The constant quick side commenting was magical in this vid - you kept it brief enough so that veterans won't feel bogged down by old redundant info but also that newbies will highly benefit from it. World class tutorial protocol.
and for those who are stuck with 4GB VRAM, just use a large virtual memory and about 768 x 768, it takes up to 8 mins depending but hey, we did it before on 1.5. and SDXL, we can keep going, you can do this.
A node that really helps to manage latent/image sizing btw is an underappreciated little extension from Ser-Hilary, called SDXL_sizing. Automatically spits out the right size for any base resolution (e.g. 512, 1024, 2048), at any given aspect ratio. I use wildcards from Impact Pack as an input to the aspect ratio setting, which works *really* well with Flux.
Excellent job on this workflow! Playing w/ it now after making a few of my own/using some common on civ/discord, but your incorporation of the LLM Party node + autosizing/etc is simply brilliant. Hope all is well!
Even if outputs were non-commercial, education and reporting are protected fair use. (Education and reporting are inherently commercial. Teachers and journos have to get paid.) Fair use is the doctrine all AI model training is built on. 😊
Here’s my theory, buy the biggest (consumer) gpu available, 4090, and be good for a few years. People will spend the minimum and expect to be good forever. In the AI world models are only going to keep getting bigger and more reliant on vram. It’s unfortunate but if we want better AI then it comes with a cost
Hello. Great tutorial! :) Though... I am trying to follow your tutorial and get llama 3.1 up and running in comfyui, but it is not working. I am not sure what to put into the base_url field of the API Large Language Model Loader. Any help? Also, why are you using this node and not the "Local Large Language Model"? Thanks
Does your patreon have workflows that work with the largest flux dev model? I have the regular one working fine, but it looks quite different from what is here, (and uses the 24GB file), and I would like to get all the extra stuff on your patreon if it works with that as a basis (or basically works with 24GB related files). BTW: I use a RXT 4090. Thanks!
How do you add/remove Pin(s) to slots of a node? For example, the pin to the t5xxl slot (wich become hidden) in the CLIPTextEncodeFlux node, or the user_prompt pin in the API Large Language Model? (sorry, complete noob in Comfy). Thanks.
I don't have comfy installed yet... but I've been wanting to try flux img2img with my DAZ3D renders... I'd be interesting to see how subtle I can make the changes... I just want to make the renders look like either painted/illustrated or more realistic.. while keeping as much detail as possible.. I got some good results a1111 by simply upscaling renders with certain models and prompts in effect... I'm wondering if I can do the same with Flux
@@NerdyRodent damn, im already on fp8 with the smallest Llama 3.1 model, it goes just enough over the vram to slow everything down with Florence + Llama.
you can change the amount of time all models are loaded into memory by setting the `OLLAMA_KEEP_ALIVE` environment variable when starting the Ollama server. You'll need to wait for it to load again each time, but if you've got other things eating that VRAM it may help
@@kkryptokayden4653 I've used both and tested both intensively. The Dev version is really good for photorealism, but the Schnell model seems great for illustrated outputs. And of course, it's much faster. Both models have their uses. I prefer the Dev, but I use the Schnell for img2img.
Even having 32 GB of ram it will max fill and swap memory and slow down! Close all other apps and tabs! i just got some extra ram yesterday to 48 gb. i just found the FP8 version and got a bit of speedup.
Can someone help me understand why the fp16 model has to be used from unet and the fp8 can just be used as a normal checkpoint? What even is the difference between unet and checkpoints and why does it matter for what is essentially the same model?
Hi buddy, what GPU are you using? I have an rx6800 which I am really happy with and it's 16gb for half the price of Nvidia as of last week and hopefully I will be able to get it to work with a few tricks here and there....just wondering...thanks! ;D
I’ve got an old 3090 as VRAM is where it’s at. AMD cards should work for the most part on Linux, though on MS Windows it’s likely to be a slightly more bumpy journey and may not work at all for many things!
