I Tested SD 3.5 and ComfyUI.exe - Here's What's Best!

[@OlivioSarikas]

#### Links from my Video #### SD3.5 Workflow: drive.google.com/file/d/1flk6frqp43nO9sSRtlS4RXA6JhbDmB6Y/view?usp=sharing blog.comfy.org/sd3-5-comfyui/ stability.ai/news/introducing-stable-diffusion-3-5?ref=blog.comfy.org civitai.com/models/878387/stable-diffusion-35-large?modelVersionId=983309

[@shun_almya]

There is a new model called SANA that will bring about a new revolution. Please check it out.

[@havemoney]

I’ll add, open a new process in a new tab and pin it, so that the next time I start it I already have ready-made processes, inpaint, instant, etc.

[@brootalbap]

Olivio, please also prompt a woman laying on the grass with SD 3.5 sometime. Curious to see if they really fixed things with anatomy!

[@LouisGedo]

Hi 👋

[@maximood-tired]

i love that this niche hobby has so many great professional news-channels, its quiet of unique in itself

[@DurzoBlunts]

Love me some ComfyUI, since the beginning of it, it's been a great and fun system to play with and test. Found a lot of fun making complex workflows for specific generating.

[@Lucifael]

Hello Olivio Sarikas, First of all, I'm thrilled that my generators made it into your video! I still have to figure out my long generation times compared to other creators. Not quite sure why mine are significantly longer! I have a 4090. great video!

[@devnull_]

Pretty much all the new things are already usable in browser version too, major difference is that ComfyUI can now run as Electron-based standalone "app". You can already use the same new UI, configure keybindings, and I think you can force model download from settings too (Show missing models warning) but I haven't tried that one yet.

[@mojoduster]

Thanks! So far yours is the only workflow I've downloaded which actually works! :)

[@johnfraser21]

The license is basically free for any company under 1 million annual revenue now. Also unlike flux this isn't distilled so it can be trained to be better, though I think flux for the moment is obviously still better.

[@OlivioSarikas]

True. But if this can be trained as well as SD 1.5 and sdxl, it's really good

[@ripleyhrgiger4669]

Yup, flux is great but… one flux model is just like any other for the most part. If SD3.5 can come back swinging with wins then they will back on top in no time. Pony is obviously the best and most popular model out there right now for various reasons; and an sd3.5 model with that on it will solidly make it the users choice.

[@DaniDani-zb4wd]

@@ripleyhrgiger4669too bad the greedy developer of pony won’t do that. There is no incentive for him to pick sd over auraflow or the de-destilled flux libre

[@Vestu]

@@ripleyhrgiger4669 "for various reasons" 😏

[@NotThatOlivia]

4:19 - that white bunny! I love it

[@gigiwow99]

Only competition will lead to progress, and users are very happy to see this phenomenon. Thanks Olivio share. 👍

[@havemoney]

I really liked what Comfy did, I was waiting for it

[@claudioestevez1028]

FYI. You can change the lines from wiggly to straight in the options. I'm new to ComfyUI, so I don't know if this is something new. I still prefer the wiggly ones. It looks like the back of my TV.

[@baheth3elmy16]

Thanks! The large is a base model unlike Flux. Finetuning off SD3.5 will be great. Also, SDXL Loras don't work with SD3.5 but the Loras that will be made for it will make it superior.

[@TR-707]

Olivio thank you for a SWEET VIDEO, you rock!

[@Elwaves2925]

SD 3.5 is okay so far, for a base model. It works better with celeb names over SD3 and Flux, skin is a bit better than Flux but fingers are worse and prompt adherence is worse from what I've tested. It has potential but that depends on the creator community.

[@Lykon]

Hi Olivio, maybe I can answer for the negative conditioning thing. So basically image quality is at its best with no negative or negative zero out. However, with that trick you can still use a negative promot by having it only influence the first 10% timesteps of the generation. If your negative is empty you don't need it.

[@OlivioSarikas]

Awesome. Thank you

[@tetsuooshima832]

This is all very confusing so thanks for the explanation, it makes more sense now. However having those extra nodes will make a slight difference wether the negative is empty or not. Which brings more confusion indeed lol

[@saarhadad]

big tnx , i learn alot from u

[@andigold2694]

This new interface is also available in the browser version of Comfy for quite some time. You can activate it manually

[@gosammy1971]

I use the new GUI some time, the only new in the EXE is that you have tabs, but that can be done with the browser tabs similar

[@AkoZoom]

@@andigold2694 yeah sure some weeks ago. Then I stopped using this new menu bar coz there were no cancel or stop generating as it in Queue Ui still. Is that added?

[@Freeak6]

What I'm mostly interested is to know if SD 3.5 uses less VRAM than flux. With a 12GB graphic card, I feel very limited by flux (I cannot add many loras nor controlnet otherwise my VRAM goes over the limit, and it becomes incredibly slow), even using GGUF models.

