New FLEX AI Model!?

Let's look at Flex1.Alpha finetune base.
Download model huggingface.co...
Place in /models/checkpoints/
Download Turbo Lora huggingface.co...
Place in /models/loras/
Workflow (download at bottom) / 120375562
Recommended online AI platform for Stable Diffusion & Flux (affiliate): www.thinkdiffu...
ComfyUI workflow for 1.5 models: / comfyui-1-5-86145057
ComfyUI Workflow for SDXL: / comfyui-workflow-86104919
Chat with me in our community discord: / discord
Stable Diffusion for Beginners Playlist • Stable Diffusion Begin...
My Weekly AI Art Challenges • Let's AI Paint - Weekl...
My Stable diffusion workflow to Perfect Images • Revealing my Workflow ...
ControlNet tutorial and install guide • NEW ControlNet for Sta...
Famous Scenes Remade by ControlNet AI • Famous Scenes Remade b...

[@jsnavely76]

More models is always good. I'll certainly download it and add it to my collection of models, but I couldn't really care less about licensing. I'll do whatever I please with the Dev model. It's the pirate's life for me.

[@fluffsquirrel]

Just keep embedded data in mind. Models may use steganography to stamp every image generated with their signature.

[@yellowboat8773]

Argh me hearties 🏴‍☠️

[@deadlyrobot5179]

AI image development has slowed down a lot, SD1.5 and SDXL are still more popular than Flux and SD3.5 because they're more consumer friendly. I hoped SD3.5 became more popular because I'm tired of the Flux and its unsolvable issues, it's practically untrainable.

[@MarioSedlak]

Flux is the most stable one. Not the fastest on the poor gaming gpu's. I trained more than 5 models and more than 20 personal loras for flux and it's pretty stable. You just need to find right way.

[@deadlyrobot5179]

​@@MarioSedlak Every single flux models that are "trained" on Civitai give you messed up body proportions. As soon as you load a Lora, it gets even worse. The "better" AI models get, the worse they perform in anything other than Realistic images. Flux is so tuned to the max if you try to make it "better" you'll end up doing a side grade not an upgrade, but SD3.5 on the other hand has a big headroom for improvement by the community.

[@MarioSedlak]

@@deadlyrobot5179 Who cares about civitai? They sucks pretty bad.

[@888felipe]

​@@MarioSedlakhow did you train? Do you know a good tutorial?

[@spacetimeranus]

Will it help me get bricked up?

[@rjw8631]

good to see you again. i haven't seen much from you lately.

[@sebastiankamph]

Good to see you in the comments! End of year business and all that. Hopefully more to come.

[@ZelpinAI]

Thank you for great tutorial.

[@MilesBellas]

"The Flex.1-alpha model is an 8-billion-parameter rectified flow transformer designed to generate images from text descriptions. It was fine-tuned from FLUX.1-schnell and retains an Apache 2.0 license, ensuring open-source compliance. The model includes a guidance embedder, eliminating the need for classifier-free guidance (CFG), and supports fine-tuning. It is compatible with popular inference engines like ComfyUI and Diffusers, making it highly accessible for developers." - Perplexity

[@musicalbug.]

Thank, I learned some thing new

[@glitchfactory404]

Thank's Sebastian

[@sebastiankamph]

😊💫

[@Nob1ej0n]

This is good to see. I haven't used Flux Dev at all because of that license. And support for Schnell is lacking right now. Hopefully Flex improves soon.

[@Holymalus]

Is there any model out there that does skin better than a fine-tuned SD1.5?

[@ParvathyKapoor]

flux 1.1pro or ultra but not available for download

[@dkemil]

many 😂

[@DivinityIsPurity]

7:11 Yes, drop the cfg to lose the plasticky output from 1.5 to 1.2 or 1 for example

[@sebastiankamph]

Tried but sadly didn't work for me like it does with Flux dev guidance.

[@zengrath]

I don't see much point in playing with this now, but will see if it gets fine tuned

[@Lahouel]

Add Ultra Real Photo Lora weight 0.8. No more plasticity. Hands are good as is.

[@sebastiankamph]

Great tip, I'll try it!

[@blackvx]

Thanks

[@sebastiankamph]

Happy to help my dear doggo

[@lesslyrics]

Error in Workflow? Load Lora CLIP output not connected to inputs of CLIP Text Encode.

[@sebastiankamph]

Didn't make a difference here, did comparison test. Feel free to test further and report back.

