NEW Flux NF4 for ComfyUI. FASTEST FLUX!
Вставка
- Опубліковано 2 жов 2024
- Flux tutorial guide / flux-how-to-109332332
Flux original full steps video: • NEW BEST AI MODEL! How...
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...
Those who are getting CheckpointloaderNF4 error
-First follow all the instructions of the video.
- after that edit the config.ini , edit "security_level = normal" to "security_level = weak" and then Open comfyUI manager and on the left side Set Channel : default to dev, then restart your comfyUI and open COmfyUI manager then select install custom nodes then search "checkpointloader" there you can Find the NF4 press "try fix" button and it will turn your red NF4 to normal
Tried try fix - not helping, it is still red. Ofc, i restarted browser, server, and it does not help.
@@nesharm did you select the the channel to dev, and install the nf4 from install custom nodes in confyui manager
It WORK !!!!!
thanks
Thank you. I also noticed that he posts the full instructions in the description 'Flux tutorial guide / flux-how-to-109332332'. Wish he would tell you that in the video.
When loading the graph, the following node types were not found:
CheckpointLoaderNF4
Nodes that have failed to load will show as red on the graph.
follow the instruction in this video at timestamp 3:00
@khalifarmili1256 I did fix it the thing that was missing is changing the Chanel to dev
@@Drazzzy Thanks, this info make my day
There are so many issues not mentioned here. Firstly, installing ComfyUI via the Windows portable method, means the NF4 loader will install but not function. Installing ComfyUI manually will do it, but leads to errors elsewhere. It not as straightforward as it's made out to be.
Thanks, i have portable installed version - so that might explain my my NF4 Loader will not appear for love nor money- guess I have to pass on this...
@@v11cu96 u must do it manuały - look at 3:00 but line "pip install -r required.txt" substitute by: " ..\..\..\python_embeded\python.exe -s -m pip install -r requirements.txt"
It worked when i installed with manual but i also needed to install numpy which also meant i needed to lower the security of the comfy manager. More steps but wasn't too difficult.
@@IIStaffyII How on God's green Earth do you lower the security on CM please.
@@KINGLIFERISM You need to edit the config.ini in the comfy ui manager addon in ComfyUI\custom_nodes\ComfyUI-Manager. Add a line called security_level = weak for example. I would wait for an official version though.
This is great for anyone who can't run fp16.
In comparison to fp8 it's supposed to be similar or even better quality because NF4 is a much better format than fp8. Ilyasviel explained why on Forge discussion thread #981. It's also the most lightweight variety yet by far for anyone with a GPU new enough to support NF4. (Which mostly means RTX 3000 or 4000 series.)
No 2000 series like my RTX 2080? Bummer. If it is floating point 4 the 2080 series should support it. Actually, a lot of older Gpus should support it too... Unless I am missing something.
@@MegaGasek I haven't got a clue. I'm just echoing what's written by Ilyasviel, who know about a million times more than I do.
Great post. Thank you.
From a lot of reddit posts the NF4 has a bit lower quality and "logic" than fp8. But it is a great way to find a good seed at a faster speed and then use a slower, better version for the actual generation.
Is flux-dev 16bit, flux-schnell 8bit and this one 4bit?
As a Automatic1111 user, ComfyUI confused the heck out of me. I found Forge which is almost the same as Automatic1111 but works with Flux models.
For anyone having issues - to get the NF4 Loader to show up, I had to change the config.ini file for the Comfyui-Manager to have weak security. (Do so at your own risk!) Then change the channel to dev. It should then be searchable in the custom nodes section. I then installed and restarted...now working fine, so that could work for others.
what changes did you make to the config.ini file? I tried, but it didn't work
@@lovelyfunnytunes2051 There should be a line in it, security_level = normal...I changed this to security_level = weak. Then once I had got the extension and all was working I changed it back.
@@lovelyfunnytunes2051 I changed security_level = weak
do u use comfy portable?
@@hagemaru96 yeah, I’ve got the portable version 👍🏼
Anyone has the same issue and solution? Thanks for your feedback.
