How to Upscale with FLUX. Comfy Workflow.

[@TransformXRED]

I really appreciate the fact that you're giving your workflows away for free. Yesterday I deleted the folder of my Comfy install, forgetting that I had a Workflow folder in it, with everything gathered over the past months and the ones I did myself. I've made the mistake of clicking on update_comfyui_and_python_dependencies instead of doing the proper update. It messed up my whole install. Krita and the stable diffusion plugin stopped working because of a flash attention error. Anyway, I nuked the folder, but before that, I made a second install from scratch in a new folder. And deleted the old one, forgetting to move all the workflow files. I'm glad that we have access to almost all of your workflows, thank you for that. PSA for anyone reading this. Never keep models, checkpoints, and the workflows in the Comfy folder (or auto1111 for that matter). Have the most important folder on a separate disk or dedicated folder outside of the install folders. And you like them with symlinks. By doing that, a git pull or deleting the installation folder of a1111, my mistake or not, will not delete the important files.

[@sebastiankamph]

Glad to hear you're enjoying the content! Too bad about your mess there 😅

[@klb1113]

As @TransformXRED said, it's really nice of you to provide these workflows to the community at large rather than behind a Patreon paywall. (Which I will definitely consider joining....being a fellow lover of "Dad jokes"!) I usually am able to run (barely) the f.8 model, as I have VRAM of 12GB, so hopefully this will work for me without completely shutting down my machine!

[@Hermila-v2p]

dopepics AI fixes this. Upscale with FLUX: Comfy Workflow.

[@RandMpinkFilms]

Amazing! Thank you for this fantastic workflow. So greatly valued!!

[@RoguishlyHandsome]

I have been simply using a separate workflow that takes an image as input, then upscaled and converted to latent. So copy-paste from one workflow to the other. Works pretty well if you're just looking to improve quality and upscale very lightly.

[@jroc6745]

Would you mind sharing that workflow?

[@cyril1111]

You can also use a SDXL tiled upscaled nodes after the Flux image is generated - Better results for upscaling

[@Master_Z2]

Thanks for the new video Sebastian, i have learnt a lot from you friend. God bless you and your family , thanks for your time making these videos it helps a lot.😊

[@mashedpotatoes7068]

OMG, these FLUX tutorials are so great! For Ultimate SD Upscale I tried height width 512 instead of 1024 that might get better results! Anyone else tried this?

[@BelgianGurista]

You need to keep within flux resolutions, otherwise you get artifacts. Google: Aplakka flux resolutions. 2.0 MP (Flux maximum) 1:1 exact 1448 x 1448, rounded 1408 x 1408 3:2 exact 1773 x 1182, rounded 1728 x 1152 4:3 exact 1672 x 1254, rounded 1664 x 1216 16:9 exact 1936 x 1089, rounded 1920 x 1088 21:9 exact 2212 x 948, rounded 2176 x 960 1.0 MP 1:1 exact 1024 x 1024 3:2 exact 1254 x 836, rounded 1216 x 832 4:3 exact 1182 x 887, rounded 1152 x 896 16:9 exact 1365 x 768, rounded 1344 x 768 21:9 exact 1564 x 670, rounded 1536 x 640 Rounded means the resolution you should set it in ComfyUi.

[@hunkrid]

In my experience, the fewer tiles, the better, so I set the tile size to the same resolution of the source image.

[@mashedpotatoes7068]

@@hunkrid Thanks for info!

[@IneKeee]

Hello! First of all, thank you so much for your videos AND your written tutorials (I usually prefer reading to go over it a couple more times) This is so helpful and I appreciate the time you put into everything! My question: I did notice that negative prompts are not used in your workflow. Is this because it is not a thing with Flux or just because you dont use them?

[@sebastiankamph]

Per default Flux doesn't use negative promps. Some users have found ways to force negatives but I'd start using it without them first.

[@damarcta]

thanks great video. what i like to do is, setting the tile size higher so that the image size isnt a multiple of the tile size (i.e. for 2048x2048 images something like 1256x1256 tiles) to create an overlap (or increate the padding option, but this doesnt work as good). most of the time the blending of the node removes the seams (if the denoise isnt insanly high).

[@techviking23]

the tattoo on the face is insanely good

[@LikquidDutch]

works pretty decent on my 3080. The NF4 model made a huge difference in speed rendering time for me like 80secs on average on 1024x1024 without upscale with upscale it added like another 3 mins thank you .

[@clarksoft]

have you trying SIMPLE schedule and only 2 steps? it's work faster for me.

[@boo3321]

3080 10gb can run fp8 60~70s,schnell 13s if you use unet version with 64gb+ ram.

