Software Engineering Courses (SECourses)•3y ago

can you passivly train with lora captions? like using the captions of the images which dont all repr

can you passivly train with lora captions? like using the captions of the images which dont all represent the same object but the same theme make that theme bleed into the composition even the lora was focused on a very different object.

RRae can you passivly train with lora captions? like using the captions of the images...

Furkan Gözükara SECourses•5/4/23, 9:30 PM

sure you can train with captions. i prefer captions only if fine tuning model

RaeOP•5/4/23, 9:31 PM

na iam not about training with captions but passivly train the lora and composition to be biased

RRae na iam not about training with captions but passivly train the lora and composit...

Furkan Gözükara SECourses•5/4/23, 9:31 PM

what you mean by that

RaeOP•5/4/23, 9:32 PM

lets say you train a lora on car rims, just rims. but u use mainly pictures of cars with rims to train the lora and have the captions be very detailed about the cars to train the lora about the rims to have it focus on the rim, but will the cars bleed into a composition later if you only want to show a stack of rims

RaeOP•5/4/23, 9:33 PM

can you use to passivly train the model to be more defined in regards to cars?

RaeOP•5/4/23, 9:34 PM

so basicly like priming LLM

RRae can you use to passivly train the model to be more defined in regards to cars?

Furkan Gözükara SECourses•5/4/23, 9:43 PM

passive training is unet

Furkan Gözükara SECourses•5/4/23, 9:43 PM

so as you do unet training, the unet will likely to output your trained images

FFurkan Gözükara SECourses i think you can do masked replace but i haven't done yet

Astuary•5/4/23, 9:57 PM

I found this tutorial (https://www.youtube.com/watch?v=rlfhv0gRAF4) and this extension (https://github.com/CiaraStrawberry/TemporalKit) that makes the face replacement a lot quicker. I haven't gotten perfect results but the process is simpler

Furkan Gözükara SECourses•5/4/23, 10:26 PM

started a twitch channel too : https://www.twitch.tv/secourses

Twitch

SECourses - Twitch

How to do Free Speech-to-Text Transcription Better Than Google Premium API with OpenAI Whisper Model

randomd•5/4/23, 11:13 PM

reinstalled A1111

randomd•5/4/23, 11:13 PM

everytime I start webui-user.bat

randomd•5/4/23, 11:14 PM

it spends a lot of time "Installing requirements". And its always the same requirements

randomd•5/4/23, 11:14 PM

Does anyone know the arguments to skip this?

Rrandomd it spends a lot of time "Installing requirements". And its always the same requi...

Furkan Gözükara SECourses•5/4/23, 11:28 PM

i also looked for wiki for this

Furkan Gözükara SECourses•5/4/23, 11:28 PM

there werent option

FFurkan Gözükara SECourses started a twitch channel too : https://www.twitch.tv/secourses

Uzy•5/5/23, 12:43 AM

with the same contents of YT?

Rrandomd Does anyone know the arguments to skip this?

Spydah•5/5/23, 4:21 AM

--skip-install is used in the webui-user.sh file. Seems to work for me in runpod.

FFurkan Gözükara SECourses anyone reading can you please upvote i appreciate it : https://www.reddit.com/r/...

Khang•5/5/23, 5:47 AM

That's a very usefull video thank you very much for sharing. I asked ChatGPT to make shorter reddit title using click-bait type (imo it's fine since you always deliver what you said) and it gave some interesting ideas:

Original "How To Find Best Stable Diffusion (any AI) Generated Images By Using DeepFace AI - Step By Step Full Tutorial - Let The AI Find Best Images For You When You Generated Images With LoRA or DreamBooth or Textual Inversion Or Any Way - Can Be Used For Professional Tasks As Well"

>> Effortlessly Discover the Most Striking Generated Images with DeepFace - Step-by-Step Guide Inside!
>> Stop Wasting Time! Let AI Find the Best Generated Images for You - Step-by-Step Guide Included!
>> Find the Best Images with SD and DeepFace - Save 100 hours this week - Full Tutorial Inside!
>> I let DeepFace find Perfect Images among 1000 results - Save tons of time"
>> I'm tired of checking 1000 SD images manually so I let Deepfake do it for me
>> SD on Steroid: Check 1000 images in seconds with Deepfake

