How to Install Forge UI & FLUX Models: The Ultimate Guide

[@pixaroma]

You can now support the channel and unlock exclusive perks by becoming a member: kzbin.info/door/mMbwA-s3GZDKVzGZ-kPwaQjoin Join the conversation on Discord discord.gg/gggpkVgBf3 or in our Facebook group facebook.com/groups/pixaromacommunity.

[@guy.incognito]

Your tutorials are pretty much the best I've seen ever. Most others either leave out important details or if they don't the whole thing is 2 hours long. You are the only one who keeps it short while still including every single step and lots of useful information. Much appreciated. Thank you!

[@tjteino26]

What a great and clear introduction to Forge-Flux Combo, up and running on the first try :) Thanks!

[@liquidmind]

yes!

[@TheMisterNebo]

I don't comment often on people's tutorials because most of them suck. Yours are excellent. They are detailed, clean, and to the point. You clearly took the time to create these videos and I thank you for that.

[@tapirko1]

Thanks!

[@pixaroma]

Thank you so much for your support ☺️

[@tetragrammaton3]

and thank you from a viewer for supporting him. You help all of us as well.

[@GenoG]

I've been playing around with Flux on Forge for a few days, this is the best "how to get started" video I've seen yet! And a learned a bunch of other basic stuff too!! Thank you! You earned Fi Dolla!! 🤑

[@pixaroma]

Thank you very much, I always try to do videos that can help, and how i wish to find videos on any subject, explained so beginers can understand and yet not too boring for advanced users to skip. Since the channel grows slowly and I don't earn much, any contribution helps me, so thank you for your support 😊

[@sandnerdaniel]

You do not need to restart UI to add styles, just use the Edit style button and refresh the styles with the icon. Well made video tutorial!

[@dangerszewski9816]

the best "how to get started with offline genAI" tutorial I've found! fantastic stuff! This is what we REALLY need. People to teach other people how to get their feet wet and be participants not observers in the AI revolution, or at least play with it and understand what it is-- and what it's not -- when politicians and breathless "tech communicators" start ranting about it.

[@evilchucky69]

very nice tut buddy. easy to follow and provides all the little extras that come in handy. not many on you tube have this talent. please don't change your style.

[@pixaroma]

Thank you ☺️

[@DrDaab]

Excellent not only for Flux but for basic Forge installation and usage. Best! Thanks ! ! ! !

[@Huang-uj9rt]

Hello brother, may I ask where you are running flux? Ever run it on mimicpc? Everyone says it works great with it, but maybe due to my limited ability, the images generated are not to my satisfaction, and I would like to know where I can tweak it in terms of details?

[@Neveen.SalahAl-Din]

This is the first video I saw on the channel and I really want to thank you for the detailed tutorial and the amazing styles u provided 🙏

[@Turije]

This was an excellent all round tutorial for lots of concepts and comparisons, thanks for taking the time to do this, very muc appreciated.

[@JP0p0san]

Thank you for this information because my head is exploding with building up my model but also the installation of it all. I tried SDXL (Which only has Flux Lora's as the moment), ComfyUI (Which almost fried my pc due to a faulty setting) and now Forge which is an absolute beast. i will be looking foward to your next upload

[@nicoteen974]

Awsome, precise guide without being too technical. Thank you so much.

[@andrewiwinski3225]

This is awesome work man. I never comment on stuff but this is hella detailed.

[@Farah_AI_Film]

Thank you so much for this tutorial. I have been trying to get Flux installed locally for a week ( I'm kinda slow) and this is the first tutorial that clicked with me and now I'm generating locally. So happy... although I definitely need a new GPU LOL.

