The power of goon can't never be stopped by hardware limitations! In an attempt to force Wan to run on my pc having less than 16GB of VRAM and extremely slow AI GPU I present with you with this caveman easy workflow.
I M P O R T A N T: I have no idea what I'm doing, if you know how to speed this workflow up and how to improve it, let me know right away! I'm using a fresh portable installation.
Requirements:
I'm not sure about the requirements, but my setup is using max of 8.1GB of VRAM and 16.2GB of RAM - After the first load, which requires nearly 28GB of RAM.
Minimum of 12GB of VRAM - There's no GPU with 10GB or 9GB anyways.
Minimum of 28GB of RAM - It will not work with 16GB or less, trust me.
Features:
Basic LoRa Support for now.
Models Needed:
Model is Rapid Wan AIO 14B SFW - Q4_K_M / Download Link
Model is Rapid Wan AIO 14B NSFW - Q4_K_M / Download Link
Vae is Wan 2.1 Vae - Default / Download Link
Text Encoder Clip is umtt5-xxl-encoder - Q5_K_M / Download Link
Read Me:
It is better to run 6 steps rather than 4 for more details, or use a detailer lora.
Negative chinese prompt is obligatory, otherwise results will be shit, add your own english tags after.
I use "No camera movement" instead of long sentences.
"Detailed movement" and other quality tags seems to work.
It is said that NSFW version is worse than SFW at physics and accurate animation despite understanding more concepts. But SFW looks much worse for me.
Seed "42" gives better results for static light novel-like animations.
You can simply bypass Save Video and LoRa loading if needed.
Performance:
Average of 3-6 minutes for 4 seconds at 800x400 resolution.
Average of 3-6 minutes for 5 seconds at 700x500 resolution.
Average of 3 minutes for 5 seconds at 400x400 resolution.
Average of 56 seconds for 5 seconds at 200x250 resolution.
Average of ~2 minutes at 416x512 resolution at 6 steps, fastest.
Performance Tips:
If you can set up Sage Attention you will have extremely faster performance.
SSD/NVME will be better for RAM file management. Mine is running on HDD, yes.
My RAM modules have scrambled frequencies, expect better performance.
Issues: For weird reasons, if iterations are above 100it/s, gen takes 8-14 minutes. Interrupt and start it again for normal speed.
Best Practices:
No advanced prompt needed, straight describe as if using Flux or SDXL.
describe first frame, describe next action to perform+details+adjectives, repeat, repeat; quality tags, negative tags
[the scene despicts a scared anime girl, she turns to face the camera, she closes her legs and touches her own chest; the camera moves farther away, highres, absurdres, detailed movements]
[the scene despicts a anime girl wearing a blindfold, she removes the blindfold, holding it close to her chest, her eyes are empty; no camera movement, highres, absurdres, detailed movements]
[a 3d model of chibi anime girl is holding two guns, she points the gun around while her hair sways with the wind, she looks at the camera]
Showcase: https://civarchive.com/posts/24529277
Description
First template share.
FAQ
Comments (16)
I'm still a noob at AI videos too. I tried using the workflow as is but with an image I provided. But she ended up taking off her blind fold and then her boobs popped out and she started holding them... XD
Try "The scene shows a girl removing her blindfold from her face with her hands, her breasts suddenly pop out her ripped clothes, she then starts to grope her now exposed breasts with her hands. No camera movement." It works best when you specify everything, spoonfed.
@MacumBX13 I wasn't trying to make her do that. That was the results of trying to use the default positive prompt! lol
Dude, this is a great workflow. When I saw the title, I immediately knew it was exactly what I needed. Can you tell me what kind of Lore I can use in the disabled nodes? Which one would work well with this workflow?
And I want to thank you. My friends are horrified when I send them videos of them getting slapped in the forehead by huge black dicks.
You're simply the best. Your workflow is what started me learning about generation.
your showcase link is dead :-/ About to try your wf now, downloading models aws :)
For me its working, civitai is buggy
Excellent. My set up is 4080 super 32GB RAM. About 10 GB offloaded but generating videos of 5 sec in about 3 minutes. I've tried all versions but I have to come back to v10 and forget mega ones. It's the only one I've tried truly non censored.
How much VRAM you've got?
hey, i love that this worklow and model is so fast, but is it normal, that it looks highly compressed. like should there be a high amount of noise in the output? 6 steps, didn't change anything for me.
I'm not sure myself, Wan is weird asf, specially AIO. But scaling the res up or keeping at 5 steps then sending to a video upscaler solves it
@MacumBX13 i think it's because of the high quantization. usually Q4 isn't too small, but i think in this case it is. i used another workflow, with the 'regular' Q4 high and low models. the image quality is much higher, but i had problems with slow-motion and loras not executing concepts properly...
anyways. your workflow is nice, but i want to add one thing. if you swap the set resolution node, with the wanimagetovideo node, it should fix the sometimes very long processing times. it at least did for me.