## ComfyUI-Manager: EXECUTE => ['/ComfyUI/venv/bin/python3', '-m', 'pip', 'install', 'bitsandbytes>=0.43.0']
ERROR: Could not find a version that satisfies the requirement bitsandbytes>=0.43.0 (from versions: 0.31.8, 0.32.0, 0.32.1, 0.32.2, 0.32.3, 0.33.0, 0.33.1, 0.34.0, 0.35.0, 0.35.1, 0.35.2, 0.35.3, 0.35.4, 0.36.0, 0.36.0.post1, 0.36.0.post2, 0.37.0, 0.37.1, 0.37.2, 0.38.0, 0.38.0.post1, 0.38.0.post2, 0.38.1, 0.39.0, 0.39.1, 0.40.0, 0.40.0.post1, 0.40.0.post2, 0.40.0.post3, 0.40.0.post4, 0.40.1, 0.40.1.post1, 0.40.2, 0.41.0, 0.41.1, 0.41.2, 0.41.2.post1, 0.41.2.post2, 0.41.3, 0.41.3.post1, 0.41.3.post2, 0.42.0)
ERROR: No matching distribution found for bitsandbytes>=0.43.0
I come here for the dad jokes, then for your news and tutorials.
2x dad jokes today!
it Requires installing bitsandbytes.
Make sure your ComfyUI is updated.
i installed the bitsandbytes nf4 and the requirements.txt inside it before updating the comfyui, but the CheckpointLoaderNF4 just won't appear from the search bar and if I use your workflow, it'll mark red the supposed node because my comfyui can't find it :(
@@ridhosiregar138 for me just showed after I change the manager the channel options to "dev" and then will show in missing custom nodes but I have error for some reason, I pip install bitsandbytes but doesn't help.
@@ridhosiregar138 add new node named NF4 and use comfy manger to install custom nodes
@@ridhosiregar138 I actually faced problem with the NF4 loader, the custom node import failed, later I solved it manually. I will be making a video about this soon and upload on my channel!
You didn't explain that the comfy manager must be installed first and that we need another comfy node..
Yeah how download manger
Can some one just please share this workflow?
Updated all, NF4 not showing. Used git clone into custom nodes, downloaded, still not working
same for me, custom node isn't showing and I've installed everything
Not showing too after updating comfy 😕
same😶🌫
I haven't tried cloning anything yet but like you, Update All didn't cause the node to show. Think I'll likely try Forge next before doing any more with Comfy
Same.
we need Flux Lightning
That's what I was expecting when I saw the title
Theres flux schnell which is a fast version
Maybe add a pixel quality comparison with Dev and Schnell?
there is nf4 for bith of them dev and schnell and even bnb and normal one
dev is way better but has a non comercial license
ive updated everything but still no NF4 coming up
when i do it via git, (This action is not allowed with this security level configuration.)
Same deal, no idea what to do about the security issue. Msybe a case of running forge in admin mode though.
same
go to the manager tab and click custom nodes manager, there should be TRY FIX in there... then restart
@@AnotherPlace youre a genius man !
@@AnotherPlace Even this isn't working for me :(
Thanks for this Sebastian. Great content!
Checkpoint loader NF4 doesn't show up after my update.
You need bitsandbytes first
@@johnedwards7655 got it work finally, I had to do some mental gymnastics
same here.but it was added to the tutorial the following:
TROUBLESHOOTING - NF4 Not loading after all steps completed:
Double check you have installed bitsandbytes (pip install -r requirements.txt)
Go into the folder for the Manager, edit the config.ini and change the security setting from 'normal' to 'weak'. Run Comfy, go into the manager, change the channel to 'dev' then search for the node 'NF4'.
@@Emperor.of.Mankind I confirm this works
@@Emperor.of.Mankind I tried that but that also didnt' work. Finally I got it working and It's even slower than before, so I went back to dev.
I have rtx3060ti, did as the video and copied his workflow on his website, i still get 53s/it. What am i doing wrong?
