Software Engineering Courses (SECourses)•12mo ago

no single config fits all cases, you always have to play around with creativity and normalisation se

no single config fits all cases, you always have to play around with creativity and normalisation settings depending on composition, upscale factor and how close you need upscale to be to original

TimsonOP•1/26/25, 10:13 AM

This is what I would start with

Wwardensc2 Flux tend to make image more polished which people claim it's artificial. It dep...

N7-Blackbird•1/26/25, 12:32 PM

there is a huge difference between a raw photo (digital or analog) and a stylized photo (with filters) of a person (a portrait). If you work with photography your eye is trained on raw images and you notice when something is even slightly artificial. FLUX at it's current stage struggles a lot with creating raw portraits. With elaborate prompt you can get better results, but currently it is a struggle, whereas accordingly-trained stable diffusion 1.5 or XL custom models do this with ease. It would be great if one day there will be a more flexible FLUX version. That said, if you are looking to create studio-like portraits with retouched skin and filter effect, FLUX is very good. Just not for natural portraits, same as Midjourney btw.

AiInfluence•1/26/25, 1:17 PM

@Furkan Gözükara SECourses @Timson I'm planning as a next project to train shoes, jewellery and other objects. Should I follow the same concept of diversity as I did with the logo ?

AAiInfluence @Furkan Gözükara SECourses @Timson I'm planning as a next project to train shoes...

TimsonOP•1/26/25, 2:57 PM

you'll have to figure out the peculiarities of each class, but the starting point would be the same process. It all depends on how each of your base classes are trained into flux

Ppwndnoobcake tweaks like? my only trick is to prompt clip and t5xxl on their own format separ...

TimsonOP•1/26/25, 3:02 PM

Do you see real non placebo benefit in this? Do you use SwarmUI comfyUI flow customisation for this?
I did not follow up this research avenue because BFL is not leveraging it for flux pro API, seems counter-intuitive that they would ignore it if it was a viable way to improve outputs with no additinal costs

TTimson Do you see real non placebo benefit in this? Do you use SwarmUI comfyUI flow cus...

pwndnoobcake•1/26/25, 3:11 PM

comfy, yeah benefit is pretty decent

sergio_gonzo•1/26/25, 3:30 PM

Dr Furkan I try to use the copuon in massconpute but have a error,, dont accept it, surely i make a error

Flakdragoon•1/26/25, 5:31 PM

hi, I'm trying to use flux fill in SwarmUI to replace an orange wristband over a subject in a photo with some ISO noise. I'm using the basic generation flow and it looks perfect except it's too perfect and doesn't have any of the ISO noise in the original photo.

Is there a way I can get the generation to also add in similar noise to the inpainted mask that I'm shrink growing? I'm at max creativity since otherwise the orange wristband still shows. you can see in this image clip

FFlakdragoon hi, I'm trying to use flux fill in SwarmUI to replace an orange wristband over a...

TimsonOP•1/26/25, 7:12 PM

try reducing Mask Blur

TimsonOP•1/26/25, 7:13 PM

or just second pass it with Flux Dev

AiInfluence•1/26/25, 7:14 PM

Is it possible to inpaint an image straight forward?

AiInfluence•1/26/25, 7:15 PM

As init perhaps ?

AAiInfluence Is it possible to inpaint an image straight forward?

TimsonOP•1/26/25, 7:16 PM

what do you mean by "straight forward"?

TTimson what do you mean by "straight forward"?

AiInfluence•1/26/25, 7:16 PM

let's say I have a white board

AiInfluence•1/26/25, 7:17 PM

and I want to inpaint a statistics graph

AiInfluence•1/26/25, 7:17 PM

with details and stuff

AiInfluence•1/26/25, 7:17 PM

can I do that by providing swarm the image

AiInfluence•1/26/25, 7:17 PM

and let flux handle the inpainting?

AiInfluence•1/26/25, 7:18 PM

instead of training a Lora like I did with Stussy logo

TimsonOP•1/26/25, 7:18 PM

Do you also have statistics graph or you want it generated?

AlexGilSeg•1/26/25, 7:22 PM

Anyone that has trained loras locally with a 3060 using kohya?

TTimson try reducing Mask Blur

Flakdragoon•1/26/25, 7:24 PM

this is with mask blur 0. if I just gen it with the same prompt in flux dev, will it make the image more cohesive?

