Required a clean install for me, but its definitely working. Very nice speed increase.
Required a clean install for me, but its definitely working. Very nice speed increase.
TThe speedup for later iterations is phenomenal. Makes increasing step on the generations more attractive.
TKLING 1.6 Pro mode now generates 5s@30fps img2vid in 60 seconds.
How fast is this model, insane compared to WAN 2.1
How fast is this model, insane compared to WAN 2.1
Teven assuming they use h200's on backend
Xgetting this error trying to run 720p image to video
Xthis is the pod im runing it in
Xis multi-gpu supported for wan on runpod?
JAnyone has a robust workflow/settings to fix fingers and toes?
When I try editing/masking or upscaling it may make things slightly better, but doesn't consistently make a big difference.
Take this image for example. Anyone who could tell me how they'd fix the toes?
When I try editing/masking or upscaling it may make things slightly better, but doesn't consistently make a big difference.
Take this image for example. Anyone who could tell me how they'd fix the toes?
JThe best I could do is this, but toes are still not great, and it took way too long time
Jthe key here was using "Init image" with 0 creativity and the prompt was just "<segment:feet,0.9,0.3>feet".
But yea not the best and don't have any control over the mask.
Manually masking and prompting gave worse results for me
But yea not the best and don't have any control over the mask.
Manually masking and prompting gave worse results for me
JI tried the Flux fill model as well but didn't get better results than my own fine-tuned model
Twith fingers - you just have to get lucky, or inpaint img2img.
Toes - simply doesn't work with flux so just avoid them
Toes - simply doesn't work with flux so just avoid them
Jah shit okay
TIt seems they were purposely censored just as some other anatomy parts during base model training
JGot it
Jsome other parts
XI'm starting to think that it might be smarter for me to do wan inside of massed compute because even if i have to download wan every time i will use it inside of massedcompute it will still be fast and cheaper the runpod 
Jprices are almost the same?
Xmaybe but runpod has so many issue
Xfor now i will stick with runpod since i already have money in there but i think soon i will try to move to massedcommput
JBeen using Runpod a lot of past few days, works (almost) as well as Massedcompute IMO, but the templates just makes everything 20x easier to setup
Jhmm yea dont know about that
Jbut from my testing the RTX 4000 Ada and 2000 are kinda shit
JA40 is 2x faster on Runpod from my tests
Ni'm seeing about a 25% increase in speed in the latest update to the WAN app on my 3090 
XHmm i think I might have to try it out
X@Dr. Furkan Gözükara my 720p image to video took around 2 hour to genrate on this kind of pod
XCan you recommend a pod that can help me generate wan much faster?
FH100
Frtx 4090 would be faster too
For L40S
SDr Furkan how many as increase the speed with teacache?
Ximage to video 720
Fit depends on your threshold
Fbigger faster
Fbut lower quality
Feven better
Jteacache update is not great here. I hope it will be faster in the end
Non my 3090 i was getting about 60s/it in the early steps but starting at about step 35 it started getting faster and was down to 33s/it for the last 4 or 5 steps
Jnot getting better. I'm using img to video 480p with 16gb VRAM and 32gb RAM (+100 GB of virtual RAM)
JI think it's even worse than without teacache
Fit gets faster as progress
Fwait
Fi am coding multiple lora and new auto scale feature now
Fby the way i hope original authors adds better teacache than me 
Jnot getting faster