He got a 4090
@@MultiOmega1911 you missed the point, i am not trying to get 0.8s/it, but my current speed is 3 times slower than regular than regular flux dev. This method is supposed to be faster than dev, not slower. No matter the hardware
When u ask questions on the internet leave out useless details cuz dummys will reply to those. Instead. Idk tho
@@zapfska7390 what do you mean by useless info? I only provided relevant info
sane for me, no noticeable difference of rendering between regular flux or NF4 flux.
Nothink work. Read comments - stopped trying :\
I actually faced problem with the NF4 loader, the custom node import failed, later I solved it manually. I will be making a video about this soon and upload on my channel!
@MentorAi-x wow, the ad in the comments. Anyway, I hope you help. My device shouldn't support the SDXL model. But it works. I'm just wondering if Flux will do the same.
@@alexblrus9825 Done
Thank you!
But how do you open the Search window, to search for the NF4 Loader ?
EDIT: I got it, just double-clicking on the board opens this search window
click two times at the back
Double click
For those that have the error of 'ForgeParams4bits' object has no attribute 'module', follow these instructuions:
In manager custom node folder, change the security leverl to weak in config.ini file.
Restart comfyui, go to manager, in the install PIP packages option (a button on the left), type exactly the following: bitsandbytes>=0.43
In my case, the package has been updated from 0.43.2 to 0.43.3. This minimal update does the trick. Then restart and it will work!
Change the config.ini security level to normal again. ;-)
Thank you my friend. After endless attempts, finally a solution. 👍
Thanks a lot. in my case this solved the problem: AttributeError: 'ForgeParams4bit' object has no attribute 'quant_storage'.
THANK YOU! After a ton of failed searches, pip install fails, git updates, Comfy updates, etc., this simple fix did the trick. You sir, deserve many thanks.
thank you, it worked like magic :)....love to you
I'm able to generate images with the new NF4 model in ComfyUI, but it's actually taking 1 second longer compared to the fp8 model. Any ideas?
I have the same issue: Tried it on SD-WebUI Forge and did same time or just 1-2 seconds longer. I also want to know what is the thing there.
maybe the number of steps? are you using the same in both?
No NF4 Loader herem how to fix???
You might need to change the Channel to DEV. Find this in the right side of the Comfy Manager.
@@Bicyclesidewalk i did that and restarted the comfyui, still with the red undefined node..
@@ridhosiregar138 you need to open up config.ini in custom_nodes/comfyui-manager folder and change security_level from "normal" to "weak"
You need comfyUI manager, you install it in custom_nodes, you need to instal bitsandbytes, then in comfy/custom_nodes/ComfyUI-Manager/config.ini set security to weak. After update and restart set channel to dev and you will be able to find nf4 node in custom nodes.
I actually faced problem with the NF4 loader, the custom node import failed, later I solved it manually. I will be making a video about this soon!
Is there a checkpoint loader for NF4 that will work on Apple Silicon? CheckpointLoaderNF4 relies on a newer version of bitsandbytes that is not yet supported on Mac.
Installed the nodes and the requirements just as the video shows but still getting error after running the workflow.
Error occurred when executing SamplerCustomAdvanced:
'ForgeParams4bit' object has no attribute 'quant_storage'
Please help.
I'm using Stability Matrix to run comfy.
Loras don't work on Flux NF4, currently.
can you do a comparison video of base flux dev, base flux s, the fp8 versions of those fluxs, and the new nf4 version? of the same prompt and fixed seed, just comparing a side-by-side or xy plot type. it would be very beneficial for many of us
how do I add negative promts to the base flux workflow?
flux isn't really compatible with negative prompts since it's always meant to be at CFG 1, people are working on workarounds but nothing great yet
I finally got it all working. But! why so long? 1h.30min 600*600 maybe the connection nodes are wrong?
