Software Engineering Courses (SECourses)•12mo ago

When generating a dataset for a character with flux for lora training, would developing a certain gu

When generating a dataset for a character with flux for lora training, would developing a certain guideline for poses be beneficial? or random images is better.
I'm thinking let's say: 8 close up images from different angles + 8 full body different angles + 8 waste up different lightning .

MacsalimOP•1/23/25, 11:16 PM

The reason I'm asking is that if we take this systematic approach, one click image to lora would be possible. because control net can be used to generate very certain poses.

MMacsalim When generating a dataset for a character with flux for lora training, would dev...

Furkan Gözükara SECourses•1/23/25, 11:19 PM

you should have every angle and pose you would like to generate after training

Furkan Gözükara SECourses•1/23/25, 11:19 PM

that is the key

Furkan Gözükara SECourses•1/23/25, 11:19 PM

and also emotion

AiInfluence•1/23/25, 11:20 PM

Doctor does training on .jpg or .png have any major difference?

FFurkan Gözükara SECourses and also emotion

MacsalimOP•1/23/25, 11:22 PM

then add 8 for emotions, or maybe sprinkle some emotions across the dataset.

MacsalimOP•1/23/25, 11:22 PM

as if it's that easy!

AAiInfluence Doctor does training on .jpg or .png have any major difference?

Furkan Gözükara SECourses•1/23/25, 11:30 PM

shouldnt have any diff

MMacsalim then add 8 for emotions, or maybe sprinkle some emotions across the dataset.

Furkan Gözükara SECourses•1/23/25, 11:30 PM

MMacsalim then add 8 for emotions, or maybe sprinkle some emotions across the dataset.

AiInfluence•1/23/25, 11:31 PM

you can train a stage 1 model and generate pictures of your subject with emotions

AiInfluence•1/23/25, 11:31 PM

then train with a new dataset

AiInfluence•1/23/25, 11:32 PM

or use inpainting on flux but haven't tested that method

MacsalimOP•1/23/25, 11:34 PM

I am testing Pulid to do so, without having to do multiple stages, it's pretty good

MMacsalim I am testing Pulid to do so, without having to do multiple stages, it's pretty g...

AiInfluence•1/23/25, 11:35 PM

make sure no noise is being generated/blur

MacsalimOP•1/23/25, 11:37 PM

Is this considered noisy? (I am still trying to get rid of the plastic skin)

MacsalimOP•1/23/25, 11:38 PM

this is a low res, here's the detail:

MMacsalim this is a low res, here's the detail:

Furkan Gözükara SECourses•1/23/25, 11:39 PM

it looks like cgi though

MacsalimOP•1/23/25, 11:39 PM

Agree, still working on that.

Furkan Gözükara SECourses•1/23/25, 11:50 PM

SDXL training now goes to as low as 10200 MB on Kohya

AiInfluence•1/23/25, 11:57 PM

@Furkan Gözükara SECourses what's the other unique token for sd called other than ohwx ?

AAiInfluence @Furkan Gözükara SECourses what's the other unique token for sd called other th...

Furkan Gözükara SECourses•1/23/25, 11:58 PM

bbuk

AiInfluence•1/23/25, 11:59 PM

here's another shot I made

AiInfluence•1/23/25, 11:59 PM

Messenger_creation_A0AA0389-937C-4891-A9BE-F31FB0C6874F.jpg

roosterpolice•1/24/25, 6:37 AM

Hiya - just wondering if anyone knows why SwarmUI inside latest massive compute instance would have trouble finding Flux model? I can see some funky alias stuff happeing towards download folder. I am a comfy user and can point to flux using that the normal way

roosterpolice•1/24/25, 6:59 AM

I got it working - some funky directory pointing going on in SwarmUI - also i am new to it coming from Comfy background and learing how the Dr does it. Cheers

J2_Nedo•1/24/25, 7:12 AM

this new Hunyuan3D 2.0 looks really interesting!

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•1/24/25, 8:32 AM

@Dr. Furkan Gözükara Will you be testing training on Hunyuan video? I am waiting for that.

