Software Engineering Courses (SECourses)•10mo ago

A upscaler would be nice feature to generate 480 and render to higher resolution as is really fast

ᴏᴏᴍɴɪ A upscaler would be nice feature to generate 480 and render to higher resolution...

Furkan Gözükara SECourses•3/30/25, 11:34 PM

yes upscaler coming hopefully soon

Furkan Gözükara SECourses•3/30/25, 11:34 PM

with best AI + fast ones

Original message was deleted

ᴏ

ᴏᴍɴɪOP•3/31/25, 1:25 AM

ᴏ

ᴏᴍɴɪOP•3/31/25, 1:25 AM

OC Openai model follows prompt well

ᴏ

ᴏᴍɴɪOP•3/31/25, 1:36 AM

Never mine I think they blocked it from generating styles...
I'm sorry, but I can't generate that image as it violates our content policies. Please feel free to suggest a different prompt or idea—I'm here to help with any creative concepts that follow the guidelines. Let me know what you'd like to try next!

cangir01•3/31/25, 4:47 AM

@Furkan Gözükara SECourses I have installed with conda same like previous versions without errors, but to be 100% sure I have installed also with your script, also the same result. The videos are being generated extreme blurry, after uninstalling sage attention, it works again....I think I will stay at version v38 untill, which works for me good enough. Thank you and have a nice week

ᴏᴏᴍɴɪ Never mine I think they blocked it from generating styles... I'm sorry, but I ca...

Furkan Gözükara SECourses•3/31/25, 8:58 AM

yes everday they block more

Ccangir01 <@205854764540362752> I have installed with conda same like previous versions wi...

Furkan Gözükara SECourses•3/31/25, 8:58 AM

sage attention supposed to work but you know it is hard to tell will work on every case

FFurkan Gözükara SECourses our repo dont have it but it looks not very good

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•3/31/25, 9:02 AM

It works good when you set zero init steps to 1

cangir01•3/31/25, 9:47 AM

@Furkan Gözükara SECourses if I disable Tae cache, then I don't have blurry results. But this is new, I used always Tae cache with v38 and in previous versions. The problem occurs only in 14B text to video model, which is very curious in my opinion. Did you test also yourself with this model?

Ccangir01 <@205854764540362752> if I disable Tae cache, then I don't have blurry results. ...

Furkan Gözükara SECourses•3/31/25, 10:21 AM

yes right now generating let me shoıw

Furkan Gözükara SECourses•3/31/25, 10:21 AM

screencapture-8447a4423530048b36-gradio-live-2025-03-31-13_21_36.png

Furkan Gözükara SECourses•3/31/25, 10:21 AM

compiled sage attention and testing it on

Furkan Gözükara SECourses•3/31/25, 10:22 AM

Furkan Gözükara SECourses•3/31/25, 10:22 AM

i will post result here once completed

Ccangir01 <@205854764540362752> if I disable Tae cache, then I don't have blurry results. ...

Furkan Gözükara SECourses•3/31/25, 11:24 AM

here

cangir01•3/31/25, 12:34 PM

Thank you very much, I will post also in the evening my results then here, in order that you can see what the problem is. But this means the problem is in my setup..very strange

FFurkan Gözükara SECourses you need to make a fresh install

Aikage•3/31/25, 7:22 PM

Thank you this fixed it and also wow the new version cut generation time in half and quality is somehow much improved. Amazing!!!

AAikage Thank you this fixed it and also wow the new version cut generation time in half...

Furkan Gözükara SECourses•3/31/25, 8:57 PM

awesome

Ccangir01 Thank you very much, I will post also in the evening my results then here, in or...

Furkan Gözükara SECourses•3/31/25, 8:57 PM

sure

cangir01•3/31/25, 11:39 PM

0003.txt2KB

cangir01•3/31/25, 11:44 PM

here is the same prompt with the same model text to video 14B but v49. It is not perfect but good enough to use it

Ccangir01 Click to see attachment

Furkan Gözükara SECourses•3/31/25, 11:55 PM

let me test

Ccangir01 Click to see attachment

Furkan Gözükara SECourses•3/31/25, 11:56 PM

ah your error is 70 steps

Furkan Gözükara SECourses•3/31/25, 11:56 PM

use 20 steps with teacache

Furkan Gözükara SECourses•3/31/25, 11:56 PM

teacache is steps sensitive

Furkan Gözükara SECourses•3/31/25, 11:56 PM

i am testing

cangir01•3/31/25, 11:58 PM

But the same prompt worked before in version v49, the last video is the proof

Ccangir01 But the same prompt worked before in version v49, the last video is the proof

Furkan Gözükara SECourses•3/31/25, 11:58 PM

because teacache changed

Furkan Gözükara SECourses•3/31/25, 11:58 PM

now we use official implementation

Furkan Gözükara SECourses•3/31/25, 11:58 PM

i am generating a video right now

cangir01•3/31/25, 11:58 PM

söylese daha önceden, anlasildi

cangir01•3/31/25, 11:59 PM

ama o zaman 20 Stepin kalitesi ile 70 in ki bir olmaz ki, bu gelisme gercekten iyi mi oldu?

Ccangir01 ama o zaman 20 Stepin kalitesi ile 70 in ki bir olmaz ki, bu gelisme gercekten i...

Furkan Gözükara SECourses•4/1/25, 12:01 AM

if you want max quality you can turn off tea cache and still do but quality almost same.

cangir01•4/1/25, 12:02 AM

What is the max step with Taecache enabled with the newer version? 20 Steps

cangir01•4/1/25, 12:03 AM

Without cache it takes too long time

Ccangir01 What is the max step with Taecache enabled with the newer version? 20 Steps

Furkan Gözükara SECourses•4/1/25, 12:07 AM

i didnt test but after some point it reduces quality for some reason

Furkan Gözükara SECourses•4/1/25, 12:07 AM

this prompt didnt work great also :d

Furkan Gözükara SECourses•4/1/25, 12:07 AM

you sure it works?

Furkan Gözükara SECourses•4/1/25, 12:07 AM

Furkan Gözükara SECourses•4/1/25, 12:08 AM

cangir01•4/1/25, 12:08 AM

I will give you the original prompt from last video which worked:

Ccangir01 I will give you the original prompt from last video which worked:

Furkan Gözükara SECourses•4/1/25, 12:09 AM

sure

cangir01•4/1/25, 12:09 AM

Here is the video without sound and prompt

00006.txt1.65KB

cangir01•4/1/25, 12:10 AM

another similar video:

00002.txt1.57KB

Ccangir01 another similar video:

Furkan Gözükara SECourses•4/1/25, 12:11 AM

testing without teacache

Furkan Gözükara SECourses•4/1/25, 12:11 AM

Furkan Gözükara SECourses•4/1/25, 12:12 AM

cangir01•4/1/25, 12:12 AM

There are a lot of results, which were also not good, but the promlem is the text in most cases not the visuals, as you know transformer models have still difficulty to create letters, sometimes they are really good but never perfect, always small mistakes

cangir01•4/1/25, 12:13 AM

Without cache in 9 min is fast enough I think.

A upscaler would be nice feature to generate 480 and render to higher resolution as is really fast

Similar Threads

Similar Threads

Similar Threads