Took me days, but i managed to make it work. The only thing i do not understand is how you make it run without a connection from the checkpoint to the Basic Scheduler Model. It doesn't work for me unless i link those 2. Still on my potato machine (8 gb gdu and 16 ram ) it takes 4 minutes to generate 4 steps 1024 x 1024. So a bit too long
Did I do something wrong? Takes more than 10 min to generate. I have a 4060TI with 8GB vram. I was expecting maybe 2 mins tops.
thanks for mention the forge the recreate all of ui and make it so much better and flexible and they work on it even now, every hour you find new update.
i have a question what is different between nf4 and bnb nf4 ?
Another Flux video: GGUF Flux models. How do they stack up to NF4? They apparently run in 8gb of VRAM.
Anychance to get workflow ?
It's in the guide in the description
@@sebastiankamph Yeah i saw it. Thank you
after 2 hours of work, because with comfy ui everything is complicated and basically nothing works right away, i finally got this model to run and it is just as slow as the normal one, 80 seconds per step (RTX 2070 Super 8GB)
Even though I beleive I followed all the steps I get an error on the first generation :
Error occurred when executing SamplerCustomAdvanced:
Attempted to call `variable.set_data(tensor)`, but `variable` and `tensor` have incompatible tensor type.
...
return t.to(
File "H:\Simulations\AI\ComfyUI-DirectML\custom_nodes\ComfyUI_bitsandbytes_NF4\__init__.py", line 69, in to
self.data = n.data
Unusable for me unfortunately, it takes like 1'40"/it on my 3050 ti 4gb, but at least the model has runned on it.
does it work on AMD rocm ?
Correct me if I'm wrong, but nf4 was replaced by GGUF tech.
"Potato PC" = any PC with less than 12GB of vRAM 😢
This is getting out of control.
Does this need NVIDIA gpu? I am using intel without nvidia, and it gives error or needs nvidia driver
for me everything shows up but when i try generating, the status bar doesn't change. it stays at 0% without any error. Does anyone have any ideas?
50seconds is a bit long for a 4090 and the original Flux dev. When you configure your Nvidia Settings and set the "Cuda - System fallback Policy" to "Prefer no Sysmem Fallback" you will generate an Imager with the original Flux dev on a 4090 in around 20seconds. Be sure you don't have anything else running while you generate images!
Hi, please in NVIDIA Control Panel where is that option?
50 seconds for 4090 is way too slow. Even my 4080 generates 1024 square image in 21 seconds. So 4090 should be in like 15 or so. Schnell is very fast, only 3-4 seconds (on a system with 64GB RAM).
@@eltalismandelafe7531 It's in 3D Settings/Manage 3D Settings.
@@eltalismandelafe7531 Manage 3D settings
@@eltalismandelafe7531 Manage 3D Settings
Your cute accent, and cool hat nearly make me forget all the things you forget to give us to actually make this shit work. ahhh well your a good entertainer.
Will this work in stability matrix, comfy UI?
Yes it does! But see my coment of an error that appears with the solution in this same video.
Is there a Flux Online Verison out there? I cant install Flux an my work PC.
Tried everything possible, followed the whole steps but can't get it to work, got a pause at the end or a bunch or red errors.
NF4 doesn't work on rocm5.71 ATM :(
Error occurred when executing CheckpointLoaderNF4:
Error while deserializing header: HeaderTooLarge
Help?
Edit: nvm, I had to redownload the safetensor
Nice!How about ForgeUI tutorial?
You've mentioned that it can run on it.
Yes a non spaghetti ui tutorial would be much appreciated.
@@stevehaney344you can try swarm UI,uses comfy as backend but you don't have to mess with noodles
Online on the channel now.
Mmmh, it's not much faster than Flux1-dev-ftp8, on my 3090
The advantage is that we can use the regular "load Checkpoing" with it, and no need of a new node
need help if i can get the workflow so i can upload it directly
Is 16gb normal RAM not enough for this?
How did you get the process stats bars in the control panel?
no, no, not...skipping it for now and waiting for better times
what about adding lora? and there is no NF4 in my options
will that work also based on the Schnell version?
how to get the checkpoint loader nf4? i can't find it...