@@NerdyRodent Thanks for your reply, I've had a few 'bumps' for Stable D but I did overcome then, I've discovered ZLUDA and Anaconda and peeps have things up and running but it's a lottery to how much ball ache you may have to experience, one guy is getting really fast renders with ZLUDA on AMD so eh, fingers crossed, thanks for your helpful videos and I'll let you know, just to add that the rx6800 with gaming is only 20% lower on average with FPS than the rtx3090, I know there's more to AI but that's really good and a used 3090 is still $6-700 depending on which one...so hopefully I'll get some reasonable performance as well....have a fun day!
Not mentioned, but I presume that to use the standard checkpoint loader I must move the flux models from unet to where they will be visible to the std loader. I ran the update-all before getting started tonight and flux is basically broken on my 3080 now after using it all week... so I don't have a choice other than trying to modify the workflow.
@@squallseeker-i2i The checkpoint version of the model is not the same file as the unet version. Move the unet version back and download the checkpoint version. The checkpoint version is around 18 GB.
@squallseeker. No - use the linked files above each workflow as they’re appropriate to that workflow. For example, to use the one shown in this video you’ll need to download the file as shown, directly to the location shown.
who are the people that can run this workflow? I am on 12gb 3060 gpu, is it possible to run flux and ollama at the same time? does anyone know about this?
From what I've read, Flux doesn't use negative prompts. If you add one it will often ignore it. Try what @tc8557 says and describe what you don't want.
There are multiple options for florence, but I find flash attention works just fine! Align Your Steps is just an option that works quite well - ruclips.net/video/JgYzCEzHDrc/видео.html
@@shApYT Funny. I tried to get the demo sample project to do anything at all with no luck. I use othe AI programs all the time. ComfyUI just looked like a logic flowchart mess of object oriented containers gone off the rails.
this is not the best approach - why load into RAM/VRAM LLMS and corresponding nodes, if you can create prompts with them and after run comfy to generate?
I'm usually very sceptical about "these" things, but why do most of the major RUclips channels with AI tutorials always use a political figure to explain image and video generation with AI? I mean, come on 🤐
I'd say it's because it's someone recognisable, a public figure of their own choosing, that doesn't fall into the category of actors etc who don't. Or it's something else, like them being easy targets.
@@_.o..o._ I don't think it's being done because it's political, not on this channel anyway but for me using actors should be okay as long as nothing outrageous is done with them. 🙂
I tried loading flux with the same loading checkpoint as you. It fails with: Error occurred when executing CheckpointLoaderSimple: ERROR: Could not detect model type of: /mnt/e/Projekty/_AI/ComfyUI/models/checkpoints/Flux/flux1-dev-fp8.safetensors (this model does load the "old" way, though)
2:41 WHAAAAAT That's how that works! Oh my goodness! I heard someone say the picture is the workflow, but didn't get it. Now I do :) Thank you!
It's one of the magical bits about ComfyUI :)
The constant quick side commenting was magical in this vid - you kept it brief enough so that veterans won't feel bogged down by old redundant info but also that newbies will highly benefit from it. World class tutorial protocol.
and for those who are stuck with 4GB VRAM, just use a large virtual memory and about 768 x 768, it takes up to 8 mins depending but hey, we did it before on 1.5. and SDXL, we can keep going, you can do this.
@@synthoelectro what do you mean by use a large virtual memory? Can you explain?
@@riflebird4842 swap file.
Canny for flux has just been released today
I thought someone might notice that sneaky screenshot I put in 😉
OMG its really going good!!! 🎉🎉🎉
Whats canny?
@@PhotoBomber A controlnet model type. Lines and such, google it.
What is canny
A node that really helps to manage latent/image sizing btw is an underappreciated little extension from Ser-Hilary, called SDXL_sizing. Automatically spits out the right size for any base resolution (e.g. 512, 1024, 2048), at any given aspect ratio. I use wildcards from Impact Pack as an input to the aspect ratio setting, which works *really* well with Flux.
Excellent job on this workflow! Playing w/ it now after making a few of my own/using some common on civ/discord, but your incorporation of the LLM Party node + autosizing/etc is simply brilliant. Hope all is well!
Nerdy, you always deliver!!! 👌🏻👏🏻👏🏻👏🏻👏🏻
Thank you. LLM party sounds awsome 😄
love this narration haha thank you!