[@murphylanga]

Thank you for your videos, I've found the crucial tip here again. You're great!🥰

[@sebastiankamph]

You are so welcome!

[@Invictus_Wisdom]

Hey Seba thanks for new video do u prefer comfyUI or forge?

[@sebastiankamph]

Both! Depends on what I need to do.

[@yu-gg]

hey Sebastian thanks for the nice tutorial! For some reason NF4 is 3 times slower than Flux Schnell, do you maybe have something in mind i couldve done wrong?

[@jakkalsvibes]

Thank you for all these detailed explanations! Would it be possible to do a video about creating consistent characters using Flux? I'd love to see your insights on that.

[@electrolab2624]

Installed NF4 + Had a problem finding the node 'CheckpointLoaderNF4' - SOLUTION: In 'ComfyUI Manager' (main screen) - change 'Channel' settings to 'dev'. NOTE: I did NOT need to change the security to 'weak' as some people had suggested (In some .ini file) Thanks for all you do! - cheers!

[@CMDRRustyDog]

security level goes to 'normal-'. I posted a solution to help people but I think Sebastian might have removed it, which was nice.

[@electrolab2624]

​@@CMDRRustyDog I did not change any ComfyUI security level, only changed the manager settings to 'dev' and got the node to work - perhaps others will need that solution too - Neat that many ppl can help each other! BTW: *Instead of Upscaling, I use 'FaceRestoreCFwithModel' Node, and 'codeformer' as a model - (noticed that it's mainly faces which it does not do perfectly) - Thank you, COMRDogRusty !

[@AyuK-jm1qo]

I still have the problem, When loading the graph, the following node types were not found: CheckpointLoaderNF4

[@Tobygroberts]

@@AyuK-jm1qo I also have this issue

[@randymonteith1660]

@@AyuK-jm1qo try shutting down ComfyUI and restarting it again. This worked for me

[@supernaturalsubcontinent6282]

Need a Tutorial for ControlNet specially in Thinkdiffusion as i can't seem to install custom nodes there

[@perfectionatic]

Great tutorial! I am interested on your take on doing a latent image upscale, followed by a sampler. I know that beyond a point you will hit memory issues, but before you get that point, wouldn't it be better than ultimate upscale?

[@Renzsu]

Ohh, I do like those performance bars you have in your main menu.. what add-on is that? Time to give these models a go later, thanks for the walkthrough!

[@sebastiankamph]

Crystools :)

[@Renzsu]

@@sebastiankamph Sweet, thank you!

[@_FutureFocus]

Amazingly easy to follow tutorial as always! Noob question. Two in fact. 1: Would it be possible to upload an image to this workflow rather than generate one with text, if so how? and 2: What if you wanted to upscale multiple times the way that was shown here but only realised after you had upscaled once already? What I mean is could I add a second upscaler and THEN have it continue the process of upscaling that existing image or would I need to start the whole process again (essentially losing that image) by generating a new one?

[@sebastiankamph]

1. Load image node, send to upscaler node. 2. Fixed seed

[@_FutureFocus]

@@sebastiankamph Wow! Realtime support! I'm hitting you on patron for this. Tha KS so much man.

[@sebastiankamph]

@@_FutureFocus that's very kind, thank you 🥰

[@Asan-b5o]

can i use Sunnel model ??

[@baddealrage]

Thank you for the insight your tips help us to to stay cutting edge! :)

[@swannschilling474]

Thanks so for keeping us updated so fast ❤

[@swannschilling474]

One thing that came to my mind is that rather than leaving the upscale prompt blank, you could also have Florence2 describe the oart of the image? 😊

[@BlueQualityRhythm]

Why are there no overlapping tiles? That would solve the seams issue.

[@therookiesplaybook]

How are you getting the live preview in the SamplerCustomAdvanced window? And how are you getting the progress bars in the Queue Prompt?

[@runebinder]

For the live preview go into Manager and on the top left there's a drop down called Preview Method, choose one of the options on there. I could be wrong but I think the Progress bar is part of the Crystools Custom Node as it shows system usage as part of that as well,

[@therookiesplaybook]

@@runebinder Awesome thanks.

[@vjunicorndreams]

Thanks Sebastian for what you publish in your channel. One question, if you just download “Flux Upscale” to the PC, can it run as a separate module without having Flux installed? Is there a link to run “Flux Upscale” online without downloading it? Thanks.

[@sebastiankamph]

You can run the workflow in thinkdiffusion online

[@vjunicorndreams]

@@sebastiankamph Thanks!

[@oceaco]

Thank you, let's try this

[@sebastiankamph]

Have fun!

[@beyond.matter]

Saweet! Thank you! How do we go about img to upscaler with flux?