[@otternas3]

Heared about ForgeUI Yesterday, finding your Video was a terasure of Information how to get goin Appreciate the clear and nobs instructions and all the Tips You Sir, earned a Sub and a new Fan today! Have a great Weekend

[@dreamzdziner8484]

I was exhausted running cmd commands. Thank you so much for that roll back technique.🧡 Forge + Flux seems unstoppable. 🤩💪

[@liquidmind]

FINALLY!!! A WAY TO USE FLUX WITH ONLY 6 GB VRAM!!! It takes 20 seconds for the schnell model and 1:30 minutes for the dev model!!!!

[@cybersphere]

I have an RTX3060 with 12GB of VRAM and 32GB of system RAM. Using the dev.bnb.nf4.v2 safetensor with an Async swap method and shared swap location, I get a render time of around 1 min 35 sec. With Schnell and the right step count of 4, my render time is 30 seconds. Some of my schnell renders have come out pretty good. Considering the pricepoint of the RTX3060, I'm happy.

[@pixaroma]

That is good, on my rtx2060 6gb of vram for schnell nf4 i get 27 sec, i wonder why you don't get a faster speed for that since you have a better card. For dev nf4 i get 1min and 40sec, i have been using auto default

[@cybersphere]

@@pixaroma As well as changing the default swap method and swap location, I changed the image dimension to 1024*1024. Changing my swap method back to queue, I get a render time of 1 min 22 seconds for dev-nf4-v2. I think you just have to play around with the options to see what works best on your system.

[@moelleunbelievable]

1 min 35 sec sounds insanely long to me. I got i9 13900k 128gb ram and either 4090 or 4070ti super to compare, so yeah I understand the hardware difference but mine takes about max 20 seconds to render an image.

[@cybersphere]

@@moelleunbelievable 1-2 minutes render time is typical for the 3060 using the dev model. If you check the comments, you'll see that some people's systems are taking over 3 minutes to render an image with similar or superior hardware. I just updated forge and rested. dev is now down to 1 min 20 sec and schnell is down to 21 sec. For most images, the schnell(fast) model is fine. Later model cards will of course be faster, but they have a poorer price/performance ratio.

[@amatrixa2923]

Great video. It's so hard keeping up with the advances of Flux and trying our best to use it outside of Comfy UI makes life much easier. The problem is using the best Flux model that won't drain the life out of your VRAM in a another UI. Nothing against Comfy UI, there are just times when you want things less complicated and Forge UI to the rescue.

[@MegaSuperJaBaTo]

Thank you for the styles. Also thank you again for the tutorials. I was thinking that "Forge" is much to complicated for a beginner, but with the informations from your videos it's becomming a piece of cake. 😉

[@Xenodermus-V1]

Great video with detailed info. Since I've discovered it, Forge is definitely my favorite platform. I am still on a slow RTX but it's a trooper, will try the SHNELL flux model really hoping it is good enough to hold it. Fingers crossed! :)

[@vonBlankenburgLP]

Thank you very very much for this insightful video. I used it as a tutorial to install flux on my machine.

[@baheth3elmy16]

Great video! I feel nostalgic watching Forge again.

[@TechneticDreams]

Thank you for the tutorial friend, keep creating and being awesome!

[@epick-studios]

This is very comprehensive. Thanks for the hard work 😊

[@sparktuneonline]

Great tuto, i love how you simplify things, keep going

[@JP0p0san]

I'm running a 4070 TI with 32RAM and 2TB of M2 but my renders take about 3 minutes. Tho i must say the flux is out of this world. The quality but also the reduced requirements of your system really makes this the best at this moment

[@pixaroma]

try more updates, for example when I finished the video I did another update, and on older pc was faster with Diffusion in Low Bits on automatic, for dev nf4 only took less then 2 minutes on rtx 2060 with 6gb of vram, so try updating and see if is faster, since rtx2060 is weaker card then yours so it should not take 3 minutes for a nf4 model, schnell nf4 model is 27 second on rtx2060 and 3.4 second on rtx4090

[@liquidmind]

@@pixaroma sorry, i want to make sure - i have that 2060 RTX CARD, you said on the video i can generate a photo with DEV model, but takes 9 minutes, and 20 seconds for a schnell model?

