Software Engineering Courses (SECourses)•10mo ago

I'm checking wan 2.1 in replicate and the default is 30 steps with a maximum of 40. Maybe the defaul

I'm checking wan 2.1 in replicate and the default is 30 steps with a maximum of 40. Maybe the default 50 steps we have here is too many.

JJaviEM I'm checking wan 2.1 in replicate and the default is 30 steps with a maximum of ...

Furkan Gözükara SECourses•3/16/25, 12:26 AM

surely you can reduce

Furkan Gözükara SECourses•3/16/25, 12:26 AM

replicate is always about making more money = lower quality

JaviEMOP•3/16/25, 12:28 AM

I have tried 15 steps and the quality is decent. 30 should be enough to get good quality

Gasby•3/16/25, 12:41 AM

New WAN app works flawlessly doc. Thanks

asap17z•3/16/25, 1:23 AM

What’s teacache and why is it only giving me errors on mass compute and runpod but not on my system( I used conda and pip for my system)

asap17z•3/16/25, 1:34 AM

Comfyui vs gradio. I prefer Comfyui for everything else, but ever since I started using your wan2.1 example I could control how it use my gpu, thereby making it possible to use my system for other things outside of wan, with Comfyui it almost impossible to do so, even on a 4090. Although there are bugs on the runpod and mass compute example you provided. I would appreciate it you can find a solution for them

JJaviEM I have tried 15 steps and the quality is decent. 30 should be enough to get good...

J2_Nedo•3/16/25, 9:21 AM

i tested 20,30,40, and 50 steps with the same seed, 50 gives the best quality, at least thats what i found out, but for quick testing i use lower steps

GGasby New WAN app works flawlessly doc. Thanks

Furkan Gözükara SECourses•3/16/25, 11:30 AM

awesome

Aasap17z What’s teacache and why is it only giving me errors on mass compute and runpod b...

Furkan Gözükara SECourses•3/16/25, 11:30 AM

which zip file you used on massed compute and runpod?

Furkan Gözükara SECourses•3/16/25, 11:31 AM

also guys i am about to fix an error with image to video models

Furkan Gözükara SECourses•3/16/25, 11:31 AM

we were supposed to load models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth as FP32

asap17z•3/16/25, 11:31 AM

wan21_v31

Furkan Gözükara SECourses•3/16/25, 11:31 AM

we were loading as BF16

Furkan Gözükara SECourses•3/16/25, 11:31 AM

i am testing

Aasap17z wan21_v31

Furkan Gözükara SECourses•3/16/25, 11:31 AM

v35 should work

Furkan Gözükara SECourses•3/16/25, 11:31 AM

but i will test newest version soon

Furkan Gözükara SECourses•3/16/25, 11:31 AM

which I will publish

asap17z•3/16/25, 11:32 AM

is it on patreon?

Aasap17z is it on patreon?

Furkan Gözükara SECourses•3/16/25, 11:34 AM

yes. wait me update to latest

Furkan Gözükara SECourses•3/16/25, 11:34 AM

give me few min

Furkan Gözükara SECourses•3/16/25, 11:35 AM

it will also fix fp8 loading for image to video

Furkan Gözükara SECourses•3/16/25, 11:38 AM

wow FP8 really reducing VRAM

asap17z•3/16/25, 11:40 AM

okay sir

Furkan Gözükara SECourses•3/16/25, 11:43 AM

I should re-set params BF16 vs FP8

FP8 really reduces VRAM a lot

asap17z•3/16/25, 11:46 AM

okay, I just downloaded it.

asap17z•3/16/25, 11:50 AM

I'm going to test this Gradio version along side with my Comfyui workflow on Runpod and mass compute. but I prefer using your version on my system because it's better optimize for consumer pc

Aasap17z I'm going to test this Gradio version along side with my Comfyui workflow on Run...

Furkan Gözükara SECourses•3/16/25, 12:00 PM

just give me few min

Furkan Gözükara SECourses•3/16/25, 12:00 PM

almost done

Furkan Gözükara SECourses•3/16/25, 12:03 PM

ok now you can install

Furkan Gözükara SECourses•3/16/25, 12:03 PM

it will install v36

Furkan Gözükara SECourses•3/16/25, 12:08 PM

v36

Furkan Gözükara SECourses•3/16/25, 12:10 PM

asap17z•3/16/25, 12:10 PM

okay

Furkan Gözükara SECourses•3/16/25, 12:10 PM

fixed VRAMs

Furkan Gözükara SECourses•3/16/25, 12:40 PM

GPU VRAM Usages Table With Respected To Number of Persistent Params

1280x720 px, 81 frames, 14b text to video model - BF16 precision

All tests made on secondary GPU thus they are 100% accurate

0 is : 14168 MB
1 is : 14168 MB
1,000,000,000 is : 15658 MB
2,000,000,000 is : 17664 MB
3,000,000,000 is : 20134 MB
3,750,000,000 is : 21134 MB
4,000,000,000 is : 21500 MB
4,250,000,000 is : 22000 MB

JaviEMOP•3/16/25, 12:48 PM

Img to video 480p BF16 with 16GB preset is using 11,6-11,8GB of VRAM

JJaviEM Img to video 480p BF16 with 16GB preset is using 11,6-11,8GB of VRAM

Furkan Gözükara SECourses•3/16/25, 12:48 PM

i am testing that right now :d

Furkan Gözükara SECourses•3/16/25, 12:48 PM

480x832 px, 81 frames, 14b 480p image to video model - BF16 precision

0 is : 8618 MB
1 is : 8618 MB
1,000,000,000 is : 9354 MB
2,000,000,000 is : 11370 MB
3,000,000,000 is : 13080 MB