Even if outputs were non-commercial, education and reporting are protected fair use. (Education and reporting are inherently commercial. Teachers and journos have to get paid.) Fair use is the doctrine all AI model training is built on. 😊
"nice beaver"
ok, thanks for that.
I've just had it stuffed.
My god ! You re good :)
"Pandora's Box of Oddities!"
😅🤣👍
:) cute things AI makes are cute.
Bill didn't seem to mind the multiples of 16! Stay nerdy!
OOOOOooh NErdy rODEnt...
Almost 50K subs. I'll do my part. (just subbed)
Amazing, tutorials to make a custom lora soon?
I hope FOOOCUS adds support for Flux
Do you have a copy of the final workflow that we can download?
Of course! www.patreon.com/posts/ai-enhanced-flux-109665789
Hi, thanks for the awesome tutorial! How did you get the labels updated for the switches input (ex: "1_llm_enhanced", instead of "text1")?
Right-click on the small dot next to the label and you should see "Rename Slot".
I really like Flux, but I'm disappointed that I keep having to buy a new GPU to keep up with the demands of new AI art models. 😁
Here’s my theory, buy the biggest (consumer) gpu available, 4090, and be good for a few years. People will spend the minimum and expect to be good forever. In the AI world models are only going to keep getting bigger and more reliant on vram. It’s unfortunate but if we want better AI then it comes with a cost
Hello. Great tutorial! :) Though... I am trying to follow your tutorial and get llama 3.1 up and running in comfyui, but it is not working. I am not sure what to put into the base_url field of the API Large Language Model Loader. Any help? Also, why are you using this node and not the "Local Large Language Model"?
Thanks
A face detailer workflow for flux would be really useful.
Can we get the img2img workflow?
it says to install flash attention. where to get it and where to put?
The ability to load the flux models via "Load Checkpoint" thows up an error for me ...Could not detect model type of ... did you have this problem?
Are you using the special fp8 checkpoint downloaded via the link above workflow, like in the video?
Lets Party 🎉🥳
hello my LLM party problem: Error code: 500 - {'error': {'message': 'llama runner process no longer running: -1 ', 'type': 'api_error', 'param': None, 'code': None}}
goodbye mj🤭👍
Does your patreon have workflows that work with the largest flux dev model? I have the regular one working fine, but it looks quite different from what is here, (and uses the 24GB file), and I would like to get all the extra stuff on your patreon if it works with that as a basis (or basically works with 24GB related files). BTW: I use a RXT 4090. Thanks!
Yup, there’s a whole boatload of workflows 😉
@@NerdyRodent Thanks, signing up!
How do you add/remove Pin(s) to slots of a node? For example, the pin to the t5xxl slot (wich become hidden) in the CLIPTextEncodeFlux node, or the user_prompt pin in the API Large Language Model? (sorry, complete noob in Comfy). Thanks.
You can write click on any node for a variety of options,
@@NerdyRodent Yup, of course, but i don't find any options to add/remove pins to slots :-/
I don't have comfy installed yet... but I've been wanting to try flux img2img with my DAZ3D renders... I'd be interesting to see how subtle I can make the changes... I just want to make the renders look like either painted/illustrated or more realistic.. while keeping as much detail as possible.. I got some good results a1111 by simply upscaling renders with certain models and prompts in effect... I'm wondering if I can do the same with Flux
install it already
is there a way to reduce the memory usage of Llama ? it's peaking heavily on my 3090 and makes everything painfully slow.
Using fp8 and a smaller LLM should reduce the load. With 64gb RAM and a 3090 you should see images generate in around 20 seconds
@@NerdyRodent damn, im already on fp8 with the smallest Llama 3.1 model, it goes just enough over the vram to slow everything down with Florence + Llama.
you can change the amount of time all models are loaded into memory by setting the `OLLAMA_KEEP_ALIVE` environment variable when starting the Ollama server. You'll need to wait for it to load again each time, but if you've got other things eating that VRAM it may help
@@NerdyRodent Done, no more long queue, thank you :) the results with florence and Llama are simply amazing
@@DrMacabre yeah, it totally changes how you can prompt!
