It's only good at 1mp resolution. Very good at styles, but anatomy is terrible. It's a step up from SDXL, but not as good as flux. Unlike flux, it can be properly trained tho.

not yet, we're all waiting for lllyasviel. Rumors say he'll add support after the 3.5 medium release.

Is that better than large turbo or fp8?

@P_Universe Its fp16 in "zip" basically, Q8 is always better than fp8, if we dont mind somewhat slower diffusion due compression.

@Mescalamba thank you so much for the explanation

I always make my own GGUF out of the base models - quantification is pretty simple ... ComfyUI-GGUF/tools at main · city96/ComfyUI-GGUF

on my 3090 its a little faster than flux but not much.

I think the comprehension is a bit better then flux. What I've read.

Better prompt understanding than Flux.

Slightly worse quality than Flux but not much.

Better quality than SDXL.

You can hardly immediately blame 3.5 for the behavior you experienced. Did you try again? Are you sure everything necessary for 3.5 is installed and up to date? In general, Flux is more resource intensive than 3.5, so it doesn't make much sense to put the blame on the checkpoint. I've been finding it pretty fun to work with so far, so I hope you get it figured out!

@FauxRealDoe It happened again. Maybe something wrong with my Comfy, even though it works fine with Flux and SDXL.

@StargateMax what flux model are you using, the defult one or some smaller sized?

It sounds like you filled up the system ram and entered a swap storm. A couple of things, watch the memory usage when it's running to confirm, and check where your swap file is - move it to the NVMe if it's on a mechanical drive. Specific fix? Maybe try smaller quantized models for text encoders etc.

Its AI, just do it anyways? Flux is clearly originally trained on screenshots from the internet so who cares?

Sd35 is free for both commercial and non commercial is you make less that 1 mil. But flux dev 1 is free only for personal use not commercial without a license from them. Flux schnell is free to use commercially. But like the person above said go ahead just don't get popular... That probably have a watermark system. Lol

@pychobj2001741 You right ！ haha

@pychobj2001741 I would assume SDXL and SD1.5 is the same? Free for both Commercial and Non Commercial

@AikoMatsui I couldn't rember for SDXL so I looked at it license at hugging face, the base sdxl 1.0 and summed up that it's license w/ perplexity: The CreativeML Open RAIL++-M License allows both personal and commercial use of the model and its derivatives, provided that users comply with specific use-based restrictions. These restrictions prohibit certain uses, such as generating harmful content or exploiting minors. Users must also share the license terms when redistributing the model or its derivatives 

It looks like SD.15 is on the same license.

In short, yes, they are both free to use.

@pychobj2001741 The minor part I get it. It should be part of ethical use. Good to know that it's "free" to use.

use a dual clip loader and download separate clip models for flux models without clip

yes you can

not sure about the 3.5 though, a1111 was not updated recently

I got error on my a1111:
: ('Cannot copy out of meta tensor; no data!',).

While copying the parameter named "first_stage_model.decoder.conv_out.bias", whose dimensions in the model are torch.Size([3]) and whose dimensions in the checkpoint are torch.Size([3]), an exception occurred : ('Cannot copy out of meta tensor; no data!',).

You can't atm

But in the next decade? maybe

@Mr_Jinguji in 10 years AI will be sending our generated images in the form of WestWorld copies to our homes. Hopefully earlier.

use webui forge

automatic1111 cant handle it sorry

yes. it should be 30. but running the dpmpp_sde vs the 3m_sde is much slower. So expect 2x to 3x the time needed to run. The results are fairly close unless you push the sharpness up.

Would be great to get a GGUF version of this!

Perhaps a GGUF Q5 if you have a minute hehe

I'm getting that too, and I believe that has something to do with the 1MP limit for SD3.5L. I saw a video of some sort of control node for it, and need to find it again, because it helped out greatly. Perhaps someone within this page can point us in the right direction.

same here!

Same here too

On my 2070 (8Gb) graphics card in Forge, the generation process takes about 60-90sec at the usual recommended settings.

@bsdesign does this hurt the gpu

@LoliGarden Tough question... I haven't found such information on the net yet.

@LoliGarden The more fancy stuff your PC needs to run SD, the worse it gets. (100 add-ons, 100 diff. tweaks etc.)