Hello and thank you for all the tutorials! I must admit I watched your videos extensively last year, I've taken a break from image generation since last summer with A1111 and SD 1.5. Question: Should it take 30 minutes to generate an image with this ComfyUI and NF4 setup? Or am I doing something wrong? The only thing I can think is maybe I need more RAM? I'm using Nvidia 4080 but I only have 32g RAM? I have ComfyUI and models installed on a Samsung 860 EVO SSD. Thank you and keep up the good work!
NF4 should be quick for you, like 30-60 seconds tops. 1024x1024?
@@sebastiankamph Yes! Even smaller. Should it say Loading Model, every time I try to render an image? It seems like that takes 29 minutes and the actual generation takes one minute.
@@sebastiankamph I put --highvram in the bat file and now it is running as you describe! Thanks again
Please compare with Main dev version
i keep getting "This action is not allowed with this security level configuration." when trying to download NF4 with URL since I cant seem to find it in the manager, tried googling how to fix it but no luck any ideas?
also when i opened config.ini to change security settings, the file didn't have security settings, checked 3 times to make sure I opened the correct file
nevermind, if anyone else had this issue... change to channel:dev lol
How are you getting the in-place previews on the “SamplerCustomAdvanced” node? I don’t see it after passing that node. (I’m new to ComfyUI and AI image generation in general)
I will add a +1 for this info - unsure how to do this also. Anyone?
In the ComfyUI Manager, you have on the left a "Preview method" and some options to choose from: 1. "Auto", 2. "TAESD (slow)", 3. "Latent2RGB (fast)", 4. "None (very fast)"
Did he just call his 4090 a potato pc 😅
I wonder if there is a way to add a node for swapping a face like in ControlNet, where I can add a photo, then generate another photo with Flux (or Flux NF4) modelbut with the same face as the added photo.. Is there any possibility to do that?
Hello UP, my graphics card is 1080 8G, can it be used normally?
Can you provide a video tutorial on installation?
Can you share your workflow? ? ?
Thats all on his patreon, i use a 1080ti without issues, just not able to use loras with nf4 on 1080ti rn i get cuda error.
No difference of speed rendering at all for me. do I miss something? I used the same nodes
1:22 go here where? What is this shit guide?
me getting error:-
load_checkpoint_guess_config() got an unexpected keyword argument 'model_options'
update comfy
It works, but I have questions. If it seems possible that the flux 1 dev model can be slimmed down so that it now runs on a smaller GPU, why didn't the developers from the Black Forest include such a model? To what extent do the licenses change for the now increasingly modded flux variants?
In fact, Mimicpc is now a product that can run Flux online, and I think it operates out of the details and generates them at a commendable speed.
Can you run a comfyUI workflow or it is just an online text to image generator ?
@@francoisneko I was downloading the FLUX model to install into Mimicpc's Comfyui, and as far as I know its Comfyui has an upgraded version, in order that we can run FLUX better!
link for model doesn't work
Thanks! I installed all the custom nodes cmd and ComfyUI_bitsandbytes_NF4 cmd etc. Update the Comfy but can not find the Check Point Loader nf4
Edit: I fix thanks to @stevecastle1730
To get the CheckpointLoaderNF4 node I had to turn the ComfyUI-Manager security setting to "weak" in the config.ini file. Then use the "dev" channel in ComfyUI manager to Install Missing Custom Nodes.
Where is this security settings? In the manager, folders, script, etc?
@@helveticafreezes5010 ...\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-Manager\config.ini
Last line should read security_level = weak
@@helveticafreezes5010 first go to custom nodes folder and ComfyUI-Manager, find configini, change security_level = weak in the last line, than open comfy ui, manager, find the channel at the top left. (Two below the skip update check). Make it dev
@@helveticafreezes5010 first go to custom nodes. Than manager folder. Find configini. Open it with notepad. Change Security level weak on the bottom line. Save close. Open Comfy. Open manager. Top left you will find the "dev" its hidden. Revealed with Click.
It's a good solution for me.
I haven't been able to get flux working at all in my ComfyUI install, I keep getting all sorts of pyTorch errors of non matching matrices. I guess I need to restart from scratch. I trust the model(s) are great .. but there's a lot of teething issues.
