Anonymous
6/20/2025, 8:45:51 AM
No.105648757
I have concluded that after wasting an entire night trying to get wan to follow prompts whilst using any so called speed up loras. That all speed up loras are fucking trash, why is that? Its probably because wan is very good at following prompts but it requires cfg, real cfg other wise it won't produce the video you prompt.
Phantom model works much better without anything like causvid(any version) fusionx what ever its called and lightx2v. A cfg of 4 and 20 steps and its working with loras correctly and not mangling everything, at least that is what it seems, i'll have to do a few gens. Yes it takes longer at 20 steps but it follows my prompt, if it does not follow my prompt then that has been a waste of electricity and time. optimalsteps is still the best because it does not mess up the video when using loras provides 40 steps quality at 20 steps, does not change reference image face.
Also
--lowvram
--reserve-vram 1.0
+ native comfyui methods
is still the best method of running wan on a 12GB card, everything else is horrible, i oom on just 5 frames using phantom model Q8 if I try to use the offloading node that every retard recommends. But if I run my comfyui with the options above, i can do 1280 x 720 videos 41 frames or more, why? I'll explain, comfyui offloads correctly and more efficiently natively, but you have to reserve some vram, because if you don't it tries to use all vram on the card before it will then use the system ram. Then it will oom when something else such as preview nodes or anything else on the system tries to use the vram.
So just use native and set those options and it will work on lower vram cards, also increase your virtual ram. All other methods are slow as fuck, also those memory saving nodes? Fuck them off they are retarded and don't do shit.
Phantom model works much better without anything like causvid(any version) fusionx what ever its called and lightx2v. A cfg of 4 and 20 steps and its working with loras correctly and not mangling everything, at least that is what it seems, i'll have to do a few gens. Yes it takes longer at 20 steps but it follows my prompt, if it does not follow my prompt then that has been a waste of electricity and time. optimalsteps is still the best because it does not mess up the video when using loras provides 40 steps quality at 20 steps, does not change reference image face.
Also
--lowvram
--reserve-vram 1.0
+ native comfyui methods
is still the best method of running wan on a 12GB card, everything else is horrible, i oom on just 5 frames using phantom model Q8 if I try to use the offloading node that every retard recommends. But if I run my comfyui with the options above, i can do 1280 x 720 videos 41 frames or more, why? I'll explain, comfyui offloads correctly and more efficiently natively, but you have to reserve some vram, because if you don't it tries to use all vram on the card before it will then use the system ram. Then it will oom when something else such as preview nodes or anything else on the system tries to use the vram.
So just use native and set those options and it will work on lower vram cards, also increase your virtual ram. All other methods are slow as fuck, also those memory saving nodes? Fuck them off they are retarded and don't do shit.