[@pixaroma]

@@liquidmind yeah that time taken but in the next day i updated again forge and now take 27 sec for schnell nf4 and 1 min 40 sec for dev nf4 so probably was a bug when i tested and they fixed that

[@liquidmind]

@@pixaroma ok thats great! one more question, if there are many demo webpages for free and paid subscription, some of them even use a BLEND model of DEV and SCHNELL, why is it differente to use locally? is it uncensored only locally?

[@pixaroma]

@@liquidmind mostly ask money because you are using their resource, the power cost money, computer running, and big video cards all cost money. Probably online they put more safety and make it more censored. But the models are the same or a mix of models ot add some extra lora to get more style and make a fancy interface. For example you can take the flux schnell model, make an interface for it, put it on your server and put it online and ask people money for generating with it. Other just use api, there is pro.model that give you certain nr.of generation for certain amount of $ , so you could build a fancy interface and call it how you want and ask people money to generate, more than the api and server maintain cost so you are on profit.

[@ninocrudele]

Thank you! best tutorial around for sure!

[@llllIIIlIll]

Quick note if you want to revert or update torch or xformers ver you can use command in .bat file with notepad add --reinstall-torch or --reinstall-xformers(in webui folder)

[@pixaroma]

Thank you ☺️ good to know

[@guy.incognito]

Danke!

[@pixaroma]

Thank you so much ☺️

[@thehandleiwantedwasntavailable]

I can't tell if the voice is AI or not - it flips between sounding real and fake constantly 😂 Great video all the same!

[@pixaroma]

yeah is AI :) is hard to keep it consistent since I do text in a few sentences, i dont do a text for all video, i watch the recording and write the text for 1 minute of video, then i convert to audio then an synchronize it with video, then repeat :)

[@thehandleiwantedwasntavailable]

@@pixaroma Makes sense! Thanks again for the video. I've been using AUTOMATIC1111 with SD for about 18 months and this helped with the minor differences switching to Forge and FLUX.

[@MarcusPrimeX]

I just fresh installed Forge on my pc, did follow both the 1 pack folder and the manual way. Both methods does not show me the TRAIN, SVD and Z123. I also followed the guide @16:24 but still the same, when i run the forge it still does not have those 3. I used the same exact commit as shown in the video to the rollback file, still same issue. Please take note that this is the very first time iam using and installing forge. I havent used Train, SVD and Z123 on Automatic1111. Do i actually need to install those first on my pc? thanks!

[@pixaroma]

So yesterday when i did the video it worked ok, did you get a message or something that say it couldn't swich to that commit? I mean if all was done like in video and the forge ui was the same , you should get the same result. When you started forge look if it start with the version you put in the commit, if it starts it should have that old interface. But that is useful if you need svd, but i didn't use it in months since are better ai for video like kling and lume , raleway

[@AscendantStoic]

Great tutorial, thanks for the info and links.

[@gilles49280]

You rock big time !!! Thank you for your valuable explanations 🥰

[@CsokaErno]

Thank you very, very much. You change my life :) More, more of this, please!!!

[@genin69]

Prob the best video on flux I've seen

[@TheColonelJJ]

Thank you for this! I've been waiting. I just can't wrap my old mind around ComfyUI.

[@pixaroma]

You should check the comfyui series after 2-3 episodes will make sense, i was like you and avoided comfyui for months, but now you can really do much more with it, and i am trying to simplify the nodes so it looks similar to forge

[@Tuugii47]

Thank you. That was really great introduction.

[@Tuugii47]

@pixaroma but after installed all files image section is blank not generated images shows, what shall i do ?

[@justinharubra]

Thanks for the video and for the rollback script! Now Tiling works again.

[@neomicryo]

okay i managed to extract the file ! it works, thanks !