[@ronbere]

bad configuration or workflow, I run flux-dev 16 with 12g of vram.

[@zebmac]

I have 12GB VRAM and use ComfyUI, but tried out InvokeAI and it handels Loras in a different way. In Invoke I can use 3 or 4 Loras without problems with FluxD. In ComfyUI I run out of ram (64GB) and VRAM when I try to use 2 Loras. Perhaps there is a way replicate Invokes settings in ComfyUI, but InvokeAI is growing on me.

[@ronbere]

@@zebmac i use a workflow with 3 or 4 loras, i have between 1.80 and 2.5 of it/s... with the biggest version of flux-dev

[@Freeak6]

@ronbere I'd like to see your workflow, because even with the most basic workflow (prompt and sampler) it takes at least 2min / image with 20 steps). What size is your latent image ?

[@dennishmiller]

Nice overview. Thanks very much. Is the complete UI with all of the windows and all of those parameters and other elements of the model available as a comfy preset? If so, could you give me a link to that? I have never seen such a thorough display of the elements of these models. Looks brilliant. Thanks.

[@cowlevelcrypto2346]

It's all very exciting, however, do you not feel somewhat overwhelmed by how fast it is changing? Trying to keep up.

[@FreqtradeFR]

Very interesting!

[@ysy69]

Thanks for keeping us updated. Would you recommend installing the new ComfyUI inside own (conda) environment ?

[@KlimovArtem1]

I was hoping for some in-depth side to side comparison with other models 😔

[@z1mt0n1x2]

Wake me up when there's something new that I can run locally without selling a kidney.

[@mailkeeper450]

Can use with A1111?

[@OVERLOARD949494]

I hope there will be a linux version

[@mufeedco]

The blog post, they said it is available on Linux.

[@AkoZoom]

​@@mufeedcobut? In fact this changes just the way it is installed. Here in one easy app ?

[@brootalbap]

As long as Stability AI doesn't even think about backstabbing the community with legal trapping pits again, i might give 3.5 a try.

[@-Belshazzar-]

all these new models show very little progress. Flux had a nice leap (for free models that is) but i don’t see anything mind blowing or inherently better or impressive in this model. meh

[@Marc42]

Any chance for A1111 coming soon as well?

[@mirek190]

a1111 not even get flux and you expect sd 3.5?

[@OlivioSarikas]

ForgeUi is basically A1111. But not every extension works there

[@alexmehler6765]

can we generate something useful for movie and vfx or is it just another desktop wallpaper generator ?

[@MikevomMars]

Unfortunately, many questions remain unanswered: - Is the .exe version faster than the browser version? - If migrating from browser version to .exe, do we have to download and install ALL nodes again? - Is the .exe version portable, so we can easily back up the ComfyUI folder with all models and nodes installed? - Are browser version workflows fully compatible with the .exe version?

[@大支爺]

It would be much faster If the exe was written by C++ language.

[@AlexanderZhukov]

Hi! Could you help me please? I can´t see that button Manager. On the previous version of Comfy "Manager" button also didn´t exist in my case.

[@OlivioSarikas]

please google "comfyui manager" and follow the instructions to install it. it's really easy :)

[@omegablast2002]

i think instead of these flux comparisons which make it look like a weak and undertrained model...youtubers should be focusing on SD3 vs sd3.5 THATS where you can point out clear improvements.

[@basic204]

Can you put together a video of how to run it on forge web UI?

[@policani]

Juggernaut seems to produce much better images for me. Am I missing something?

[@4.0.4]

Before SDXL tunes like Juggernaut, SD 1.5 tunes like CyberRealistic were better than (base) SDXL. The base models from Stability are never something you'd use as is.

[@Elwaves2925]

That's because you're comparing a tuned model against a base model. A fair comparison is the SDXL base to SD 3.5 base. It's like comparing a stock model car, straight off the factory floor to a version tuned and modded for track racing. :-0

[@petrimannisto9318]

@@Elwaves2925 A tool is a tool. If the new tool is worse than the old one, it is not worth using. Simple as that.

[@4.0.4]

While it kinda looks like hot garbage, here's hoping the license attracts people into finetuning some taste into it. Though Flux is probably a better bet.

[@oneanother1]

Where the girl lying on the grass comparison? I hope sd35 gains tractions so we can get more tools for it. Sd35 is faster than flux right?

[@ookamimizo6231]

Do you have a dedicated forge Playlist i can't find it

[@OlivioSarikas]

Not yet, but good idea

[@ph0enixph0enix65]

What about that absolutely absurd license model of SD 3? Did they change it with 3.5 to something acceptable? Is SD 3 still banned on CivitAI?

