) at non-singleton dimension 4. Target sizes: [1, 16, 4, 60, 104]. Tensor sizes: [16, 4, 4, 64]
And my tiledksampler node is not seeing "tiling", but seeing "tile_height", "tile_width", and "tiling_strategy"...what have I set wrongly?
@pza Are you sure you’re using the correct node? I updated it to check but the tiledksampler is the same but the tiledksamplerprovider does have tilling_strategy and the other settings.
@ziriuss1 you can turn off compile transformer blocks see if that helps but the settings look fine to me.
@The_frizzy1 well, turns out they are different nodes. ur workflow is actully using the tiledKsampler by FlyingFireCo. While the link in ur post is by BlenderNeko
@cigg Thank you I’m gonnna correct that I didn’t notice.
why would you upload a zipped .png of your workflow instead of the .json?
it's the same thing, ComfyUI can read the .png
I cant open workflow it says custom missing node wan image2video, please help?
Do you have the comfyui manager than install the missing nodes or install them directly via git in the text I have them listed.
Just update the gguf node and update the comfyui and the nodes will work.
Hi, I tried version 1.1 and it works fine, but very slow. while version 1.5 generates distorted contents with a huge mismatch from the original image. What might be the problem?
What settings are you using for generation?
Q5_K_M, umt5_xxl_fp8_e4m3fn_scaled, 480*960, length:33, all other parameters the same as what you shared in your workflow
@rockiecxh maybe set the denoise higher to 0,95-1, change up the prompt, lower end percentage for tea ache.
@The_frizzy1 Thanks for your tips, it works great now! I changed the denoise to 0.95 and end percent to 0.9, got a very good result!
@rockiecxh my pleasure!
@rockiecxh So that's the reason. I'll go test it
Unexpected architecture type in GGUF file, expected one of flux, sd1, sdxl, t5encoder but got 'wan'
It sounds like you downloaded the text encoder from the model repository it won’t work you need to use the split files text encoder.
!!! Exception during processing !!! Unexpected architecture type in GGUF file, expected one of flux, sd1, sdxl, t5encoder but got 'pig'
got this error
using wan2.1-tv2-14b-q6_k.gguf
Make sure you download the text encoder from the split files and you have newest update of all the nodes the model loader should be lowercase gguf others are not supported
@The_frizzy1 The diffusion model was not working from the link provided but i downloaded the model from
https://huggingface.co/city96/Wan2.1-I2V-14B-480P-gguf/tree/main
now the model is being loaded but i am getting some other errors
@kavirap_23 Are you using the text encoder from the split files?
@The_frizzy1 Yes
If you still get the error, try to use normal GGUF loader
@jackcreatordesign782 Yeah that might help I had it the other way around where the loader I was using before didn't work. Im not sure what the actual issue is..
@The_frizzy1 I have no idea what it is either. I'm guessing it's something to do with the checkpoint. Other checkpoint works
I have 6 gb of VRAM so what model should i use ?
Q5_k_m should work pretty good
@The_frizzy1 Thanks
getting error "Found NVIDIA GeForce GTX 1070 which is too old to be supported by the triton GPU compiler, which is used as the backend. Triton only supports devices of CUDA Capability >= 7.0, but your device is of CUDA capability 6.1"
EDIT: setting "compile_transformer_blocks" to false in torch node fixes this.
Yeah not everything on the workflow is supported by everything, it still works without torch compiling but it’s not as fast.
thank you i had the same issue
"EDIT: setting "compile_transformer_blocks" to false in torch node fixes this."
where should I place this ?
@xirecentxz635 Do you mean like what node? The "torch wan compiler"
@xirecentxz635 there is only one node in the workflow that has this property (and initially its "true")
How can I change sampler and scheduler? and what are default sampler and scheduler in this workflow? Thanks!
well just change it in the node.. The default is not what you should be using I would suggest uni_pc_bh - Beta denois 1
Probé su modelo en mi laptop de las siguientes características 32GB RAM y RTX 4050 6GB VRAM.
Llevo utilizando Confyui dos días así que diría que realmente no se casi nada pero hice lo que pude.