[@nvadito7579]

12:13 I have a 2060 with 12 GB, and still, making the cat image in 4 steps with that model takes me 23 seconds. Is that normal? The settings are the default ones from Flux. And if I go a few seconds without making images, it can take up to 40 seconds when I resume

[@pixaroma]

depends on your system, i have also 2060 with 6gb but i have 64gb of vram and the forge is on fast ssd. So are many things that it takes in account, and first time is slower, so it might, not sure, or maybe is a bug and in a few days if you update maybe works ok, is hard to tell

[@SelenaSella-x2s]

What could be wrong here? at 2:33 in the video, on his checkpoint, he has a model, while I don’t have any model even though I did the same steps he did , pls i really need help

[@pixaroma]

Maybe you don't have long path activated in windows and wasn't able to download it, you can download it manually , there are a lot of modles on civit ai

[@turjoturjo7422]

Great tutorial! Please make a video on- flux controlnet, flux ipadapter, flux faceID or InstantID all inside Forge UI 🙏

[@pixaroma]

I rarely use forge, i prefer comfy UI, i have more control, and forge still has a lot of bugs and is frustrating

[@LanceWinder]

Great stuff man, thanks!

[@jennifertsang6572]

Awesome video! Thank you!!

[@jeffwads]

3080ti 12GB Vram here. 6:25. This flux worked for me. Generation time is around 15 seconds or so.

[@pixaroma]

Yeah i updated again forge next day and got a faster speed for rtx2060 6gb also on older pc, compared with a few days before, probably i got a bad version on older pc, but now i get decent speed :)

[@paulcurry8383]

I'm suprised you were able to get flux to work immediately. I get AssertionError: You do not have CLIP state dict! which seems to imply I need to install other things for the model to work.

[@pixaroma]

Make sure you are using the right t5 and clip models, they changed the interface after i done tutorial, so you need the right clip models for the right flux model

[@ZuckerbergDaReptilian]

Which TTS did you use for this video?? ...It sounds pretty natural to me, nicely done!

[@pixaroma]

voiceair and elevenlabs, voiceair has api from elevenlabs.

[@marvellousimagination]

short and sweet. thank you very much sir!

[@kKhanMC]

About what you mentioned at 5:32, I've got a 1650 on the PC I'm trying to run this. Will there be any issues other than slow generation of images?

[@pixaroma]

Only if you try it, the system can influence a lot how things work, for some might crash for others work, depends on many things so just try it. See if works with sd, then try fkux schnell maybe

[@ralf.starke]

Thank you, this helped a lot!

[@thevoidchronicles-r5q]

So question, I get this: AssertionError: You do not have CLIP state dict!? Any ideas why?

[@pixaroma]

check this github.com/lllyasviel/stable-diffusion-webui-forge/issues/1075 and this github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1050

[@rechellesaavedra9318]

THANKS SO MUCH. YOU DA MAN! SUBCRIBED!

[@GrandFromage1]

Thank you for this, I needed the discussion on samplers. (so far DPM++ 2M / SGM Uniform seems best for speed+quality!) I'm running on a 4070 (12GB vram) and 64GB memory - with the full DEV model and T5xxl_fp16 - and I get between 3.2s/it to 6.5s/it (so renders take 2-3 minutes each, but at really amazing quality! Strangely resolution isn't a huge impact on speed, obviously steps are.) So yes, ForgeUI can run the full DEV FLUX.1 without a 4090!

[@liquidmind]

rare, i have 2060 6GB VRAM, and i get faster renders,,,,,, ? why? i can schnell a 4 steps photo in 20 seconds and DEV a photo in 1:20 minutes - 20 steps and im using the bigger DEV fp8 model, - might the sampler you are using is slower...... i use SIMPLE instead of SGM UNIFORM, maybe thats it.

[@GenoG]

I didn't have any luck with Samplers other than Euler... Maybe with ...NF4...V2 the DPM samplers work?? That would be great! Good details! Thx!

[@liquidmind]

@@GenoG DPM++ 2m is fine, but SIMPLE, not KARRAS. Karras sometimes gives grey results. EULER is nice on 8 steps schnell or 22steps DEV,.

