@vrgamedevgirl Believe me, you don't want that xD explaining it to me will only give you grey hair...

Rent GPU time with a cloud service. You can you wildly fast GPU's for very little dosh.

My understanding is you need the Native workflow if using a GGUF model. Wrapper uses the regular model and includes a node you can use if you have low VRAM to offload to system RAM. I'm sure there's more than just that though...

Wrapper = beta
Native = production.

Wrapper is like a beta set of nodes that Kijai built. Super powerful but complicated. Most stuff comes out on wrapper first and you get more control over a large number of nodes. Nothing to do with models.

@be20449 That's really helpful to know, thanks!

@be20449 Ah okay. I just haven't really come across Wrapper flows using GGUF model and think I saw somewhere here to use Native flow if using GGUF model.

@vrgamedevgirl ah. That would explain why then. Thanks for clarifying.

I'm curious, what resolution of videos are you generating? I'm working with a 12gig card and 32gb of system memory and I can do 8 seconds of video in around 20 minutes, depending on settings, and with @vrgamedevgirl models that's down to 5 minutes which is just ludicrous honestly but it happens.

@sarashinai those are t2v, im generating at 720p at 9 seconds with sliding windows at 737 frames, so thats 9 frames slt at 40 windows.

I run my own local model, all i need is flux, sdxl and thats it, no loras or nightmares needed, just a story board, my work flow is veo 3 and local, so 720p is my minimum don't go below that, unless im using phantom wan 1.3b for fusion of images controlnet struggles at like batman sitting on a toilet or girl sitting in a chair, and then i use also 720p but only for the transition to feedback into either veo, or local model for certain things, cars i leave to veo, people talking i do local, its all post production.

What you should be asking is what is your goal, give me the best image you have, and i'll do it for free, please make it interesting

lol wut

@markdalias its spelled what, put some effort in on sarcasm

lul wot

Seriously though. Of course you're right to run anything close to the full model needs some serious hardware but thanks to the inventiveness and ingenuity of the community people are managing to make some great content on far less powerful hardware. Some people only want to create under 5 seconds long clips for some fun. Others that are hoping for longer creations are using other creative and artistic ways to achieve this by either creative post generation editing or by tech work arounds. When people use and enjoy horribly compressed images online and social media platforms. A professional would never "work" with jpegs if they access to the original RAW files and the hardware to process them, but it doesn't stop the non experts from enjoying viewing the images in compressed jpeg format and even messing around with compressed formats in creative ways.

What the community are doing with WAN with pruning and quantization, fine-tuning, baking-in and optimisations, like we are seeing on display here, open the model up to more people to be able to be "creative" without the barrier of exclusive levels of hardware requirements. To tell people to "not bother" seems a strange stance to take on creative endeavours.

Given the ever changing landscape then WAN may be obsolete in 6-12 months. But in the mean time we've had some fun with it, and all learned a lot about the basics and sometimes intricacies of making AI work on a local machine.... to cram a half decent model on under half decent hardware and run it to 99% capacity.

I'm managing really nice outputs at 16:9 720p on a 4060ti 16gb and a CPU that's 7 years old now. And I'm enjoying the process of creating and learning how to use all this.

I'm very glad you've got the hardware to do what you're doing and actually I'd love to see what you're producing if you've got it published anywhere? It sounds like you're probably producing some great stuff.

anyway, if anything I'd be encouraging people to get involved and work with whatever they've got. Demand and experimentation en-masse is what drives innovation and very importantly, what drives down or knocks down price based entry barriers.

@vrgamedevgirl +4

1024x576 and 720p (1280x720) are different resolutions, with 720p being the higher resolution. 720p, often referred to as HD, has more pixels (both horizontally and vertically) than 1024x576, resulting in a sharper and more detailed image.

Let me know when it gets to 1080p at 10 seconds near 30 fps

LOL Ur funny

@vrgamedevgirl while I believe having fun is important the model wan14b cannot fit entirely on a gpu that's consumer grade, there's a reason it's quantised and CPU offset, however that's not to say a ada 6000 can run it effectively you would need a rtx pro 6000 to really benefit with most tencent and Alibaba models, given China's embargo getting a cluster of rtx pro let alone h100 has set these monolithic companies back 6 months, for good reason it's banned.

If upscaling and pushing it to 30fps is necessary then 48gig for the wan14b was just kind words.

It's important people know so they know the difference between fun and production.

@reponum13625 In a matter of weeks what people can do on consumer grade hardware has gone from god awful glitchy graphics with terrible consistency, continuity and weird inhuman movements to being able to produce linked and consistent clips of any length with character continuity AND with cinematic qualities. Is it production level yet? no. Are people "producing" content that people are viewing and enjoying, and in some cases paying money for? Yes.

Again, the more people that get involved at their "consumer" level inspire more and faster innovations in the technology that enables even more people to not be priced out of being creative.

The arts have long been the playground of the rich as they have the time and money to be able to do it at leisure with all the best equipment and staff available to them on a whim. Disruptive technology breaking those kind of based art entry requirement structures should be applauded and encouraged by all.

If you are not getting pleasing results on 16gb Vram and 64GB ram then you haven't bothered to learn enough about the constant innovations that are being championed almost daily within the community.

I can use sliding windows to produce full 720p (1280x720) at any length. It's slow yes, but its 6-8x faster than it was just a few weeks ago for far higher quality results. If I'm clever with my test runs and previews then i can create really enjoyable results.