[@IIStaffyII]

Dropping Flux videos back to back to back now are we? Thank you for the hard work Seb!

[@sebastiankamph]

You bet! 💫

[@hotlineoperator]

Some say that upscale is no need in Flux - you can somehow render image in multiple parts, like bucket render ??

[@giovannifama1089]

Hi Sebastian, thanks for your videos! Can you please make a tutorial on flux face swap? I don't even know if it's a thing since flux is so new. Thanks in advance

[@icchansan]

Hi Sebastian, love your stuff! would u consider making a tuto about loras in flux?

[@2008spoonman]

The nf4 model does not support lora’s.

[@erans]

Anyone have any idea about this error? "ValueError: not enough values to unpack (expected 2, got 1)" I tried connecting it to the flux model after a facedetailer node

[@techviking23]

nice video!

[@sebastiankamph]

Thank you! Cheers!

[@Andro-Meta]

Wild. I cannot get ComfyUI to update and run Flux models at all in Think Diffusion. Kind of frustrated with Think Diffusion because of it.

[@titanium5]

Truely awesome. Thank you SK! I love your videos. Sub / Like!

[@sebastiankamph]

Thanks for the sub! And welcome aboard! 💫

[@MysteryFinery]

@Sebastian Kamph if im new to comfey what vid do I watch first?

[@sebastiankamph]

I got a how to install and use comfy somewhere

[@kamruzzamanuzzal3764]

it's better to use an SD1.5 varient model instead of flux for the ultimate upscaling, gives better colors. just a tip

[@sebastiankamph]

Where's the fun in that. Also, I did an hour long comparison videon on Patreon and SD3 beat both 1.5 and XL for upscaling.

[@kamruzzamanuzzal3764]

@@sebastiankamph lol true. As for sd3, is it base sd3 or a fine tuned one? Cause I tried it with base sd3 (when it first came out) and compared it with my fav sd1.5 (with some embeddings), and to me it felt like sd1.5 was better. Since then I didn't feel like getting any news on sd3 (I also lost my hope on new open source models until flux came out)

[@DaveTheAIMad]

Did they fix the contrast issue in sd upscale? I stopped using it a while back as all my upscaled images came out slightly whiter than the image input, like the contrast was tweaked and washed the colours a bit.

[@scarletpenn]

put into the folder like your guide said but when I load json file it says "When loading the graph, the following node types were not found: CheckpointLoaderNF4 UltimateSDUpscale Nodes that have failed to load will show as red on the graph." I also cannot update all on comfy manager, it gives an error message saying it cannot update, any idea whats going on?

[@AILOFICHILLMUSIC]

May i know NF4 can use Lora ?

[@klb1113]

Ooops...forgot to ask. I tried loading my upscaler into an Models/ERSGAN folder, but it's not showing up in the interface. Any ideas? Am I putting it in the wrong place?

[@ВладимирБондарь-т8ь]

Hello! Tell me how to load the node "CheckpointLoaderNF4"

[@eddiej.l.christian6754]

There is no Checkpoint loader NF4.. Where do I find this????

[@loquacious1956]

I am also trying to find it

[@schert5648]

@@loquacious1956 same

[@RoguishlyHandsome]

Something changed in the last couple of days. I can't use the nf4 models in comfyui on my rtx 2080 super 8GB anymore. Getting only OOM, even with --lowvram. So much for this being the "low vram" model, while I can still use the fp8 ones, but you know, they are freaking slow on a GPU like this.

[@AncientShinrinYoku]

I have the same problem on 1060 3GB VRAM: torch.OutOfMemoryError: Allocation on device It looks like SamplerCustomAdvanced might be causing problems. Any way to replace it with another node?🤔

[@RoguishlyHandsome]

4x ultrasharp makes plasticky looks worse, also. Removes details, smooths everything into oblivion.

[@SpikeColgate]

Any suggestions for an alternative?

[@RoguishlyHandsome]

@@SpikeColgate I use the siax one we see in this video. I could not tell you which one is best. I just know that it bring a lot more details than 4x ultrasharp.

[@marhensa]

use 4xFaceUpSharpDAT, that's for sure the best for removing plastic-y looks.

[@marhensa]

@@RoguishlyHandsome 4xFaceUpSharpDAT, very much better than plastic looking 4xUltraSharp. Google it.

[@kfayayay2674]

So as i understand, this is for upscaling a txt2img, not for img2img. Any way to add an img loader ?

[@sebastiankamph]

Just add a 'Load image' node and connect it to the upscaler

[@madmed6672]

Does Flux with any resolution or ?