Do Loras work yet (and I am just too stupid for it) or are those also still on the waiting list?
Yup! Loras and controlnets are available, though still early days!
yes but of course they need additional VRam controlnet too
Can you please compare the different flux gguf versions please
Is it possible to run it locally with an RTX4060 8GB VRAM and with 16GB RAM ?
@mexihcahcoatl4105 that's with Schnell, right?
@mexihcahcoatl4105 I'll try and I'll share my result and opinion. Thanks!
@@BlackParade01 I would use dev version is way better, slower but is worth it
@@kkryptokayden4653 I've used both and tested both intensively. The Dev version is really good for photorealism, but the Schnell model seems great for illustrated outputs. And of course, it's much faster.
Both models have their uses. I prefer the Dev, but I use the Schnell for img2img.
Even having 32 GB of ram it will max fill and swap memory and slow down! Close all other apps and tabs! i just got some extra ram yesterday to 48 gb. i just found the FP8 version and got a bit of speedup.
Im starting with flux can you share that workflow??
Sure! The pre-made version you can grab from www.patreon.com/posts/ai-enhanced-flux-109665789
Llama 3.2 for Comfyui FLUX ?
Prompt enhancement node?
Yup, works great with llama3.2… or indeed much larger models
Can someone help me understand why the fp16 model has to be used from unet and the fp8 can just be used as a normal checkpoint? What even is the difference between unet and checkpoints and why does it matter for what is essentially the same model?
Think of the collection of files like a zip file, where lots of things are collected together - whereas the unet is just one of those files.
Excuse me, but I didn't understand, where can I find the workflow you have shown in the video?
On his patreon page
@@SouthbayCreations Thank you! Wasn't clear :)
Hi buddy, what GPU are you using? I have an rx6800 which I am really happy with and it's 16gb for half the price of Nvidia as of last week and hopefully I will be able to get it to work with a few tricks here and there....just wondering...thanks! ;D
I’ve got an old 3090 as VRAM is where it’s at. AMD cards should work for the most part on Linux, though on MS Windows it’s likely to be a slightly more bumpy journey and may not work at all for many things!
@@NerdyRodent Thanks for your reply, I've had a few 'bumps' for Stable D but I did overcome then, I've discovered ZLUDA and Anaconda and peeps have things up and running but it's a lottery to how much ball ache you may have to experience, one guy is getting really fast renders with ZLUDA on AMD so eh, fingers crossed, thanks for your helpful videos and I'll let you know, just to add that the rx6800 with gaming is only 20% lower on average with FPS than the rtx3090, I know there's more to AI but that's really good and a used 3090 is still $6-700 depending on which one...so hopefully I'll get some reasonable performance as well....have a fun day!
@@CapaUno1322 nice!
Works with flux1 dev.sft?Because I don't have the extension.safetensor in the model I use
Yup! Remember to use the workflows at the top when using the individual files though
@@NerdyRodent thanks bro
Not mentioned, but I presume that to use the standard checkpoint loader I must move the flux models from unet to where they will be visible to the std loader. I ran the update-all before getting started tonight and flux is basically broken on my 3080 now after using it all week... so I don't have a choice other than trying to modify the workflow.
@@squallseeker-i2i The checkpoint version of the model is not the same file as the unet version. Move the unet version back and download the checkpoint version. The checkpoint version is around 18 GB.
@squallseeker. No - use the linked files above each workflow as they’re appropriate to that workflow. For example, to use the one shown in this video you’ll need to download the file as shown, directly to the location shown.
who are the people that can run this workflow? I am on 12gb 3060 gpu, is it possible to run flux and ollama at the same time? does anyone know about this?
Hi, do you accept sponsorship to your video?
can you somehow simulate negative prompts??
@@sinayagubi8805 since it's an llm you're prompting, just say what you dont want.
Try like 'with no eyeglasses or makeup on'
From what I've read, Flux doesn't use negative prompts. If you add one it will often ignore it.
