Software Engineering Courses (SECourses)•10mo ago

You dont need 50 steps to test, nor 81 frames. Drop it to 20 steps and 33 frames for testing purpo

You dont need 50 steps to test, nor 81 frames.

Drop it to 20 steps and 33 frames for testing purposes. will save tons of time.

TTomfoolery You dont need 50 steps to test, nor 81 frames. Drop it to 20 steps and 33 fram...

JaviEM•3/15/25, 9:03 PM

I tried 20 steps once and I got a completely bugged video

JaviEM•3/15/25, 9:03 PM

I'll test 20 steps again next time

Furkan Gözükara SECourses•3/15/25, 9:04 PM

i can test faster if it is not NSFW. send me output and image

FFurkan Gözükara SECourses i can test faster if it is not NSFW. send me output and image

JaviEM•3/15/25, 9:04 PM

Don't worry, it's already in progress

JaviEM•3/15/25, 9:05 PM

I am in no hurry

JJaviEM I tried 20 steps once and I got a completely bugged video

TomfooleryOP•3/15/25, 9:12 PM

Hm, you definitely have something weird going on. This is 1.3b model, 25 steps, 49 frames for "A cinematic air-view tracking shot of a train waving through a forest, peaceful and serene."

Not a masterpiece, but a quick way to generate outputs to test prompts / functionality. This one took 104 seconds.

TTomfoolery Hm, you definitely have something weird going on. This is 1.3b model, 25 steps, ...

Furkan Gözükara SECourses•3/15/25, 9:12 PM

really good

Furkan Gözükara SECourses•3/15/25, 9:12 PM

i am gonna add <random:X, Y, Z, etc> now

TTomfoolery Hm, you definitely have something weird going on. This is 1.3b model, 25 steps, ...

JaviEM•3/15/25, 9:14 PM

Does it work the same with 14b? I tested with 14b model. I have no problems with 1.3b

JJaviEM Does it work the same with 14b? I tested with 14b model. I have no problems with...

TomfooleryOP•3/15/25, 9:22 PM

sure does, takes a bit longer to generate (10 minutes or so at 20 steps, 33 frames), but still works.

TomfooleryOP•3/15/25, 9:23 PM

ill generate the same prompt now

TomfooleryOP•3/15/25, 9:26 PM

you can also 'scale down' the 720p model and it handles it wonderfully, which speeds up generation times for testing

TomfooleryOP•3/15/25, 9:28 PM

10 minutes at full resolution for 20 steps, 33 frames, teacache at 0.15

TomfooleryOP•3/15/25, 9:29 PM

Again - not the best output in the world, but definitely lets you test things without having to spend an hour per generation.

TomfooleryOP•3/15/25, 9:29 PM

if you wanna save the headache of waiting each time.

TomfooleryOP•3/15/25, 9:34 PM

This is same inputs (different seed) as before, but using 16:9 low size. 3.5 minutes to generate. This oue actually turned out really nice

TTomfoolery This is same inputs (different seed) as before, but using 16:9 low size. 3.5 min...

Furkan Gözükara SECourses•3/15/25, 9:37 PM

very cool

JaviEM•3/15/25, 9:52 PM

@Tomfoolerywhat gpu and ram do you have?

TomfooleryOP•3/15/25, 9:55 PM

3090 Ti, 64 gb RAM.

TomfooleryOP•3/15/25, 9:56 PM

But the basic premise that lower steps/frame/size to speed up testing runs should work on all hardware.

Furkan Gözükara SECourses•3/15/25, 10:05 PM

i am about to add random word sytnax

Furkan Gözükara SECourses•3/15/25, 10:05 PM

and also fixing auto scale

TTomfoolery 3090 Ti, 64 gb RAM.

JaviEM•3/15/25, 10:05 PM

you have more vram and ram than me (4080 16GB VRAM and 32 RAM)

Furkan Gözükara SECourses•3/15/25, 10:10 PM

man i hate this orientation problme

trying to solve this too

Furkan Gözükara SECourses•3/15/25, 10:10 PM

during batch

Furkan Gözükara SECourses•3/15/25, 10:16 PM

ok fixed :d

Furkan Gözükara SECourses•3/15/25, 10:16 PM

time to push

Furkan Gözükara SECourses•3/15/25, 10:24 PM

v35 published : https://www.patreon.com/posts/123105403

Patreon

Wan 2.1 Ultra Advanced Gradio APP for - Works as low as 4GB VRAM - ...

Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon

Furkan Gözükara SECourses•3/15/25, 10:24 PM

@Tomfoolery

Furkan Gözükara SECourses•3/15/25, 10:25 PM

FFurkan Gözükara SECourses can you try same seed same prompt and image without teacache?

JaviEM•3/15/25, 10:50 PM

without teacache worked fine with a lot of hallucination but no image corruption. I think teacache is sometimes bugged

JJaviEM without teacache worked fine with a lot of hallucination but no image corruption...

Furkan Gözükara SECourses•3/15/25, 10:51 PM

could be. can you open an issue thread here and report if possible?

Furkan Gözükara SECourses•3/15/25, 10:51 PM

https://github.com/modelscope/DiffSynth-Studio/issues

GitHub

Issues · modelscope/DiffSynth-Studio

Enjoy the magic of Diffusion models! Contribute to modelscope/DiffSynth-Studio development by creating an account on GitHub.

Furkan Gözükara SECourses•3/15/25, 10:52 PM

i am working on FP8 model weight loading for image models

JaviEM•3/15/25, 11:04 PM

I just realised that despite having the same seed, prompts and settings (except teacache) the generated videos are different

JaviEM•3/15/25, 11:04 PM

I don't know if the test is 100% valid with different results

JaviEM•3/15/25, 11:06 PM

I will continue to do more tests and if I confirm the problems with teacache I will open the issue

Gasby•3/15/25, 11:23 PM

Another issue here with my 5090. At the end of the generation, it says ffmpeg cannot be found but I have it installed and on path. Is this a common issue?

GGasby Another issue here with my 5090. At the end of the generation, it says ffmpeg ca...

Furkan Gözükara SECourses•3/15/25, 11:39 PM

yes you have to install it

Furkan Gözükara SECourses•3/15/25, 11:39 PM

your installation wrong

Furkan Gözükara SECourses•3/15/25, 11:39 PM

https://youtu.be/DrhUHnYfwC0

YouTubeSECourses

How to Install Python, CUDA, cuDNN, C++ Build Tools, FFMPEG & Git T...

If you are interested in using AI, generative AI applications, open source applications in your computer, than this is the most fundamental and important tutorial that you need. In this tutorial I show and explain how to properly install appropriate Python versions accurately, how to switch between different Python versions, how to install diffe...