FFurkan Gözükara SECourses what video i should i make next

AlexGilSeg•5/5/23, 8:20 AM

Thanks, my latest DB model is almost completed after your latest video instructions, will be intressting.. I think your next video should be about EveryDream2 and comparing it with dreambooth.. I've been messing with it for a few weeks and I find it to be able to reproduce much better images than dreambooth and a much more styled manner BUT, it is not as consistent as dreambooth, in my experience one has to produce more images to get "good" images, with dreambooth you get more similarity but the price is styleization.. But that's just me, you are much more adept at these things, it would be intressting to see what you come up with..

AlexGilSeg•5/5/23, 8:30 AM

btw, about that info, are the images you are showing there in the video upscaled? I could see from your details that it's using a model called analog madness, you didn't mention any of this.. seems it might be important..

UUzy with the same contents of YT?

Furkan Gözükara SECourses•5/5/23, 9:22 AM

yep

SSpydah --skip-install is used in the webui-user.sh file. Seems to work for me in runpod...

Furkan Gözükara SECourses•5/5/23, 9:22 AM

thanks i am going to add this to upcoming tutorial

KKhang That's a very usefull video thank you very much for sharing. I asked ChatGPT to ...

Furkan Gözükara SECourses•5/5/23, 9:23 AM

thank you so much. i asked mods of the reddit to delete old and repost with a better title

AAlexGilSeg Thanks, my latest DB model is almost completed after your latest video instructi...

Furkan Gözükara SECourses•5/5/23, 9:24 AM

EveryDream2 is same of dreambooth extension without using classification images in theory. but practical implementation matters

FFurkan Gözükara SECourses EveryDream2 is same of dreambooth extension without using classification images ...

AlexGilSeg•5/5/23, 9:50 AM

but it seems one can train much further than with dreambooth, and use many more images, and no need to resize and make 512x512 etc.. I get much better stylization there anyway..

AlexGilSeg•5/5/23, 9:51 AM

on another matter, something has happend with my xyz plot, when I use the method for checkpoint name it loads weights from the wrong model, a model that I haven't even selected.. It's very strange, do you have any ideas?

AlexGilSeg•5/5/23, 10:07 AM

btw, the images in your vidos, are they before or after the upscale?

AlexGilSeg•5/5/23, 10:26 AM

not really sure how to upscale with the analogmadness model, that would be useful.. and again I'm not at all getting the same amazing results you are.. that's why I'm asking if your videos you are showing in the vid are upscaled..

AAlexGilSeg on another matter, something has happend with my xyz plot, when I use the method...

Furkan Gözükara SECourses•5/5/23, 12:20 PM

model loading system could be broken not sure. there are also browser caching issues i encountered

AAlexGilSeg btw, the images in your vidos, are they before or after the upscale?

Furkan Gözükara SECourses•5/5/23, 12:21 PM

they are not upscaled raw

AAlexGilSeg not really sure how to upscale with the analogmadness model, that would be usefu...

Furkan Gözükara SECourses•5/5/23, 12:21 PM

only thumbnail upscaled with 4x ultra sharp

Furkan Gözükara SECourses•5/5/23, 2:15 PM

purchased 13900k for to be able to do more experiment and make better videos

Furkan Gözükara SECourses•5/5/23, 2:16 PM

anyone have anything for me to test? i plan testing

Furkan Gözükara SECourses•5/5/23, 2:16 PM

core i7 10700f vs 13900k

Furkan Gözükara SECourses•5/5/23, 2:30 PM

a new community post is made : https://www.youtube.com/@SECourses/community

YouTube

SECourses

Welcome to Software Engineering Courses (SECourses) – the ultimate destination for skillfully curated insights into state-of-the-art technologies and programming paradigms. We demystify the realms of Artificial Intelligence, Stable Diffusion, DreamBooth, LoRA, ControlNet, Textual Inversion, Software Engineering, Programming, C#, .NET, ASP .NET, ...