FFlakdragoon this is with mask blur 0. if I just gen it with the same prompt in flux dev, wi...

TimsonOP•1/26/25, 7:27 PM

you can just gen over this specific area with no prompt and flux should fix inconsistencies

TimsonOP•1/26/25, 7:27 PM

if you play around with creativity and normalization settings

TTimson if you play around with creativity and normalization settings

Flakdragoon•1/26/25, 7:27 PM

just to confirm, by normalization do you mean "init image reset to norm"?

AAiInfluence and let flux handle the inpainting?

TimsonOP•1/26/25, 7:27 PM

pretty trivial impaint

mini-whiteboard-mock-up-for-copy-space-3d-rendering.png

1924-white20board20with20colorful20very20detailed-sasb-1-2400-517543648.png

TTimson pretty trivial impaint

AiInfluence•1/26/25, 7:29 PM

something like this but with an image of my own , how did d you do that

AAiInfluence something like this but with an image of my own , how did d you do that

TimsonOP•1/26/25, 7:31 PM

just regular inpaint. This is generatied, it won't work with your own image

TTimson just regular inpaint. This is generatied, it won't work with your own image

AiInfluence•1/26/25, 7:32 PM

Ok so it's better to make a model of it got it

AiInfluence•1/26/25, 7:32 PM

Do you know where I can find a list of stable diffusion classes?

FFlakdragoon just to confirm, by normalization do you mean "init image reset to norm"?

TimsonOP•1/26/25, 10:03 PM

yep, sometimes it helps

AAiInfluence Ok so it's better to make a model of it got it

druhl•1/27/25, 7:35 AM

Why not just do this with photoshop/ Canva?

druhl•1/27/25, 7:36 AM

You can also composite the image and do a very light Ksampler pass to blend it in, but if it still changes stuff for you, why not just use photoshop?

Ddruhl Why not just do this with photoshop/ Canva?

AiInfluence•1/27/25, 10:55 AM

blending issues

AiInfluence•1/27/25, 10:56 AM

Also I'm experimenting on fluxs capabilities

druhl•1/27/25, 10:59 AM

So there is a compositor V3 node. You need to feed it images, place 'em where you need and then do a very light refiner-like pass

AAiInfluence Also I'm experimenting on fluxs capabilities

druhl•1/27/25, 11:00 AM

Give that a try

AiInfluence•1/27/25, 11:00 AM

Is there any tutorial I can watch?

Ddruhl Give that a try

AiInfluence•1/27/25, 11:00 AM

Also in on SwarmUi

TimsonOP•1/27/25, 11:23 AM

Has anyone tried installing swarmUI on ramdisk to speed up t5xxl/clip/vae switching for fp16 workflows on 24GB VRAM cards?
Current process seems grossly inefficient on machines with huge RAM - why this models can't be always loaded in RAM for near instant swapping to VRAM (6GB/s)?

TimsonOP•1/27/25, 11:27 AM

I don't seem to understand what is the bottleneck for this operation. Suspect comfyUI always unload models from RAM cache after loading to VRAM that causes extreme slowdown of this process

druhl•1/27/25, 2:36 PM

You can install things on the RAM? Lol

AiInfluence•1/27/25, 3:09 PM

@Furkan Gözükara SECourses any comments on Deepseek R1?

42069•1/27/25, 4:53 PM

https://github.com/deepseek-ai/Janus

GitHub

GitHub - deepseek-ai/Janus: Janus-Series: Unified Multimodal Unders...

Janus-Series: Unified Multimodal Understanding and Generation Models - deepseek-ai/Janus

42069•1/27/25, 4:54 PM

DeepSeek Janus Pro just released

42069•1/27/25, 5:44 PM

Multimodal LLM capable of text to image

daroth•1/27/25, 9:58 PM

not impressed

daroth•1/27/25, 9:59 PM

daroth•1/27/25, 10:00 PM

Hunter•1/27/25, 10:05 PM

@Furkan Gözükara SECourses thank you for all your work. I have created a finetuning model of myself and the result so far have been quite awesome! I need to learn a lot more on prompting, but its very easy to get quite good results! If you have any tips to learn realism prompts that would be cool!

no single config fits all cases, you always have to play around with creativity and normalisation se

Similar Threads