Support NeoGAF

[Mr1999]

Not sure if anyones been following but Flux has been the latest craze in the ai image world the last couple of weeks and with lora support everyones making new loras. I have been converting my old datasets so I can use it with Flux and the results are amazing in comparison to sdxl. Feels like skies the limit because sdxl never really stylized as well as flux does, or even pick up on features, a lot of the times features were completely missing from sdxl lora no matter what I did, but flux does a really good job at figuring out what is what, especially complex enteities. Im using an NVIDIA A100 80GB's right now to train but I think soon, maybe even now you can train using 12GB, but not sure. Here are some of my standouts, these were created using my old datasets which is like half the fun.

 

[Sokka]

Flux is amazing, I've been using it myself these past few days.

Can't wait to see what's in the future, looking forward to their text to video model.

 

[Unknown Soldier]

I have an RTX 3090, is 24 GB enough for this and is there a guide for people who literally know nothing to teach them how to get started with this sort of thing

 

[gundalf]

Can I run this on my M3 Max?

 

[RJMacready73]

Is there any websites that offer complete newbs access to these AI image generators, what would be the best ones to use? Microsofts was an absolute pain in the hole to use if there's anything remotely violent in the text prompt, I mean I just want to fuck around and get some Judge Dredd, Megacity stuff

 

[Mr1999]

I have an RTX 3090, is 24 GB enough for this and is there a guide for people who literally know nothing to teach them how to get started with this sort of thing

My advice is to check out youtube and look for comfyui flux tutorial which will tell you to download comfyui portable and download flux dev, schnell or nf4. You can definitely run it with that hardware ,I think you can even run it using 6GB correct me if im wrong but you will likely have to run using

--lowvram

, for those with 12GB or less I believe. I know im probably not making sense, check out youtube or civitai , search for comfyui flux tutorial and watch latest videos since its changing by the day. I think you can now run nf4 which is supposedly faster but I use dev currently which is 4 minutes a gen at 1920x1080.

Can I run this on my M3 Max?

Just to run it Im pretty sure you can.

Is there any websites that offer complete newbs access to these AI image generators, what would be the best ones to use? Microsofts was an absolute pain in the hole to use if there's anything remotely violent in the text prompt, I mean I just want to fuck around and get some Judge Dredd, Megacity stuff

There are tons but in the end the downfall of all these public ai image sites like midjourney is that you can't modify it a whole lot, they look great but you can only do so much and the results can become predictable after a while which to me takes the fun out of it, I want to be surprised. https://huggingface.co/spaces/markury/FLUX.1-dev-LoRA allows you to mess around with flux and you can even link loras from huggingface, and use it for free up to a certain point just to check loras without actually downloading it, but you dont have to use lora, its optional if you just want to mess around with flux.

The results above were taken with custom datasets which is going to take some time to learn for someone starting new unless they make it easier which im sure there are easier way sof doing it, like right now im using this tutorial https://civitai.com/articles/6678/training-a-flux-lora-using-simpletuner-on-runpod for making my own loras and even though its explained and the files created for you, I still ran into issues starting the training and had to use chatgpt to find out why it was not running and then run using the commands it gave me. Kohya_ss will soon (if not already) have the ability to train locally and you dont have to have a lot of memory either, but im too lazy and spoiled to use my lowly 4070ti when I can rent a A100 instead, its a shame since I'm familiar with Kohya_ss.

The good thing is you dont have to make your own loras, there are sites like civitai that has a huge library of loras already made by users, but first you gotta learn how to use comfyui at the basic level, cool thing is you can drag and drop a .png gen into the browser and it will make the workflow for you, then you just have change a few settings like model/loader/prompt, click queue prompt and sit back.

 

[Haint]

My advice is to check out youtube and look for comfyui flux tutorial which will tell you to download comfyui portable and download flux dev, schnell or nf4. You can definitely run it with that hardware ,I think you can even run it using 6GB correct me if im wrong but you will likely have to run using

--lowvram

, for those with 12GB or less I believe. I know im probably not making sense, check out youtube or civitai , search for comfyui flux tutorial and watch latest videos since its changing by the day. I think you can now run nf4 which is supposedly faster but I use dev currently which is 4 minutes a gen at 1920x1080.


Just to run it Im pretty sure you can.


There are tons but in the end the downfall of all these public ai image sites like midjourney is that you can't modify it a whole lot, they look great but you can only do so much and the results can become predictable after a while which to me takes the fun out of it, I want to be surprised. https://huggingface.co/spaces/markury/FLUX.1-dev-LoRA allows you to mess around with flux and you can even link loras from huggingface, and use it for free up to a certain point just to check loras without actually downloading it, but you dont have to use lora, its optional if you just want to mess around with flux.