[@Redemtor62]

Thanks for this! - The new updates were pretty broken :D

[@pixaroma]

Keep.looking for recent commits that says verified, there are a few new version since i made the video 2 days ago :) they update it every hour, hope one of them will be stable enough

[@rubenbrito4166]

i mean it's the ULTIMATE GUIDE so i gotta watch it lol

[@saarhadad]

tnx alot i learn alot from u.

[@BombXXplosive]

thank you for this, only issue for me is that lora's dont work for me i get this issue, how can i fix it? Patching LoRAs for KModel: 74%|███████████████████████████████████████▏ | 225/304 [00:34

[@pixaroma]

First of all from what i know they still have problems with lora, and a few other things not sure if they are fixed. then usually when say out memory is because you don't have enough VRAM on your video card. So is your card good enough? That could cause some problems. But also the version, try update every day and see if you can get a version that work, i see updates every hour adding or fixing something on the commit page like i showed in the video

[@BombXXplosive]

@@pixaroma thank you for the response, yeah i have the 12gb version. artwork generates fine without lora but when i try use a lora, it breaks haha

[@pixaroma]

@@BombXXplosive i saw some saying the flux lora doesn't work yet with forgez only normal lora with other models, so that might be the cause, you should be able to use lora at 12gb

[@BombXXplosive]

@@pixaroma ahhh damn ok!!! So I’ll need to use it on comfy then. Thank you

[@guitrindade-m3i]

Do you know how I can fix the error ValueError: Failed to recognize model type! my stable defusion is not working

[@pixaroma]

it still has a lot of bugs, maybe you can report it here github.com/lllyasviel/stable-diffusion-webui-forge/issues as you can see it has over 700 open issues, so not sure what model you use that is not recognize, if it work with other model must be a bug recognizing that model.

[@FoodInboxAU]

what model are you using for your voice over? is it f5? it sounds so real, I could barely tell it was AI at first!

[@pixaroma]

I use voiceAir, they have the voices from elevenlabs, so elvenlabs is the source

[@WizzardofOdds]

Thanks, can't wait to put an upgraded video card in my computer.

[@synthoelectro]

tried to get the 5.95 GB version of flux1-schnell-nf4 but it doesn't work, wonder why?

[@pixaroma]

Not sure, only the bigger size worked for me the one I used in the video, probably the will make it work eventually but not sure how smaller they can make it without loosing quality

[@synthoelectro]

@@pixaroma it's a mystery.

[@yDkayVal]

Huge amazing clear video, sub, like, everything bro.

[@pixaroma]

thank you 🙂

[@tuwonga]

excellent! What is the best commit to get back Train, SVD and Z123 ? thx

[@pixaroma]

I used the one that i showed in video starts with 29

[@tuwonga]

@@pixaroma thank you so much !

[@travisgumm5861]

Great video! Following it was nice and easy. My poor 4070 super isn't happy with it though lol

[@pixaroma]

I got it to work on rtx2060 6gb of vram also, 27 sec for schnell nf4 and 1 min 40 sec for dev nf4

[@travisgumm5861]

@@pixaroma turned out I still had automatic open. Silly me 😆

[@TomLally]

How is the 4070 super? I’m thinking of getting that one hmm. Thanks a bunch.

[@travisgumm5861]

@TomLally it's plenty fast enough. Images usually don't take more than a minute to produce. If using SDXL it's even faster than that.

[@TomLally]

@@travisgumm5861 amazing! Well right now I’m only using Fooocus. The whole comfyui is confusing haha. But that certainly sounds great

[@bart.k]

Great guide! Followed all steps but flux dev on default settings is crashing during model unload. Any ideas whats going on? 4070 super and 32gb of ram Thanks!

[@bart.k]

changing virtual memory paging file size from 16gb to 32gb seems to fix the issue!