Yobin•1/24/25, 10:11 AM

random thought, can you make flux loras a gguf???

Joshcandle•1/24/25, 10:22 AM

When i for example extract a lora from Checkpoint XY can i also use it with Checkpoint YX or will i have quality loss? (Both based on Flux) ?

YYobin random thought, can you make flux loras a gguf???

Joshcandle•1/24/25, 11:12 AM

Think i never saw any lora in gguf format, why would you need that?

YYobin random thought, can you make flux loras a gguf???

wardensc2•1/24/25, 11:55 AM

Ask this author, he is the one make GGUF to use for ComfyUI https://github.com/city96/ComfyUI-GGUF

GitHub

GitHub - city96/ComfyUI-GGUF: GGUF Quantization support for native ...

GGUF Quantization support for native ComfyUI models - city96/ComfyUI-GGUF

JJoshcandle Think i never saw any lora in gguf format, why would you need that?

Yobin•1/24/25, 11:59 AM

loras are gb for flux on 128 rank+ higher unlike sd 3.5/xls

YYobin loras are gb for flux on 128 rank+ higher unlike sd 3.5/xls

Joshcandle•1/24/25, 12:41 PM

lol, just checked, ive trained full checkpoint on a FP16 Model (11 GB , which resulted in finetuned checkpoint with around 22GB, so will the Lora be 11 GB big? why is it that huge?#

FFurkan Gözükara SECourses it looks like cgi though

VAMP•1/24/25, 1:12 PM

Doc would be buying the Rtx 5090?

FFurkan Gözükara SECourses it looks like cgi though

VAMP•1/24/25, 1:14 PM

And doc what do you think about NVIDIA Project DIGITS? ..Nvidia priced it at 3000$ ...Rtx 5090 is priced at 2000$...

dxqb•1/24/25, 2:07 PM

What is the state-of-the-art of creative upscaling? still SUPIR?

Joshcandle•1/24/25, 2:07 PM

Thats still what im using

Joshcandle•1/24/25, 2:07 PM

Also tried some comfy workflows but they cant reach supir level

VVAMP And doc what do you think about NVIDIA Project DIGITS? ..Nvidia priced it at 300...

dxqb•1/24/25, 2:08 PM

let's see. I believe DIGITS at 3k for this spec when I have it on my desk. Most likely with NVIDIA, 3k is the price for a DIGITS-060, and the DIGITS-090 is 3x that

JJoshcandle Also tried some comfy workflows but they cant reach supir level

dxqb•1/24/25, 2:09 PM

thanks

JJoshcandle Thats still what im using

dxqb•1/24/25, 2:11 PM

SUPIR was using JuggernautXL when I last used it. Have they upgraded to newer models such as Flux?

Joshcandle•1/24/25, 2:12 PM

@Furkan Gözükara SECourses Is Pinokio used for anything at massedcompute or can i delete it for more space?

Ddxqb SUPIR was using JuggernautXL when I last used it. Have they upgraded to newer mo...

Joshcandle•1/24/25, 2:12 PM

Im still using XL Models for Upscaling, dont even know if Flux works with supir

JJoshcandle Im still using XL Models for Upscaling, dont even know if Flux works with supir

dxqb•1/24/25, 2:13 PM

don't know either, but I had problems with SUPIR with teeth and other things XL typically has problems with, or needs ADetailer to fix it. So using the newer models would make sense

Joshcandle•1/24/25, 2:14 PM

Never had those problems with realvis

Joshcandle•1/24/25, 3:42 PM

How can i view the metadata of a flux checkpoint?

dxqb•1/24/25, 4:37 PM

what do you mean, what metadata

Joshcandle•1/24/25, 4:40 PM

like info tab in a1111 when selecting a checkpoint

dxqb•1/24/25, 4:41 PM

did you just answer your own question?

Joshcandle•1/24/25, 5:03 PM

No does not work with flux checkpoints

When generating a dataset for a character with flux for lora training, would developing a certain gu

Similar Threads