TheWhykiki•5/5/23, 4:36 PM

I want to generate an image like this:

https://media.istockphoto.com/id/1266756979/de/foto/r%C3%BCckansicht-von-zwei-muldenkippern-f%C3%A4hrt-stadtausw%C3%A4rts-auf-einer-flachen-asphaltstra%C3%9Fe.jpg?s=612x612&w=is&k=20&c=Z7hgkxB6nLUCzbbrfe3NrNsvI89LYick75U_a70QAFQ=

Anyone some ideas what to prompt and which model to use?

The best I got so far looked like this:

parameters

a couple of semi trucks driving down a road, close-up shot taken from behind, traveling in france, full body profile camera shot, high quality detailed, white border, shipping containers, photograph”, waste, detailed wheels, head and upper body in frame, trailer, dutch, port, no long neck, panorama shot, cone heads, stock photo, masterpiece,best quality, advertisement, high quality photo
Negative prompt: (worst quality:2), (low quality:2), (normal quality:2), lowres, normal quality, ((monochrome))
Steps: 25, Sampler: Euler a, CFG scale: 4.5, Seed: 3134462633, Size: 512x512, Model hash: e6415c4892, Model: civitai_realisticVisionV20_v20

TheWhykiki•5/5/23, 4:43 PM

And some funny question also:

When using the prompt "RAW photo of woman" to generate class images mostly I get a variation opf nearly the same face - EVERY TIME!
She looks like this, in different variations - that's a little bit creepy:

TheWhykiki•5/5/23, 4:43 PM

Anyone of you experienced the same??

TheWhykiki•5/5/23, 4:43 PM

Model I used RealisticVision 20_20

crazygonuts•5/5/23, 5:16 PM

most models if you don't prompt for any specifics that would modify the face will give you their "default" face

crazygonuts•5/5/23, 5:18 PM

if you want a different face, prompting for facial features - "square jaw", "high cheekbones" - doesn't work very well. What DOES work very well is prompting for celebrities - it knows what they look like. And you can remix celebrities, so [ angelina jolie | scarlett johansson ] will give you a mix between their faces. if you prompt for a specific celebrity but only for part of the generation like [ angelina jolie :: 0.8 ] then you will usually get a softer version of that celebrities face as it mixes it down at the end of the generation

crazygonuts•5/5/23, 5:20 PM

you can even put specific celebrities in the negative prompt and it will modify the face away from what they look like - so if you want a woman, put some male celebrity in the negative and it will change the face for you

Haliphas•5/5/23, 5:32 PM

good day guys, I'm new here and was directed here from the youtube video on deepface for finding best image. I ran into this issue while installing deepface and was wondering if anyone can help guide me to resolve it? I don't have programming experience so this is all a bit confusing for me

Haliphas•5/5/23, 5:32 PM

thanks for any assistance in advance

RudraYodha•5/5/23, 5:32 PM

Hi, Hi, I need help to set up SD on Google Colab without using GPU, like Easy Diffusion can do that on local pc without any GPU. I know it's going to take a lot of time to render, but still, I am ok with that. I am ok with using only command line interface, so no GUI is needed. Any links or help ?

HHaliphas good day guys, I'm new here and was directed here from the youtube video on dee...

ashley•5/5/23, 6:07 PM

This is just a warning, not an error, you can ignore this.

Zono50•5/5/23, 6:19 PM

so my lora model looks great when i use it on the model it was trained on, but when i start using other models, i don't see my person at all, what's the best way to use my models face across other checkpoints?

Aashley This is just a warning, not an error, you can ignore this.

Haliphas•5/5/23, 6:56 PM

Thanks!

ZZono50 so my lora model looks great when i use it on the model it was trained on, but w...

YVD•5/5/23, 8:39 PM

I have the same problem, it seems like the Lora has a very weak effect even when increasing it’s weight.

can you passivly train with lora captions? like using the captions of the images which dont all repr

Similar Threads