[@pixaroma]

Glad you figured it out, sorry for the late reply it was night here. Did that happen with all versions of dev or you tired the full regular version

[@Aimkrad]

Спасибо за видео. Модель fp8 у меня генерирует на 20-40% дольше, но часто оно того стоит. Было бы здорово увидеть видео про лучшие апскейлеры для flux, так как у меня hires fix генерирует 1 картинку 5-6 минут.

[@pixaroma]

I rarely use forge, I switched to comfy UI since I can do more with it, the next episode will be about upscalers but with comfy. As for upscale speed depends on the model, the size of the image, the vram you have.

[@Aimkrad]

@@pixaromaчто вы имеете ввиду под удобным интерфейсом?

[@pixaroma]

@@Aimkrad ComfyUI this one kzbin.info/aero/PL-pohOSaL8P9kLZP8tQ1K1QWdZEgwiBM0

[@ericanderson4973]

Highly recommend a fresh forge install. updating my old version totally broke it.

[@pixaroma]

Yeah i just have different versions in different folders

[@ArcosGaming]

Indeed faced the same problem. My lora tab and checkpoints weren't showing anything other than error. Had to reinstall forge completely from scratch

[@kirill747]

thanks a bunch for the revert bat !! was having troubles with generation if the window lost focus, tabbed out etc, couldnt stop the monster!!!

[@sabirsafiullin5304]

Great video, thanks for what you do, keep up the good work!

[@pixaroma]

Thanks, will do ☺️

[@buda3d2007]

10/10 great tute!

[@VeryLoki]

when i installed it, it didn't download a checkpoint for me, no idea why. Luckily I already have some in another folder.

[@pixaroma]

Sometimes it happens might be with long path not being active in windows or was an error when trying to get it from hugging face or where was model hosted

[@RoodeMenon]

Thank you for this.

[@elefmaster]

Thank you for video, very usefull.

[@mechunz3741]

The great video, the great forge

[@brianp235]

Am I missing something, or did you forget to mention downloading the clip models and putting them in the correct location? I get errors as is. (Awesome video by the way, very helpful)

[@brianp235]

It seems my issue was using the wrong flux model that did not have VAE and clip baked in I think? I had downloaded one previously form another source (black forest labs not labeled as fp8). I got it to work using the fp8 model you mention here. Is there a way to specify the clip and VAE model so I can use another model?

[@pixaroma]

No, there are different kind of flux versions released. The first versions that need it clip the unet ones, then the other variation like fp8 and nf4 that had those clip included, and now the gguf versions that need again clip. When i did the video the interface couldn't even load the original ones with clip only vae, after that they added the clip version. So since i did the video they updated the interface like 100 times

[@vuhuy4386]

Thank you so much!

[@The-Yellow-Man]

I have an RTX4060 with 32GBs of Ram. Just curious if I should go with nf4 seeing as I have the specs, or would the quality still always be best with the nf8? Would there really be a notable difference?

[@pixaroma]

try both and see, depends on preferences, sometimes it is sometimes not so much

[@fkol4847]

Great vid! Do you know if controlnet works?

[@pixaroma]

Is possible to not work, i see they keep working on lora, hires fix also had some problems, each hour u see new updates so maybe look at the commits page to see what they fixed

[@jjjbeastfd]

can you make a run pod installation please

[@pixaroma]

I only use it locally, and I never used run pod so I don't know how to do that

[@SumoBundle]

Do you have also for mac ?

[@pixaroma]

No, only for windows, but check this out, seems other have installed github.com/lllyasviel/stable-diffusion-webui-forge/issues/314

[@SumoBundle]

@@pixaroma I will take a look. Thank you for the link.

[@imrabbit9950]

Hey can you give me info about the controlnet i should use with this ?