Try Forge. NF4 works straight out of the box there pretty much.
So, for all Apple Silicon M(X) users, we are stuck on bitsandbytes v0.42. Version 0.43 needs to be installed for the NF4 stuff, which doesn't work right now.
Go figure : ( I've been trying to find a solution for the past 3 hours. I'm on a Macbook pro M1 machine, thanks for the heads up
@@robbiepacheco753 we need to be patient. They're opening a new issue on BNB github for it.
Darn, thanks for saving me the 3 hours lol
Yes, that's very unfortunate. I'm getting stuck here as well. :/
well it sucks finding this after wasting my afternoon ;) Any Solutions for Mac users already by any chance?
Doesn't work with loras
well apparently this wont work in Comfy UI Windows portable (NF4 Loader) :(
I'm using ComfyUI Windows portable and it works. Here are some tips in the comments that might help. The only thing I needed to make it work (apart from the explanations in the video) was going to the ComfyUI Manager and on the left change Channel from "default" to "dev". After doing that, I opened the Custom Nodes Manager, searched for "nf4" and it showed up, seemingly broken. But there was a button "Try Fix" and after clicking it and then restarting, it worked for me.
my potato pc says thank you !!
FIX for (This action is not allowed with this security level configuration.) | CheckpointLoaderNF4 not loading properly!
follow these steps:
open directory of ComfyUI-> Custom_nodes -> ComfyUI-Manager
open config.ini within this directory in any text editor of your preference (notepad is fine, vim or neovim for arch users btw)
In the bottom of this config.ini file there's a line, change / edit this line (or similar)
' security_level = normal' (not good, makes security level too high for proper installation)
change it to
'security_level = weak'
(without the ' ' ofcourse..)
Restart ComfyUI and open up ConfyUI Manager.
You should be having no issues with the ' not allowed with security level config' problems preventing you from loading the model properly.
If u still run into issues. Feel free to reply to this msg and i'll try and help you out.
have fun
I'm getting TypeError: load_checkpoint_guess_config() got an unexpected keyword argument 'model_options'
did it, no good still
doesn't find it for me as well. running it in mac terminal tho
1:22 you can go in where? Downloaded comfyUIManager but how do I get that window? Searching for checkpointloadernf4 in the manager under Custom Nodes gives no results. Thank you. Edit: Doubleclick the workboard, ok, but searching for checkpointloadernf4 still gives no results. Updated everything, restarted and refreshed browser
same here. Updated everything and NF4 doesn't show up in the search.
Install like I did. It'll be in the manager soon
To get the CheckpointLoaderNF4 node I had to turn the ComfyUI-Manager security setting to "weak" in the config.ini file. Then use the "dev" channel in ComfyUI manager to Install Missing Custom Nodes.
@@sebastiankamph Hi, GREAT channel! I can always count on you to keep us up do date with the newest tech!!! I followed your instruction and have been waiting for quite some time. How long does the download typically take to get the node into manager? Thank for all your hard work!!!
@@stevecastle1730 Thank you, this is what I needed to get it to install.
Well ... dat alpaca doesnt look like "pixar style" ... but ... awesome anyway :D
I get the following error message: Error named Symbol not found at line 90 in file D:\a\bitsandbytes\bitsandbytes\csrc\ops.cu
This directory doesn't even exist because I only work on C:.
I use a Radeon RX 6700 XT graphics card and ComfyUI with Zluda. It actually runs smoothly with the other Flux models. It's just pretty slow. When I want to integrate the FLUX NF4 model, i always get the error message from above in this workflow. With another workflow I then get this error message: "Error occurred when executing UNETLoader:
"I don't know what to do anymore...
Flux nf4 6gb model review please
Where can I get the 6GB model?
@@rennynolaya2457 silveroxides/flux1-nf4-weights
@@rennynolaya2457 Huggingface
@@WorldCuppppp582 do you mean flux1-schnell-nf4 ? or a new dev nf4 model ?
Can I run it with 4070 ti?