Furkan Gözükara SECourses•3/16/25, 12:49 PM

these are exact usages so you also need to add your windows usage

Furkan Gözükara SECourses•3/16/25, 12:50 PM

i will publish as a table hopefully after finishing all :d

JaviEMOP•3/16/25, 12:52 PM

Furkan Gözükara SECourses•3/16/25, 1:28 PM

GPU VRAM Usages Table With Respected To Number of Persistent Params

1280x720 px, 81 frames, 14b text to video model - BF16 precision (torch.bfloat16)

All tests made on secondary GPU thus they are 100% accurate 

0 is : 14168 MB
1 is : 14168 MB
1,000,000,000 is : 15658 MB
2,000,000,000 is : 17664 MB
3,000,000,000 is : 20134 MB
3,750,000,000 is : 21134 MB
4,000,000,000 is : 21500 MB
4,250,000,000 is : 22000 MB


480x832 px, 81 frames, 14b 480p image to video model - BF16 precision (torch.bfloat16)

0 is : 8618 MB
1 is : 8618 MB
1,000,000,000 is : 9354 MB
1,500,000,000 is : 10090 MB
2,000,000,000 is : 11370 MB
3,000,000,000 is : 13080 MB
3,500,000,000 is : 14038 MB
4,000,000,000 is : 15140 MB
8,000,000,000 is : 22100 MB


720x1280 px, 81 frames, 14b 720p image to video model - BF16 precision (torch.bfloat16)

0 is : 18326 MB
1 is : 18326 MB
2,000,000,000 is : 20360 MB
3,000,000,000 is : 22010 MB
3,500,000,000 is : x MB
4,000,000,000 is : x MB
8,000,000,000 is : x MB

GPU VRAM Usages Table With Respected To Number of Persistent Params

1280x720 px, 81 frames, 14b text to video model - BF16 precision (torch.bfloat16)

All tests made on secondary GPU thus they are 100% accurate 

0 is : 14168 MB
1 is : 14168 MB
1,000,000,000 is : 15658 MB
2,000,000,000 is : 17664 MB
3,000,000,000 is : 20134 MB
3,750,000,000 is : 21134 MB
4,000,000,000 is : 21500 MB
4,250,000,000 is : 22000 MB


480x832 px, 81 frames, 14b 480p image to video model - BF16 precision (torch.bfloat16)

0 is : 8618 MB
1 is : 8618 MB
1,000,000,000 is : 9354 MB
1,500,000,000 is : 10090 MB
2,000,000,000 is : 11370 MB
3,000,000,000 is : 13080 MB
3,500,000,000 is : 14038 MB
4,000,000,000 is : 15140 MB
8,000,000,000 is : 22100 MB


720x1280 px, 81 frames, 14b 720p image to video model - BF16 precision (torch.bfloat16)

0 is : 18326 MB
1 is : 18326 MB
2,000,000,000 is : 20360 MB
3,000,000,000 is : 22010 MB
3,500,000,000 is : x MB
4,000,000,000 is : x MB
8,000,000,000 is : x MB

Furkan Gözükara SECourses•3/16/25, 1:29 PM

i am gonna update all presets

Furkan Gözükara SECourses•3/16/25, 1:30 PM

also now supports , in params

Furkan Gözükara SECourses•3/16/25, 1:31 PM

asap17z•3/16/25, 1:39 PM

I'm currently using 80gb

Furkan Gözükara SECourses•3/16/25, 1:39 PM

now i will test rest on 80 gb

Furkan Gözükara SECourses•3/16/25, 1:39 PM

for 80gb just enter like 20b

Furkan Gözükara SECourses•3/16/25, 1:39 PM

it will use vram entirely

asap17z•3/16/25, 1:42 PM

is it okay to increase the denoise for the sampler to 1?

GPU VRAM Usages Table With Respected To Number of Persistent Params 1280x720 px, 81 frames, 14b text to video model - BF16 precision (torch.bfloat16) All tests made on secondary GPU thus they are 100% accurate 0 is : 14168 MB 1 is : 14168 MB 1,000,000,000 is : 15658 MB 2,000,000,000 is : 17664 MB 3,000,000,000 is : 20134 MB 3,750,000,000 is : 21134 MB 4,000,000,000 is : 21500 MB 4,250,000,000 is : 22000 MB 480x832 px, 81 frames, 14b 480p image to video model - BF16 precision (torch.bfloat16) 0 is : 8618 MB 1 is : 8618 MB 1,000,000,000 is : 9354 MB 1,500,000,000 is : 10090 MB 2,000,000,000 is : 11370 MB 3,000,000,000 is : 13080 MB 3,500,000,000 is : 14038 MB 4,000,000,000 is : 15140 MB 8,000,000,000 is : 22100 MB 720x1280 px, 81 frames, 14b 720p image to video model - BF16 precision (torch.bfloat16) 0 is : 18326 MB 1 is : 18326 MB 2,000,000,000 is : 20360 MB 3,000,000,000 is : 22010 MB 3,500,000,000 is : x MB 4,000,000,000 is : x MB 8,000,000,000 is : x MB

I'm checking wan 2.1 in replicate and the default is 30 steps with a maximum of 40. Maybe the defaul

Similar Threads

Similar Threads

Similar Threads