[@pixaroma]

For the flux models you should wait a little so the flux control net model are getting better, maybe they get a union version like on sdxl

[@stephenzadach1870]

Can you tell me approximately how long it should take to extract the files with winrar in the forge folder

[@pixaroma]

A few minutes depending on your system

[@stephenzadach1870]

@@pixaroma I just don't understand. I have a rtx 2080 ti graphics card.,ryzen 9, 32mgs ram. Do you have any idea what my next steps should be.

[@pixaroma]

@@stephenzadach1870 maybe try to download again and extract. Also is faster on a SSD, try on different folder maybe, I never had problem installing so not sure what could be the cause on your system

[@wrillywonka1320]

i was told that flux is what i was looking for to try to get a consistent character in the same pose but multiple different camera views. do you know if this is possible and why someone recommended flux?

[@pixaroma]

I saw in comfyui someone using it, not sure in forge, flux is better in prompt understanding so probably you can get more characters on one image in different angles only using good prompts. you can also check this for comfyui www.reddit.com/r/comfyui/comments/1elfcef/flux_consistent_character_sheet/

[@ORGANICMINDART]

Any chance running this on m1 mac ..

[@pixaroma]

Not sure, but check the comments of this post maybe you can make it work, github.com/lllyasviel/stable-diffusion-webui-forge/discussions/270

[@atajqtsawa3293]

I have an error "You do not have CLIP state dict". Followed everything you did. Please help.

[@pixaroma]

they changed some things in the interface, make sure you have the right clip models, check this post github.com/lllyasviel/stable-diffusion-webui-forge/discussions/1050

[@Kayeandre]

None of the LoRas I installed seem to work

[@pixaroma]

I think lora for nf4 flux doesn't work yet, i saw normal lora works if you have sdxl but for nf4 saw some comments that still working on them, so maybe update from time to time to see if they fixed

[@raymondandreaswilke8176]

the most problem i have with restriktions i want to yous this for Books, Gameplot/story books. so i need to generate fighting scenes as well...

[@pixaroma]

Train maybe a lora with the types of images you want to generate, or search for different models on civitai

[@fatallyfresh3932]

I am so confused, I had no idea Forge had an update or added Flux support... Last I heard this project was abandoned and he told everyone he will no longer provide updates because he was working on something else. This came out of no where? Please more videos about Flux/Forge and cool features. I want to to learn everything! xD Btw, I just tried the nf4 version and with my 3060 gpu its taking like 2mins to generate images... Is this normal? It seems incredibly slow for my hardware.

[@pixaroma]

Flux takes a lot of resources, depending on the flux version, and your vram and ram. Depends also on the t5 clip size, etc. I only use comfyui since forge is so unstable. But it can take from 15sec to a few minutes

[@billyn7757]

Thank you for this! But im having trouble with Forge Ui generating images with hires.fx. For some reason, i can only generate images with hires.fx upscale 1.5 and below or 2.1. Everything else is a brown blank image. The same happens when i use any SDXL Pony images. I have a rtx 4090. Please let me know if there is a solution, if not, it's not a big deal since i can still generate images fine in these weird hires.fx upscale. Thank you.

[@pixaroma]

I noticed too that doesn't work, probably on feature updates, eventually look on commits if some new one appears with fixing that. But I rarely use hiresfix i just send it to img2img and use Denise 0.35 and make the same ratio but bigger like 1800 instead of 1920*1080 so i can get a bigger image similar to hires fix. I also can do inpaint or edit in Photoshop and paste again in img to img and get simila bigger image just like i did on that girl with red hair in the video

[@billyn7757]

@@pixaroma Ok thank you, yea i normally use hires.fx because like for anime images which i tend to make, the faces almost all the time gets pixelated and over all doesn't look smooth especially the eyes. But Thank you, i'll try using img2img.

[@MindCaged]

Sadly I'm unable to use this. I'm running on an AMD Mini PC with APU and I can barely run an AMD fork of Automatic1111's webui version. I was able to finally get Forge to launch, but it keeps throwing errors and crashing when I try to generate images with Flux, even low res ones. It's really disappointing since it seems like a huge chunk of the models online are Flux based, and I've tried running those Lora models on a SD checkpoint and it just doesn't work right, /maybe/ I can generate something that looks like the intended subjects cousin, but just can't get close enough. The SD loras get close enough you can recognize the face, but they don't nail it like making it look like a real photo.