Funny thing in ComfyUI, I have to try 2-3 times for each new prompt. I get OOM for 1-2 generations right from the start and after changing the prompt. That's on my rtx 2080 super 8GB. On my 4070 12GB no such issue.
Everytime i start the generating ComfyUI crashes, ridiculous. all up to date, all set right, but it crashes everytime
Just in general, since everybody is praising Flux as being so super good... for simple prompts like "portrait photography closeup of beautiful woman with black hair and bright eyes, age 35" or "portrait photography closeup of old man with beard, age 65" which give me a large variety of awesome looking portraits with SDXL, the result with Flux is very limited and all very similar. Also always very smooth and artificial looking.
And what bothers me the most, the portraits are always very oversaturated. Adding different ways of trying to tell Flux to make it less saturated are completely ignored.
Interestingly enough when I set the Denoise factor in the BasicScheduler to 0.9 instead of 1, I get massive variations in the result which all look good, but then the "photography" part is completely ignored. I get all kinds of images, photorealistic, sketches, manga, illustrations, paintings...
So since this is supposed such an awesome model, is there any way to get some control over it? Although I've seen videos now where everyone says how good this model is at recognizing what you are telling it, it is kind of annoying how ignorant it seems to me.
Flux is the best!!!!!!!!!!!!!! f sd3 hahahaah FLUX FOR THE WIN!!
hey everyone, i followed every step of this tutorial but for some reason the custom node "checkpointloaderNF4" isn't showing for me, even when i search for it or try to install it from install missing custom nodes, it's not showing, any assistance please ?
EDIT (from anther comment on here) : Here's what I ended up doing. Go into the folder for the Manager, edit the config.ini and change th security level from 'normal' to 'weak'. Run Comfy, go into the manager, change the channel to 'dev' then search for the node 'NF4'. then i went into install missing custom nodes, it showed me the node and a "try fix" button, that worked for me. hope it does for you!
imagine running 4GB VRAM with 16 GB onboard memory and a 52 GB virtual memory size, takes about 8 to 10 mins per image, depending on how big past 768 x 768 but it works that's the thing.
has anyone noticed the quality drop of this NF4 version compared to the FP8, it's very noticeable for me, I'm using the schnell versions. i did get a 1 second decrease in generation time but it's not worth it. I'm using Tesla T4.
Hello Sebastian, I followed the instruction updated the comfyui but still cant find the nf4 loader. I wonder which comfyui u are using? can u guide if I can install this loader separately by any means.
flux lora not working?
Error occurred when executing SamplerCustomAdvanced:
.to() does not accept copy argument
Not for NF4 sadly.
@@sebastiankamph thank you very much for the quick reply, maybe there will soon be a NF4 or for bitsandbytes lora support
Is there a recommended comfy tutorial for someone getting started into the local graphics generation. It looks like local generation has come a long way. Let me know.
Hi please tell me. You use Pinocchio to intall flux ai to your mac. Please help. I have M1 mac mini 8GB version. Is ot work for me. Help
Hey I'm really new to this image creation stuff. I don't really want to spend thousands of dollars on a GPU. I currently have the API to stability AI configured in my application and I'm making prompts, but how can I use custom models and modify all of the image parameters that your UI gives you access to? Is there a way to put this in the API call?
Anyone tried it with a 3070Ti 8Gb or similar and care to share the slow experiece? Any tips? Much appreciated.. =)
I'm sure this is my fault, but when I tried the nf4 loader, my computer literally died. I'm not sure what happened.
I just updated the comfyui but can't find the CheckpointLoaderNF4 node T-T
i installed the bitsandbytes nf4 and the requirements.txt inside it before updating the comfyui, but the CheckpointLoaderNF4 just won't appear from the search bar and if I use your workflow, it'll mark red the supposed node because my comfyui can't find it
Check your startup log and see if you get an IMPORT FAILED for the nf4 node.
Aren't the clip models including the T5 encoder already built in the model?
so... i did the update and refreshed and restarted but there is no node "checkpointloadernf4".