[@lesslyrics]

@@sebastiankamph I've only done a few tests with this model. Unfortunately, it has no added value for me. Unlike your videos! 😊👍

[@solomslls]

so if flux dev lora can work with this you can basically use "realistic lora" and it should improve the plastic skin thing, but the problem its still 20GB file !!!

[@sebastiankamph]

Thanks, will test it!

[@ShubzGhuman]

downloading it gonna test it, i have idea

[@solomslls]

@@ShubzGhuman please tell me if it works great

[@thibaudherbert3144]

I'm not using flux anymore. If you wan to use imagery for professional usages, you must have fast results, commerical usage, by the time goes on, a knew ai model in 6 month or 1 year will come out and solve the current problems with flux. Ideogram + midjourney + SDXL for lora is the best combo. Flux is not trainable easily, takes forever to generate an image, does not support openpose, and commercialy unusable unless you use schnell but you can't use any loras lol. Flux is dead to me

[@metasamsara]

this is what flex is about, using lora with schnell base

[@thibaudherbert3144]

@@metasamsara if you can use dev loras and controlnet with this checkpoint then yes it's interesting, but if you can't, it's useless. (90% of loras are trained for dev not schnell, so you lose all the benefits of flux...)

[@sebastiankamph]

You can use dev outputs for commercial purposes according to the license.

[@jibcot8541]

On an RTX 5090 a Flux image can take only 3.5 seconds now.

[@HolidayAtHome]

wasn't' sure if it's flux, saw the flux face in the beginning, ok it's flux ;P

[@MonDiabolique]

is forge or automatic1111 dead?

[@niceone3091]

no

[@richctv]

As someone who is pressing on to learn ComfyUI and all its intricacies; Forge is very much alive and well and actively updated.

[@biberwahn]

moooooooor

[@mikeGietExperience]

what real use is of it?

[@z1mt0n1x2]

1:14 Sees a 21GB file... Closing the video. Not interested.

[@springheeledjackofthegurdi2117]

Oh so it's like the pony illustrious merge model?

[@metasamsara]

what do you mean?

[@springheeledjackofthegurdi2117]

@metasamsara there's a model on civitai, that can use both pony and ILl models

[@chipko]

Did I miss the dad joke?!?!

[@sebastiankamph]

Oops

[@fluffsquirrel]

@@sebastiankamphWhy is this even funnier 😂

[@MauMau4448]

Still the flux chin I see

[@0A01amir]

Does it need 64gb vram and takes 10min to generate a generic image like other Flux models ? looking at downloaded models all Flux models are dead on arrival. it killed image generation AI.

[@sebastiankamph]

It's similar to Flux Schnell and dev dedistilled. Running at 8 steps in my workflow.

[@0A01amir]

@@sebastiankamph Quality is still med, amount of vram you will need to run a mediocre model, amount of power and time you waste is just a huge step backwards. SDXL take like what ? 5 second on a low to midranged GPU to produce a good quality image, SD 3.5 is also like that but horrible human bodies. Flux is like something that youtubers with their 4090s and soon 5090s can only run and it's alive because of some sort of marketing behind the scene, not by community. doesn't matter how you look at it, it's inefficient. and sorry i talked alot with my terrible english, your videos are good and really fun to watch.

[@sebastiankamph]

@ You raise some valid points. Flux is tough to run currently. Do bear in mind that people didn't like SDXL much when it was released, as 1.5 finetunes totally destroyed it at the time. But things evolve :) I do hope there will be faster/easier to run Flux variants though that can deliver above SDXL quality.

[@0A01amir]

@@sebastiankamph Thank you. yeah but SDXL was running just fine with 6gb vram and it took off pretty quickly, nothing happened to Flux since release, hell FluxChin is still present in almost all models. A pretty model (tbh flux isn't one) that needs alot of raw power to run and even with flagship hardwares it has problems will never takes off, doesn't matter how they market it, SD 1.5 anyone could run it days after release, SDXL almost anyone can run now same as SD 3.5 but when it come to Flux they laugh at you even if you have 16GB 800 to 1200$ cards.

[@metasamsara]

Downloading it out of curiosity to test the censorship. Models that are too censored to follow prompt get abandonned by me pretty quick tho. I'm not about to use 3 LORAS just to get a bagina that looks semi decent.

[@andreasknops3123]

Find god

[@solomslls]

@@andreasknops3123 hahahaha

[@artplaenan445]

We just need cats with funny hats. That's all !

[@metasamsara]

@@andreasknops3123 if you found him you'd know only him can judge me. yet you do.

[@gisabelle7724]

not good for porn i ll pass xD