[@pixaroma]

Amd usually has a lot of problems, need Nvidia most of the Ai. Also for flux even if you get it to work that still need a lot of VRAM

[@MindCaged]

@@pixaroma Frankly, even the regular SD generation was crashing compared to WebUI. I think it's because the automatic memory management wasn't exactly perfect, or there were glitches due to trying to run an AMD fork. Had trouble even getting it to run, and it was throwing a bunch of errors.

[@Fayrus_Fuma]

And what do we lose in the new version of 1111? I'm sitting on the June version before the update broke the program. Without the 1111 update, Flux won't work?

[@pixaroma]

Not sure about 1111 this was for forge, i didn't use a1111 since the forge appeared. So the new version keeps updating every hour or sometimes minutes so if you look at commits you can see what has been changed and when. The ones with verified are more stable so you can keep trying different commits until you get one that works. I have more versions installed one that i test updates, and one that is more stable that works for what i need. It was pretty unstable in the last month and they still working on it. So i am using comfy UI since is more stable and forge for when i need something quick like an image to image or something

[@pixaroma]

Without the update for 2-3 days ago flux doesn't work, that is why i said to have a different version in a separate folder so you can use flux

[@Fayrus_Fuma]

@@pixaroma Can I keep the 1111 in two different folders? So I can have two versions? When the game files are updated, are they updated only in these folders or somewhere else on the system? (C drive for example)

[@pixaroma]

​@Fayrus_Fuma You can have many in different folders. I have 3 forge ui, 3 comfy ui and one a1111 that i didn't use since i had forge . It updates in the folder it is in.

[@treewx]

Thank you!! Nice and easy to follow.. got it working

[@ranucho]

for some reason, on auto1111, all 3 models generate nothing but noise. What am I missing?

[@pixaroma]

I dont think they made it work yet with flux, they need to update it to work, i didnt use a1111 for a long time since with forge

[@ranucho]

@@pixaroma I did install Forge and it works great! thank you!!

[@dziku2222]

Doesn't work "ModuleNotFoundError: No module named 'git'" or "unterminated triple-quoted string literal (detected at line 236)"

[@pixaroma]

try maybe to install git to see if helps git-scm.com/downloads

[@amkkart]

Does it work with deforum ? Or animate diff?

[@pixaroma]

I didnt try it, but there are still things that doesnt work, if you look at commits like every hours or even minutes there is a new version, so they keep fixing and adding new stuff, eventually all will work but probably takes some time github.com/lllyasviel/stable-diffusion-webui-forge/commits/main/

[@amkkart]

@pixaroma thanks man.. I'll wait Abitur more before i install it.. Thanks for the nice videos !

[@baag78]

is possible to run "ownload Flux 1 Schnell bnb nf4 from here (Medium Quality - Fast)" with it 8750h 16ram and 1050ti 4gb ? THANKS!

[@pixaroma]

Form what i saw online people have problems even with smaller models sometimes like sdxl, since the ram is only 4gb. And flux need more, so even it you run it probably takes a lot of time. NF4 seems that is being replaced by other, gguf, I didn't make a video on that yet but if you search gguf flux you can find some tutorials, they have release some smaller flux models gguf, q2, q3 and so on, q2 being faster but lower quality, and high you go like q8 better the quality but slower to generate

[@JieTie]

Maybe next vid on how to train custom lora for flux? :D Locally :)

[@pixaroma]

If I succeed in doing that i will make a video, but i tried also in the past with sdxl and it was hit and miss. I will see in the future if there are easier way to do it since training is more advanced than my designs skills 😁

[@ScytheSalinas]

​@@pixaromaif you want to brainstorm we can chat about it, I also want to do this