Furkan Gözükara
patreon
Simple prompt 2x latent upscaled FLUX - Fine Tuning / DreamBooth Images - Can be trained on as low as 6 GB GPUs - Each image 2048x2048 pixels
🕑 Added 2024-12-09 00:16:22 +0000 UTCAI Photos of Yourself - Workflow Guide
Step 1: Initial Setup
Follow any standard FLUX Fine-Tuning / DreamBooth tutorial of your choice
You can also follow mine step by step : https://youtu.be/FvpWy1x5etM
Step 2: Data Collection
Gather high-quality photos of yourself
I used a Poco X6 Pro (mid-tier phone) with good results
Ensure good variety in poses and lighting
Step 3: Training
Use "ohwx man" as the only caption for all images
Keep it simple - no complex descriptions needed
Step 4: Testing & Optimization
Use SwarmUI grid to find the optimal checkpoint
Test different variations to find what works best
Step 5: Generation Settings
Upscale Parameters:
Scale: 2x
Refiner Control: 0.6
Model: RealESRGAN_x4plus.pth
Prompt Used:
photograph of ohwx man wearing an amazing ultra expensive suit on a luxury studio<segment:yolo-face_yolov9c.pt-1,0.7,0.5>photograph of ohwx man
Note: The model naturally generated smiling expressions since the training dataset included many smiling photos.
Note: yolo-face_yolov9c.pt used to mask face and auto inpaint face to improve distant shot face quality
Comments
Furkan Gözükara
it is explained here : https://github.com/mcmonkeyprojects/SwarmUI/blob/master/docs/Features/Prompt%20Syntax.md#automatic-segmentation-and-refining
Furkan Gözükara
hello error is easy. you need to put models into diffusion_models folder
ProudChinaLover
I have a technical question: I did follow your steps in the tutorial; however, when I used the trained dreambooth in SwarmUI, I go this error message: All available backends failed to load the model ' C:\SwarmUI\SwarmUI\Models\Stable-Diffusion\Flux\Raflux-00020.safetensors'. What is this ? I did reinstall SwarmUI or update whatever I can . Anything wrong ? I did check ae , text encorder ... I can't get any luck.
eduardo
Can you explain the parameters you used for th yolo face model 1,0.7,0.5
cool1
Yes the 2K one doesn't look that great resolution (maybe it's also because they compress images for their AI model so much (they said they compress a lot more than others) and the model is a lot smaller than others). Though the 1024px one upscaled to 2K with Lanczos, while it looks brighter in parts, at 100% size it looks like the edges of the car (eg. the bottom edge between the wheels) on the upscaled one are more aliased/stair-case like than the 2048 original one. And at the back wheel edge there's maybe a bit more of a halo+step effect around the edge of the wheel on the upscaled one than the 2048 original one at 100% size.
Furkan Gözükara
here https://github.com/NVlabs/Sana/issues/83
Furkan Gözükara
well i generated 1024 and 2048 it only did regular upscale that you can do with paint .net. it didnt generate in 2048. you can test. it is just fake i dont know why they have it :d actually i will open a issue now for this
cool1
On Nvidia's SANA website (https://nvlabs.github.io/Sana/) they talk about 4K saying "our AE-F32 outputs 16× fewer latent tokens, which is crucial for efficient training and generating ultra-high-resolution images, such as 4K resolution."..."Linear attention achieves comparable results to vanilla, improving 4K generation by 1.7× in latency. ". Though maybe it's that 4K isn't high quality with SANA (since I think it was a sample image from that that looked too compressed with artefacts). edit: and faces get distorted in it to - that's what one video of it showed but it said they'd fix it in a later version.
Furkan Gözükara
i didnt see any benefits yet . tested on human and a style.
Ross
Is there any benefit to using training photos with higher resolution than 1024x1024? I'm curious if there are any benefits or consequences to using higher quality training photos.
Furkan Gözükara
sadly with flux others always almost turns into you (class bleeding). you need to inpaint their faces after image has been generated. you can use flux fill - inpainting model for that : https://youtu.be/hewDdVJEqOQ
Sonivas Sx
If i train a dream-booth fine tune, do i need regularisation images when i want to appear with random people or other Loras ? Because when i trained a Lora of myself it tends to make all other people in the image to look like me. Also how many , and what kind of images i need for regularisation for a dataset of 20 images ?
Furkan Gözükara
sana always 1024x1024 - 4k is fake :d when last time i tried 2048 it wasnt useful for training a person or a style but you can test and compare. actually i can also compare it today with higher res training
cool1
It looks impressive, but if it's upscaling won't it be inventing details that weren't in the trained lora/dreambooth (or missing details that weren't the lora/dreambooth)? Wouldn't it be more accurate if it could train at that 2048 res (though maybe it's almost impossible with Flux/SD3.5 type models without GPUs with huge amounts of VRAM. Maybe it was Nvidia's SANA that could do about 4K but that wasn't for commercial use and I think that was the one that some images at high res looked over compressed on some image(s) on reddit and I don't know that could be trained.