Utilice su workflow haciendo los siguientes cambios:
- me salte el nodo (TorchCompileModelWanVideo) por que me causaba error la carencia de TRITON en mi PC, estoy utilizando ConfyUI sobre PINOKIO, por lo que no encontré forma de instalarlo.
- En WanImageToVideo cambie el vaor de length por 29.
- En el KSampler cambie los steps a 40 y el denoise 0.9
Resultado: me tomo 15.43, considere que los cambios que hice produjeron que mi tiempo sea mayor, considerando que TRITON es aparentemente un acelerador con el cual no cuento y por otro lado la diferencia que hay entre una RTX 3050 de 4gb y una RTX 4050 de 6GB no considero que sea la gran cosa en este apartado, por las pruebas que hice un día antes con otro workflow y otros valores en mi pc que cuenta con una RTX 2060 SUPER 8GB VRAM parece obtener mejores tiempos comparado con mi RTX 4050 que es 2 generaciones por delante.
Tiled KSampler
backend='inductor' raised:
RuntimeError: Cannot find a working triton installation. Either the package is not installed or it is too old. More information on installing Triton can be found at https://github.com/openai/triton
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
What do I do to fix this error?
You can try installing triton it improves the speed or you can just get rid of the TorchcompileModelwanVideo Node
@The_frizzy1 Does the lack of it only affect speed, or quality of the result as well?
For windows I recommend using pip install triton-windows from this github repo - everything else failed.
Also triton installation depends on your graphics card. I have a mid-high 20 series so I can't run it on windows, for example.
Hello, Master,
Why my generation result bad when using i2v models while the t2v models work fine using the same workflow.(Just change the gguf model, no other changes)~~?
I would change my settings in the sampler use 1 denoise and uni_pc_bh2 - Beta
@The_frizzy1 Yeah，much better now and THANKS~~
@Diaoshini8280128 yeah im gonna update it but im trying to make a video first for anyone having issues with torch,triton etc
when download zip file there is only image inside, not a json file with workflow? maybe im miss something ?
sorry, new on comfyui... all is ok and fixed
No worries, for everyone you just drag the PNG onto the comfy ui window
@The_frizzy1 thanx, i am fixed that... and is here chance to implement here ComfyUI-WanVideoStartEndFrames ?? to use start and end frame to generate i2v video, ?
@xxxtembel I took a look at it but it uses the actual wrapper so for now youre not gonna be able to do it with low vram if you have enough you could implement it.
@The_frizzy1 im try run that, but i am a fish... if you implement that to this template, i am think, it will be run on 12 GB + cards... i am just not good enough, to implement that to you workflow ))
Missing ApplyRifleXRoPE_WanVideo even it is installed in my custom nodes
same here, is there any solution for it? thanks a lot
Hey, i'm getting the error:
UnetLoaderGGUFAdvanced
Unknown model architecture!
Does anyone know why this is?
I'm using the q5_k_m version with bfloat16.
Yeah me too.
same i dont know what to do about it
Me too. Help please
Seems the model is bad. try downloading from https://huggingface.co/city96/Wan2.1-I2V-14B-480P-gguf/tree/main I tried q3 k m or s and managed to get it wroking
Whooply yes I have noticed some of the models are incompatible with some of the nodes. I always got it to work again downloading a differnert model like this. I have updated the download and the unetloader node I dont get this issue anymore, hope that helps.
下载的文件只有图片没有工作流的JSON
Hi, I faced the following error. How do I figure it out? :
mat1 and mat2 shapes cannot be multiplied (77x768 and 4096x5120)
https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files
我也遇到这个问题，需要去下载text_encoders中的umt5_xxl_fp8_e4m3fn_scaled.safetensors，这是特殊的
How do I install Triton? I use ComfyUI Desktop version, should I use the portable version instead to install it?
I have had issues aswell I don't know what specific issues you have but this video helped me install it make sure to just really follow every step it includes sageattn aswell: https://www.youtube.com/watch?v=DigvHsn_Qrw
Could someone help me? I don't know what folder I have to put it in or where that folder is. It's for making videos. wanAIWan21VideoModelSafetensors_kijaiWan21I2V14B480P. That's the checkpoint.