Try what @tc8557 says and describe what you don't want.
thanks for the vid, really appreciate it!, But the resources needed to run flux, It's still too just demanding for me
Oh, Nerdy Rodent, 🐭
he really makes my day; 😎
showing us AI, 🤖
in a really British way. 🫖
🎵🎶
😄
Long version when ??!!
hmm whats this ays + scheduler. Also, is Flash attention 2 required for maximum nerdyness or not really?
There are multiple options for florence, but I find flash attention works just fine! Align Your Steps is just an option that works quite well - ruclips.net/video/JgYzCEzHDrc/видео.html
@@NerdyRodent wow awesome video! I don't see AYS+ in my ksampler though. I do see Beta and Resample though ..
i have the same problem... no ays+ scheduler?
@@TR-707 Align Your Steps is just an option that works quite well - ruclips.net/video/JgYzCEzHDrc/видео.html
@@Rentoa Align Your Steps is just an option that works quite well - ruclips.net/video/JgYzCEzHDrc/видео.html
Has anyone shared this workflow already or do I have to build it on my own?
You can get the workflows here - www.patreon.com/posts/ai-enhanced-flux-109665789
CLIPTextEncodeFlux as of this day doesn't have the t5xxl input
I found ComfyUI impossible to figure out.
Then you'll have trouble using any other art software. Nodes aren't hard. Just do it. blender, houdini, substance designer, davinci resolve etc.
@@shApYT Funny. I tried to get the demo sample project to do anything at all with no luck.
I use othe AI programs all the time. ComfyUI just looked like a logic flowchart mess of object oriented containers gone off the rails.
@@choppergirl millions of artists using blender, Houdini, grasshopper, davinci resolve, substance and countless. It is the standard interface.
did you try the portable /standalone version? and what are your GPU specs?
@@shApYT That's complete rubbish about other art software.
this is not the best approach - why load into RAM/VRAM LLMS and corresponding nodes, if you can create prompts with them and after run comfy to generate?
definitely a better idea
Generating images with the checkpoint models takes forever on my 8gb card compared to the unet models for some reason.
this new checkpoint is huge! the regular Flux one is 22GB, the FP8 version is 11GB. We got so used to having SD1.5 at 2-4 gb! XL and pony 6GB
My version is:
2489 (comfyui)
2.48.5 (manager)
What 4GB vram people can do ??😢
Sd1.5 is a great choice for extra low vram cards!
Of all of the 8 billion photos out there you chose Bill fuckin Clinton 👌
How do I get the manager menu to come up. I'm ew to ComfyUI, sorry if it's a dumb question
ComfyUI manager is the first thing to install right after Comfy itself - github.com/ltdrdata/ComfyUI-Manager
the moment you don't check he mean schnell > Schn-e-ll
If it uses CLIP why dont you just use clip vision encoder to input an image and trick flux by telling it its a text embedding lol
I see rodent I click. Simple 👍
Not a nice model. Generations look like animation or drawing and the skin is like plastic or shiny! Going on with SDXL!
Don't use the Schnell version. It's plastic. Dev version is much better.
@@hadbildiren123 I agree the dev version and the schnell version are very different it's day and night
I hate comfyui with a visceral passion
Oh, so you don't know how to use it and you've never even really tried... thanks for sharing!
I'm usually very sceptical about "these" things, but why do most of the major RUclips channels with AI tutorials always use a political figure to explain image and video generation with AI? I mean, come on 🤐
I'd say it's because it's someone recognisable, a public figure of their own choosing, that doesn't fall into the category of actors etc who don't. Or it's something else, like them being easy targets.
@@Elwaves2925 I would rather see actors' faces. Lately, whatever I try to learn, I see political figures shoved down my throat everywhere 🤮
@@_.o..o._ I don't think it's being done because it's political, not on this channel anyway but for me using actors should be okay as long as nothing outrageous is done with them. 🙂
I tried loading flux with the same loading checkpoint as you. It fails with: Error occurred when executing CheckpointLoaderSimple:
ERROR: Could not detect model type of: /mnt/e/Projekty/_AI/ComfyUI/models/checkpoints/Flux/flux1-dev-fp8.safetensors
(this model does load the "old" way, though)
Could be the checkpoint file or your version of comfy? Make sure it’s the 17gb one and check the sha256