@LoliGarden I hurt my 3080 by rendering AI videos with Deforum, however I was running the videocard at 100% utilization for over 24 hours straight at a time. What happened was my thermal paste dried up and my thermal pads degraded and the card started overheating. I just had to replace paste and pads and now im back in action. The factory pads and paste most card manufacturers use are not up to par anyways so I shouldn't have to do that again any time soon. Honestly running AI on your video card is the same kind of wear and tear a card would receive mining crypto 24/7. It heats up your memory modules more than anything. A little preventative maintenance can go a long way.

maybe RTX A6000

SD3.0 is the Medium version of SD3.
SD3.5 is the Large version of SD3.
What's missing is something that normal people can run, the Small version.

@punkbuzter340 That is false? https://huggingface.co/stabilityai/stable-diffusion-3.5-medium

me too!

@antonfawkes33350 SAI said they'd release SD3 in Small Medium Large sizes, apparently they've broken their promise then.

Read information:
"The Main Download (Full 11GB) is the version containing only the Unet, so you can use it accompanied by the clips of your choice, depending on whether you prefer the fp8, the fp16 or want to experiment with new clip_l or t5.

In the file section of the sidebar you can instead find the Secondary Download (Pruned 20GB) of the AIO version with t5xxl_fp16, clip_l and VAE already included."

@mykeehu what's the point of having pruned model if you are gonna make it a 20 gigabyte download???

@antonfawkes33350 Simplicity if you use Comfy as you don't need all the extra nodes for VAE and Tripple CLIP Loader. And ease of use for A1111 users as they don't have to download all the CLIP models and the VAE.
... Would be my guess.

I hope you get pruned for not reading.

It's DPM++ SDE

The Main Download (Full 11GB) is the version containing only the Unet, so you can use it accompanied by the clips of your choice, depending on whether you prefer the fp8, the fp16 or want to experiment with new clip_l or t5.

In the file section of the sidebar you can instead find the Secondary Download (Pruned 20GB) of the AIO version with t5xxl_fp16, clip_l and VAE already included.

You should specify the environment you are/will be using (i.e. comfy, forge, SDNext, or your own setup) as there are different ways to save on GPU when running models. You should be able to by offloading components to the CPU. You also save GPU memory on quantizing components of the model. Using GGUF versions of the model, etc.

its gonna run in 4060 as in i am using now but the thing is i have 32 gb ram if you have enough ram you can run 22 gb flux dev model with t5 fp 16 without any issue . it will take around 1.30 min to 2 min depend on the workflow . with 20 steps

Same

Same here.

nobody uses automatic1111 anymore, use forge

@Rudi_aus_Buddeln I get those strange patterns on comfyui. I'm afraid maybe my downloaded file is broken or incomplete.

FP8 is a great version too. But for max details youre going FP16 with more steps (DPMPP_2m for example with 40-50 steps). In FP8 youre fine with 30.

@kaivanbi922 And it takes forever, even on a high end (at its time) GPU like 3090Ti.

@engineer 3090Ti is not high end

In the models folder under checkpoints or diffusion model depending on which you download. All resources are available online if you're able to do a quick Google search. Focus on learning from YouTube and asking model specific questions here.

@PeachCandi Thats what I have done before, nothing works

@justinof1503328 Do you use it locally or use any cloud-based service? That's gonna depend on which...

Not sure why you find it this way. I don't find them bad at all. 3.5M is fairly zippy and 3.5L is 1/2 again faster than Flux Dev. Your environment and hardware up to the task?

@condzero1950 No, that's why I'm specifically using SD 3.5 Medium and not Large. This checkpoint you're looking at is labeled Medium but it's weighing like a Large model. I have other Medium checkpoints that run 9 times as fast. You get prompt coherence and output quality like a Large model so I'm suspecting maybe the labeling is wrong.

"so I'm suspecting maybe the labeling is wrong."

Yes, I found this out with one model fine tune. Labeled SD 3.5M but clearly a large. Let the size be your guide (all things being equal) as to what it might be.

Is this really a sd3.5 Large model?? If that's the case, I use Wavespeed either way.

So, it does appear to be large just by the amount of time it takes. Typically, I have been using Absynth and sometimes vanilla 3.5L and the times are similar between all three. So, it seems quite likely this is the SD3.5Large variant, which makes sense being the quality is a bit better than Medium. Using Wavespeed I can obtain about 40-45 seconds with a batch of 2 images most times running at 40 steps. Running at 50 steps only adds about 5-7 seconds total.

Additionally, when trying out medium vs large loras, i always have issues trying out medium loras.

Hello. Out of curiosity, did you find a solution? I just started using the model and am having the same issue with the top of the images.