#HuggingFace ?

65 messages · Page 1 of 1 (latest)

sacred latch
#

Hiya, I've been loving SwarmUI , its my first AI gen - I'm a little lost however with how to get models from hugging face or how to install wan 2.2 etc. Is there like a console you can open in swarmui to auto download from huggingface or anything along those lines? Thank you in advance !

high kindle
#

You just click download on huggingface website and save the model in the models folder, simple as that

sacred latch
#

Hi Berni, thanks for answering. if there's like 163 models for example. will it compile them together ?

#

cause it says 1 of 163 - so I'd have to manually download all 163 and then stick them in the folder ?

high kindle
#

Unfortunately not. Usually someone else repacks those annoying models into a single safetensor file. There is a official ComfyUI huggingface page that has a lot of these repacked models

sacred latch
#

Okay , thats what I thought. I did see some people using Qwen but looking at the list there has to be a compiled one somewhere lol

high kindle
sacred latch
#

Thank you Berni ! Appreciate you taking the t ime

high kindle
#

Anytime. Hope you have a beefy GPU tho as Qwen is a pretty chonky model

sacred latch
#

I think its ok for now rtx 5070 ti and 128gigs of ram , not sure what it requires but we'll see !

high kindle
#

16GB of VRAM on there is pretty good, won't fit the 8bit quant of it but something like a 5 bit quant should fit, not best for performance but it will run

sacred latch
#

thank you for the advice , ill do trial and error , still learning all this. its been fun though. I'd love to try the video makers. Any suggestions?

high kindle
#

Wan 2.2
Tho you will need a pretty well optimised workflow to fit it into VRAM ( Qwen and Wan are the hardest to run, your card will easily handle anythyng older with great performance tho like Flux and back)

sacred latch
#

I dont think i can run flux its why i went with Stable diffusion

#

the 5k rtx series dont work with flux?

high kindle
#

Flux 1 Dev in fp8 format is 11GB so it should fit no problems

sacred latch
#

excuse my lack of knowledge , whats fp8 ?

#

i thought it was do the pytorch and cuda versions that wouldnt work

high kindle
#

8bit floating point format

#

Back when models were smaller we didn't care as much how much memory they take up so they mostly used FP16 back then. This means each parameter takes 16 bits to store (also known as 2 bytes)

steel topazBOT
brittle sable
#

there's links for everything in the model docs

sacred latch
#

oh okay i understand so as long as the model is flux fp8 it wont turn into a blurry mess

brittle sable
#

(some people do use gguf q8 but it's a really odd choice to ever make)

high kindle
#

Flux is a bit anoying to work with in general, it doesn't like certain settings and makes a blurry mess. Also needs a lot of steps to get a sharp result (hence slow) so you might want to look into some of the Lightning LoRAs to let you use much less steps

sacred latch
#

well I like SD more so far

#

flux seems to always make the same character faces no matter what or the same animal look. not sure if you know what i mean

high kindle
#

Yeah SDXL is our workhorse right now. It makes amazing results every time

sacred latch
#

its like copy paste but different colors lol

high kindle
#

Yeah the base Flux model is not very good, you want to be using some of the comunity made finetunes of it that are better trained for the particular style you want to generate. Main benifit of Flux is that it is a smarter model that can handle more complex scenes. It won't be any prettier than a good SDXL finetune model

sacred latch
#

if you use only SDXL , can you generate difference faces without downloading community models? example trying to make a couple , the guy and the girl will always look the same unless i use a community model even if you describe the faces down to a pixel

high kindle
#

That's part of learning how to best use AI tools.
They say it takes no skill to write a prompt and make an image... true, but the result is just as uninspiring. It takes skill to generate nice results.

sacred latch
#

I'll keep reading guides on how to make better prompts then !

high kindle
#

Expect to spend many hours experimenting with various ways of using AI tools to learn the craft, watching what others do with AI to pick up ideas etc...

#

Heck even prompting strategies can change hugely as you change between different models. Just don't be discuraged by this not being as easy as it looked at first, everyone has to start somewhere. Also it is not always about generating the perfect result in 1 shot, sometimes you might run the image trough multiple AI tools before getting to the final result

sacred latch
#

It's already been a steep learning curve. I'll keep poking at it everyday !

#

I heard Chroma was really good for prompts

high kindle
#

But to be fair. for face swapping, QwenEdit works damn well (No native support for that in Swarm yet, so you need to dive into the ComfyUI noodles, but you can turn those noodles into a custom SwarmUI workflow)

sacred latch
#

I'm still rather timid about the workflows for comfy ui. I understand it's purpose but modifying and changing it I wouldn't know where to start

high kindle
#

Yeah they are for advanced users. Best to first get familiar with the basics (prompting, collecting good models, LoRAs..etc) then onto the intermediate (ControlNets, regional prompts, segmentation prompts, wildcards, refiners..etc) and then finnaly drive into the deep end (like wrangling latest novel models in ComfyUI)

sacred latch
#

for sure!!

#

lol

#

I'm going to be in the basics for a few weeks I believe

high kindle
#

Civitai.com is the most valuable resource you have in the world of AI image generation. Not only is that where you get all the best models, but you can also see what other more experienced people are doing with those models so you can follow their methods

sacred latch
#

only one ive been able to use , huggingface is rather a mess for me and even tried getting it from ollama

#

but it was like how many times can I shoot myself in the foot to get 1 model

high kindle
#

Huggingface is good for LLM models and folowing the latest state of the art stuff (where you need to know what you are doing)

sacred latch
#

with Qwen , is there a different model I should use than fp16 ?

high kindle
#

Go register an account on Civit, you will find a lot of use out of it. Not only does an account let you see all the NSFW content, but it lets you bookmark and track your favorite models and if you like you can also post your AI creations on there to show them off to the world

#

You will not be able to run Qwen at FP16, that would need very expensive server GPU

#

As i said best to stick to SDXL based models for a start. It just works, most of the people making AI art use it as it can produce some very nice looking results

sacred latch
#

sounds good! I thought I'd give it a try for fun lol and yeah Civitai only thing I understand and know

#

Berni , once again! thank you for your time and patience. Much appreciated ! Off to further adventures with more knowledge at hand this time 🙂

young edge
#

Thanks for asking the question @sacred latch . As another new user to swarm (regular civit user), when something says get the model from huggingface, I go there and then just stare at all the files and folders with no idea what to do. Reading this thread was helpful.

high kindle
#

For Civit users trying to run the latest tech the easy shortcut is so instead set the Civit search filter to only show "Workflows" this lets you find a workflow for the model you want to run. Offten the workflow description or a comment inside the Comfy workflow will tell you where to get models needed to actually run that workflow (potentialy also have a separate ComfyUI instance for testing them, especialy if you like to just blindly click "Install missing nodes" in the Comfy Manager)

brittle sable
brittle sable
steel topazBOT
high kindle
# brittle sable QwenEdit works great and super easily in Swarm

Is there already native QwenEdit support in Swarm? (not sure how to feed it multiple input images trough the generate tab, or is it just single image from Init Image)

I made myself a custom Swarm workflow for QwenEdit and i like it a lot. (Tho with these more advanced models like editing and video it usually makes more sense to use comfy noodles anyway,)

#

Excellent summary table for models btw! 👍
Really gives you a good picture of what is out there ( I need to give HiDream and Hunyan a try, they didn't get the hyped up online so i never looked at them)

brittle sable
#

multiple image support that was added in the 2509 model was even demonstrated in the announcement post #announcements message

#

HiDream is kinda irrelevant due to qwen and hunyuan outclassing it at a similar scale