As i said before, some of the most viewed and enjoyed images and even films were produced on "non professional" hardware and actually forced a deeper level of creativity from their creators.

You are giving off a strange passive exclusivist vibe, and i genuinely can't understand why anyone involved in this arena would.

@vrgamedevgirl  4080 16gb, Model 720P, base_precision fp16, Torch off, attention mode spda, 1024 576. 21 frame took 44 minutes...what im doing wrong?

You are clearly using the default workflow. I'm sure this is a common issue with newcomers so I don't blame you for not knowing all this and it's not your fault. 16gbvram here and can run 720p q8 gguf model with full 1280 resolution, 81 length, with 10 minute generation time. I highly encourage you to incorporate optimizations in your workflow.

Optimizations like WanVideoBlockSwap helps with OOM errors allowing you to run higher res and bigger models.

Causvid lora lets you go lower steps which increases gen speed. Teacache node can 2x your generation speed (at the cost of quality).

Sageattention improves performance nicely.

You'd be surprised how good quality even at lower res's + upscaling looks, giving even better generation times as low as 200-300 seconds at 81 length. I thought the same thing as you when I first started too : )

runs great with my 3080ti 12GB, maybe you need to tweak some things

You can just chain them, worked well for me with native workflow... (Actually I just inserted the FusionX LoRA on my pre-existing workflow since it was by and large the same as the posted workflows, with some changess to the saving of the files and other tweaks.)

whats your workflow? the wrapper can do it just fine but depends on your hardware i think

WAN_Native_I2V_FusionX

@vrgamedevgirl I have the same issue with that, I downloaded yesterday the GGUF one

I have 16gbvram too and had to switch to a gguf model and can do 1024p around like 8 minutes but I do low res first then upscale later for even faster speeds. You can also add optimizations to your workflow.

First, make sure you're using a GGUF workflow. I couldn't find his so I'm using this one instead - https://civitai.com/models/1622023/causvid-2-sampler-workflow-for-wan-480p720p-i2v

Download a GGUF version of Wan2.1 model of your choice for better performance.

P.S. the Causvid LORA is replaceable with FusionX LORA bc it's built-in apparently so choose one or the other, download found on site.

Here are I2V 720p gguf models (higher the Q number in the name, the higher the quality) - https://huggingface.co/city96/Wan2.1-I2V-14B-720P-gguf/tree/main

Make sure to also download the missing VAE, clip vision, and clip

https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/vae/wan_2.1_vae.safetensors

https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/clip_vision/clip_vision_h.safetensors

https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/blob/main/split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors

Settings:

Make sure you're running 4-10 steps when using fusionx lora, 16fps, 81 length max. You'll be able to do 1024p with 720p q8. Also download 480p gguf to compare quality just in case bc some reason 480p model slightly looks better for me prolly bc im using lower res's? idk tho.

Optimization nodes:

Double click the background and add the WanVideoBlockSwap node (helps with OOM error and uses less VRAM) to the workflow and set it to 40 (since you want high res) and everything to "TRUE".

OPTIONAL:

The next node you can add is called WanVideo TeaCache (2X generation speed boost at the cost of QUALITY.) Connect it BEFORE the sampler node. If quality bad, set the thresh value to like 0.1400. I'm currently trying to figure out best settings for it.

That's it. If you want sageattention for windows for slightly more speed, that's a whole other story and it can bork your comfyui if installed incorrectly... I hope I didn't miss anything else

GOOD LUCK

@bhopping Thank you very much! I'll try to do so and reply after! Thanks again for fast anwering

@Pirate_rus No prob, I think I did forget to mention where you put the WanVideoBlockSwap node, it just connects right after your Diffusion Model Loader (aka the first node)

@bhopping I looked at the diagram. It's difficult for me so far. Replace the links to the models, no problem, set the number of steps, frames, and length. In this scheme, I did not find where the step is set. If you could share your working schema, our configuration is roughly similar, it would be much easier. If it's somehow possible here, of course.

@Pirate_rus I'll upload the workflow soon

searching the discord, looks like you have to change the base_precision of the WanVideo Model Loader to fp16 instead of fp16_fast. Not sure why the workflow is set up in this broken state out of the box but easy enough to fix.

@vrgamedevgirl I tried using Gemini for 30 minutes to install the version of pytorch it was complaining about needing and Gemini concluded that it wasn't publicly available. I wouldn't call someone not using a dev branch having an "outdated" pytorch version. How do I install it?

Also, I don't see anything about this pytorch version in the notes of the workflow. It only talks about torch compile, which isn't an issue I saw.

@vrgamedevgirl Well I would like to use fp16_fast. I already knew how to use fp16. Can you tell me how you installed the 2.7.0.dev version?

vrgamedevgirl Yikes. That amount of effort definitely isn't worth it to me. Thanks for the heads up.

@vrgamedevgirl Oh I mean, the blond women we see in the 'Phantom-Wrapper-FusionX_W' workflow, isn't she similar all the time? Or it's just prompting? Thanks for your answer and again, a huge thank for your insane work and contribution to the community!!

@kdashhoward Phantom is trained for character-subject consistency. You can load up to 4 images and it does a pretty ok job at retaining character or object. No lora training needed

There is the node WanImageToVideo. The field "length" is the duration of the video. You have to multiply the frame rate of the Video Combine node (e.g. 16) with the desired length (e.g. 5 seconds). Then enter 80 as length. It will be changed automatically to 81. That´s it