And it's in its own folder, but it keeps giving me an error. Can someone enlighten me?
I have already deleted everything, because there was a problem with installing Triton, but I remember that the model was in (comfyui > models > ) "diffusion_models", umt5 and clip vision in "clip", and vae in "vae". If something is in the right place, then after refreshing the browser tab, it can be selected in the corresponding node. Well, if you download models \ encoders through the comfy manager, they automatically get to the right folders, although not everything can be downloaded that way.
No offense, but this is all except "Simple" ....so those gguf models are like different from the "originals"? I received: "This is a different model architecture"
well what can I say simple in the open source AI world is relative...
and yes they are different there quantizations of the original.
LoaderGGUF，在什么文件夹 我的模型放在/ComfyUI/models/diffusion_models，加载不出来是为什么？
try checkpoints or unet file
已经解决，从新下载模型放进去就可以了
refresh your comfyui browser.
I would help but I don't understand it.
are you using a t2v model for i2v?
shouldn't you be using i2v?
I think it might still be on that setting but ill make a update real soon just have to do my work aswell...
Yo you need to update your listing. You're linking to models that dont work with your workflow.
And the model you're using in the example immediately skyrockets to 16gb of vram usage.. how do you fit 16gbvram into a 4gb card as you claim? Makes no sense..
I do need to make an update but I have been super busy. I don't know what you're doing but I am using the linked models and also only have 4gb of vram I don't know why it spikes for you but its not a claim its the only graphics card I have...
what version of python do you use i cant run this due to triton compatbility
who the heck to setting denoise to lower than 0.8? how are you going to generate exactly prompt to exact video you want? I remind you go denoise == 1 !
Yeah you are right, I was trying something but I will make an update very soon and have it set right from the start.
I'm running a 3050 RTX with 4GB RAM and it doesn't seem to like it. Seems like too much is going into the cuda 0 and causing it to die.
I mean if you get vram overload all you can do is lower the quality lower the frames
How about another workflow for optimizing the speed on 24 VRAM?
you know I would but I don't have that much vram and don't really want to make a collab notebook for this.
为什么我无法在加载clip的时候选择wan？
RIFE VFI is missing. Which repository can I get it from?
Might be this rep? https://github.com/Fannovel16/ComfyUI-Frame-Interpolation
4x_foolhardy_Remacri.pth
I don't have this enlarged model. Where can I download it
Which folder should it be placed in
What is this? I downloaded an... image? That's not a workflow.
You can Put the image into comfy UI and then you have the workflow…
If a person doesn't know even something as simple as this, doesn't know the basics, then he won't be able
to figure out this workflow, he won't be able to understand how it works. Only blind use.
@Rahim33 Calm down my dude. We all start somewhere
@The_frizzy1 Thank you for kindly replying to the question. It helped me too.
I had no idea that you could load a workflow from an image. I would assume the metadata is involved.
Thank you very much
Its always the dumbest that are the loudest.
when you hit a container, it's the one that empty makes the loudest noises.
you shouldn't write sarcastic riddles it will flew way above his head.
Reinstalled comfy 64 times, still triton error on Tiled Kampler, If triton gets fixed, it gives torch error.
Check this one: https://www.patreon.com/posts/easy-guide-sage-124253103 :)
Hello, is there anything for 2GB VRAM? gtx960
Haha maybe if you go really small 1 frame load your clip in your cpu and only text to video or image in this case :)
Hi i think yes you can , try this version of wan that work with 4 steps only.
it has t2v and i2v.
Evados/DiffSynth-Studio-Lora-Wan2.1-ComfyUI · Hugging Face
@hichuicomfy I tried it the normal way and I can create a 256x256 video of 2 seconds, at 16 frames, if I increase it a little it gives an error Cuba due to lack of memory, Thanks for the reply.
@syphonfilterargAI now we have causvid lora for wan1.3 t2v and i2v with great quality in 2 3 steps
Wan21_CausVid_bidirect2_T2V_1_3B_lora_rank32.safetensors · Kijai/WanVideo_comfy at main
not working, monitor black screen :), I'm using zluda, 7800 xt and 32gb ram but still no support at all, it's so hard, I don't know what else to do