[@Renzsu]

I must be doing something wrong. I updated Comfy, but when I load your workflow, the checkpoint loader is red.. can't install it as a missing node either. Then when I replace it with a regular checkpoint loader, I get a bunch of errors when I try to run it. I put the model in /models/checkpoints, although my fp8 schnell and dev are in /models/unet. Putting it in the unet folder and using the fp8 model workflows to load the model also gives me errors. Pretty sure I downloaded the whole model without problems as well.

[@themachine8229]

The same for me

[@KevinFlynn1967]

Same here, trying to figure out how to install the node.

[@2008spoonman]

Change the channel from default to dev via Comfyui Manager.

[@gurinolegg]

Doesn't work. problem finding and installing the node 'CheckpointLoaderNF4' "dev" mode doesn't help

[@m_sha3er]

This should be fast than dev or what the difference? I found that dev fast than this one for sure without upscale!

[@dushan6588]

Please help. Your all videos got same description and links. I need to try this as well. Please tell me or can you send me link. What is different comfy ui and pinokio version what i need to install. Please refer your install video of comfy. Please

[@mal-avcisi9783]

hey bro, what do i have to do, to use this for img to img upscale ? can you give me some tips ?

[@sebastiankamph]

Use a load image node that goes into the upscale instead.

[@FinalJazon]

does this work in Forge ?

[@sorijin]

So I couldn't get the flux LORAs working with the v1 but v2 they seem to be working fine or am I crazy? --edit: apparently they just don't do anything atm

[@Dingle.Donger]

So what's the verdict on NF4? Is it better than fp8? I have a 3090 so vram isn't an issue for me.

[@sebastiankamph]

I prefer nf4. Unless you want to go dev fp16

[@hououinkyouma8233]

which is better ultimate SD or SUPIR ?

[@sebastiankamph]

Supir is really really good but harder to use.

[@laeternaml]

Quite new to comfyui here, anybody knows how to control the image generator and the upscaler separately? Maybe I don't understand how this is supposed to work but I would love to just generate images and trigger the upscaler on that specific image when I want. Currently it is running the full process on every iterations so it takes quite some time considering not all results deserve an upscale. Any tips welcome :)

[@benyaminaria283]

how to add image ?

[@TheGalacticIndian]

Thank you, Sebastian!🎖🎖

[@sebastiankamph]

Happy to help!

[@ana-ot3is]

hey Sebastian, can you please tell me how to install missing CheckpointLoader NF4? I tried with the regular loader but it doesnt work. Thank you! edit: guys, you can check Rusty dog tut on how to solve this, it worked for me : )

[@rawpotatofella9654]

Will the v2 model also work in forge??

[@leonxu7575]

yes

[@jasonstetsonofficial]

How to do it in Forge?

[@AtheistExpert]

Hey does anyone know if we can use Loras with Fast Flux yet? .... I havent seen a workflow, and flux1-dev on 16gb is kinda slow for me.

[@evilshadows3870]

Hey! Can you do a workflow of Inpaint for comfy?

[@sebastiankamph]

I guess!

[@markusblandus]

Thanks Seb!

[@AgustinCaniglia1992]

Flux loras arent working on Comfy, right?

[@2008spoonman]

Nf4 model, no support for lora’s. Dev and schnell models do support lora’s.

[@KevinFlynn1967]

I had to go into my config file and change the security_level to = weak to install the checkpointloaderNF4. Currently running at security_level = normal-. Sry guys, I new to all this lol

[@schert5648]

where did you get the chekpointloader?

[@walidflux]

it's really sucks that NF4 doesn't supports loras, i hope they fix that because this can't go any further

[@NotThatOlivia]

go for NOMOS family of upscale models instead of old ultrasharp!!!

[@sebastiankamph]

Isn't that a watch brand

[@ESheridan]

Hello there! Thanks for your video, it's really nice that you share your experience, but sometimes, people like me (noooob) who follow your instructions don't understand what to do and how to do it. Where to get this or that node from and so on... Yes I am a nooob and I really want to understand all your manipulations. Thank you very much! P.S. To run your upscale did not work, because there are nuances and you did not say a word about them. It's a pity

[@EvgenyCh-th8dc]

можно стразу full HD генерировать и не возить с апскейлером

[@rafedalwani]

You know all the AI videos that ask you to pay to get the work flow result in subscription

[@59Marcel]

Thanks!

[@sebastiankamph]

Thank you so much, very kind! 😊💫

[@59Marcel]

That's ok. I appreciate the work you put into your tutorials and then to offer the work flows for free is so kid.

[@zaferustuner]

definitely follow Sebastian's workflow but put guidance to 7 and basic scheduler Steps to 50 for the clearest/sharpest image.

[@kenrock2]

can you do one for lora?

[@Joie-l1r]

dopepics AI fixes this. Upscale with FLUX: Comfy Workflow.