So basically anything below 1024x1024 makes it better i feel
So basically anything below 1024x1024 makes it better i feel
Cthat's pretty standard. it has everythign to do with whether the AI was trained on images that were all the same AR or if it was trained on images of all different ARs. if they were all the same AR - then it knows how to draw that and things get wierd when it tries to draw some other aspevt ratio
Cyou see that a lot with stable diffusion 1.5 - which was trained on 512x512 images. ask for an AR that's not 512x512, what it does is draw multiple 1:1 AR images and concatenate them together - resulting in heads on top of heads in a 9:16 image or other strangeness
Fi hope its dev version get released 
Fyou can mix
Fat worst case you can reduce timing
Fi found that it is nice for prompts like a dinasasur
Fwhat issue you have
FI have dumped T5 XXL token list and working on a new rare token
Fthat will be rare in T5 too
F@Dr. Furkan Gözükara , I trained my own LoRA Flux with the Flux FP8 model, and I want to generate images with my LoRA for free on Kaggle. Before I subscribe to your Patreon, I wanted to ask if your Patreon post about "Free Kaggle Account Notebook for SwarmUI with Stable Diffusion 3 and Dual T4 GPU support" , can use FP8 LoRAs ?. I previously tried using Forge , but the images result were blurry and pixelated.
Fyes it can use
Fbut cant train
Fblurry result happens due to cfg not being 1
Fwe use latest swarmui and comfyui as backend
Fworks perfect
Nbeen getting good results sometimes by using Flux at
512*512 or even 384*384 in the first passNthe VAE is so much better than SDXL so you can do more passes
PHow do you do mutliple passes? Are you running it in Comfy Ui? I am still on Web Ui Forge
Nyeah in Comfy, but you can do this in Forge as well
Nwhat you need to do is take the result of the first pass
Nupscale it using an external upscaler
Nand then do the standard img-to-img in Forge using a denoise % less than 1
IHi everyone, should I use regularisation images when training a lora on sdxl?
Ftesting flux rare tokens
Fsingle words makes such good images
F
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=16)
XCame out really low quality with the details on my face, maybe I’ll try with another image
XIt would be so cool if we could generate videos with our models using text to video. Is there anything out like this? @Dr. Furkan Gözükara
Cthere are lots of text to video AIs out there - and meta just annouced this https://ai.meta.com/research/movie-gen/
AI at Meta
Meta Movie Gen is our latest research breakthrough that allows you to use simple text inputs to create videos and sounds, edit existing videos or transform your personal image into a unique video.
Fnothing good
Fi am about to publish a video for it
XYes, I heard about this, but we’re unable to use our own trained models
Cyou might want to wait till it's more than a research paper
Cyes, well - you could use them with animate diff or live portrait
XHmmmm need to look into animate diff
XSent you a friend request
XThanks will check it in a few minutes
Fyes nothing published yet
FYouTubeSECourses
Movie Gen sets a new standard for immersive AI content. Meta announced the Meta Movie Gen just today which is going to change the Cinema, Video and Animation industries forever. This will shock the industry workers.
Paper : https://ai.meta.com/static-resource/movie-gen-research-paper
How Meta Movie Gen could usher in a new AI-enabled era for c...
Paper : https://ai.meta.com/static-resource/movie-gen-research-paper
How Meta Movie Gen could usher in a new AI-enabled era for c...
C@Dr. Furkan Gözükara Have you had any success with the batch 7 training? If the config is ready but no video yet would it be possible to receive the config file to try?
Fyes it is working great
Fbut i still didnt finalize
Flet me put into zip file
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
OJust a quick question on Supir upscaling: are there a collection of presets that match up with some of the tests run? I'm not seeing exactly how to configure this to match up with the grid tests run on the SECourses channel. Is there a go-to configuration that I can start with and experiment from there?
I'm likely most interested in upscaling images generated in Flux and likely 2x to 4x in resolution
I'm likely most interested in upscaling images generated in Flux and likely 2x to 4x in resolution
Fi have a default config
Fthat upscales with loyal to original image