The results above were taken with custom datasets which is going to take some time to learn for someone starting new unless they make it easier which im sure there are easier way sof doing it, like right now im using this tutorial https://civitai.com/articles/6678/training-a-flux-lora-using-simpletuner-on-runpod for making my own loras and even though its explained and the files created for you, I still ran into issues starting the training and had to use chatgpt to find out why it was not running and then run using the commands it gave me. Kohya_ss will soon (if not already) have the ability to train locally and you dont have to have a lot of memory either, but im too lazy and spoiled to use my lowly 4070ti when I can rent a A100 instead, its a shame since I'm familiar with Kohya_ss.

The good thing is you dont have to make your own loras, there are sites like civitai that has a huge library of loras already made by users, but first you gotta learn how to use comfyui at the basic level, cool thing is you can drag and drop a .png gen into the browser and it will make the workflow for you, then you just have change a few settings like model/loader/prompt, click queue prompt and sit back.

3 or 4 years on how and why is this shit still so convoluted? What is the technical reason why this stuff can't be packaged as a traditional windows exe with a standard GUI you install with 1 click like literally every other piece of software in the last 30 years?

 

[Miyazaki’s Slave]

Is there any websites that offer complete newbs access to these AI image generators, what would be the best ones to use? Microsofts was an absolute pain in the hole to use if there's anything remotely violent in the text prompt, I mean I just want to fuck around and get some Judge Dredd, Megacity stuff

Not sure if anyones been following but Flux has been the latest craze in the ai image world the last couple of weeks and with lora support everyones making new loras. I have been converting my old datasets so I can use it with Flux and the results are amazing in comparison to sdxl. Feels like skies the limit because sdxl never really stylized as well as flux does, or even pick up on features, a lot of the times features were completely missing from sdxl lora no matter what I did, but flux does a really good job at figuring out what is what, especially complex enteities. Im using an NVIDIA A100 80GB's right now to train but I think soon, maybe even now you can train using 12GB, but not sure. Here are some of my standouts, these were created using my old datasets which is like half the fun.

Did you tell it to add what looks like strawberry nerd candy to the desk?

That sells the FUCK out of that image for me. It feels like a TMNT or (REALLY bad ass) He-Man action figure.

 

[badblue]

It feels like a TMNT or (REALLY bad ass) He-Man action figure.

I thought so too. So I went looking because I remember having a figure like that when I was a kid.

Mutagen Man from TMNT.

 

[Mr1999]

3 or 4 years on how and why is this shit still so convoluted? What is the technical reason why this stuff can't be packaged as a traditional windows exe with a standard GUI you install with 1 click like literally every other piece of software in the last 30 years?

Although not 1 click yet, they've made it much easier with it being portable and now if you are missing something you can automatically install it using Comfy manager.

Did you tell it to add what looks like strawberry nerd candy to the desk?

That sells the FUCK out of that image for me. It feels like a TMNT or (REALLY bad ass) He-Man action figure.

I thought so too. So I went looking because I remember having a figure like that when I was a kid.

Mutagen Man from TMNT.

Correct

Still working out a few things on my Cain lora but here's a quick gen.

 

[WhamBamThankYouMaam]

3 or 4 years on how and why is this shit still so convoluted? What is the technical reason why this stuff can't be packaged as a traditional windows exe with a standard GUI you install with 1 click like literally every other piece of software in the last 30 years?

Because making actually working installers is a task in itself, and people working with ai aren’t necessary focusing on that particular branch of the tech tree.

 

[sankt-Antonio]

3 or 4 years on how and why is this shit still so convoluted? What is the technical reason why this stuff can't be packaged as a traditional windows exe with a standard GUI you install with 1 click like literally every other piece of software in the last 30 years?

The same reason Linux is never taking off, programmers think they are edgy when using the console to operate software.

 

[Lord Panda]

Some of my generations from today:

Spoiler: NSFW?

 

[Sokka]

I have an RTX 3090, is 24 GB enough for this and is there a guide for people who literally know nothing to teach them how to get started with this sort of thing

ForgeUI or SwarmUI are both friendly to beginners.

With ForgeUI, you go to their github page and download a compressed file. Unpack it and then click on the run.bat file.

You'll also need to download a Flux model that you put in a specific folder so that ForgeUI can find it.

If you want more exact details, you can PM me for help.

 

[Tumle]

Im using swarm UI and it’s straight forward really good results and best with text I have seen

made for my friends to post in chat if we want to play warzone. Made a face swap on the characters with our own faces

 

[AJUMP23]

AI girlfriends incoming.

 

[Tumle]

AI girlfriends incoming.

More like curated porn.. but when robots come or full body vr haptic suit the sure