[@luisbriceno5242]

They unbanned SD 3 months ago.

[@ph0enixph0enix65]

@@luisbriceno5242 So the licence has been changed, I suppose? I didn't follow SD 3 news after that licence disaster and turned towards Flux (and other models).

[@luisbriceno5242]

@@ph0enixph0enix65 Yes, the SD 3 license went from really bad to a bit bad and that's it.

[@OlivioSarikas]

Free for private and commercial for anyone under 1 million dollars per year

[@ph0enixph0enix65]

@@OlivioSarikas Ah, I see. Thanks for clarifying. 😀

[@FuZZbaLLbee]

Is ComfyUI.exe still dependent on the installed NVIDIA toolkit?

[@wrillywonka1320]

olivio. I keep hearing about comfyui and so far every attempt to install it has had errors with models and nodes. now im hearing about comfyui exe but nobody can tell me how to find it. is this like only for priviledged people ooooorrr.......? am i missing something?

[@zacharyshort384]

Didn't hear too much enticing info over SDXL or Flux... or even SD 1.5 for that matter :P

[@bobtahar]

With Flux, Forge UI is faster than comfy ui... I learned to use ComfyUI simply to use Flux but rarely use it now because of it's slowness.

[@nickd9274]

Will ComfyUI.exe still work in combination with Nuke?

[@KDawg5000]

I'm having fun playing w/3.5 so far. And in ComfyUI, I set the bar on the left to small, and I turn off that little widget in the bottom right (not really useful to me).

[@Masterpouya]

@@KDawg5000 Wow how do you do this "search button" in your comment? I'm impressed!

[@KDawg5000]

@@Masterpouya Sorry, I don't understand your question. Can you rephrase?

[@Masterpouya]

@@KDawg5000 Sure! First time I read your comment , the word "ComfyUI" in your comment was written in blue with a "magnifying glass logo" on top of it. It was clickable and when I clicked on it, it searched "ConfyUI" in my youtube app. So I was impressed that there is a specific syntax to do so. But this time, everything is back to normal and the logo and blue color are not there anymore, so I'm not sure if it was some kind of youtube comment bug or if you did something intentional to make it this way.

[@AkoZoom]

​@@Masterpouyathis seems coming from YOUR own browser as an extension.. As edge does, opera too.

[@Robert_Marks]

Only Ideogram is the best realism model for lifelike images, flux and SD you can easily say are AI generated

[@aurelienb9109]

Did you try your SUPER Flux workflow technique with Stable 3.5 ?

[@iladdiewhiskynerd4924]

Does it also have a Mac version?

[@peterpui7219]

ComfyUI I prefer on FLOW UI

[@StargateMax]

SD3.5 totally froze my PC, I'll stay away from it and just use Flux for now.

[@holdthetruthhostage]

8 billion parameters isn't up you think by now they be shooting for I'd say 4B to about let's say 8B for medium and for the large probably 16b The problem with the image AI generation is there two conservative when it comes to the parameters

[@dreamcore7]

I doubt that consumer GPU's could handle 16b parameter models

[@dowhigawoco]

The time will show, but actually it cant compare to Flux

[@flcdd]

Can either 3.5 large or the Turbo one be used in Forge?

[@jeethechamp7458]

Didn't you watch the video? He said NO at the very beginning.

[@madrooky1398]

"don't be salty in the comments" Aww, what to do now instead? 😛

[@Elwaves2925]

Be saucy? Be juicy? You could pepper the channel with comments in those styles. :-)

[@AB-wf8ek]

A spoonful of salt is what makes the advertisements go down. Rage baiting is the sugar of the attention economy. A little bit of outrage is healthy, but in excess creates diabetes of the mind.

[@lennoyl]

I don't understand why there's no easy way to save a workflow (a save button, for ex) in the new comfyUI interface. The first time I wanted to save a basic workflow I modified as a new one, that UI renamed my basic workflow and overwrote it. It's not user friendly. Hiding the links will be used by mean people that don't want to show their workflows because they sell them

[@zombieploios]

There is literally a save button that does exactly what you said...even if you leave default name without adding a name (like you should so you find it easier) it wont overplace the old one it will just add (1) or whatever number to the name..

[@OlivioSarikas]

top left ist the save, save as and export all in one button

[@deadlyrobot5179]

Pls don't have the same face issue Pls don't have the same face issue Pls don't have the same face issue Pls don't have the same face issue Pls don't have the same face issue Pls don't have the same face issue

[@videosfeoscomotucara9038]

try with realism promts

[@Mooooty]

SD 3.5 results still looks 1990's cartoonish clip art.

[@caracal4361]

meh

[@yokipop9467]

lol its alr 3.5

[@brainwithani5693]

SD 3.5 IS NOT BETTER THAN FLUX

[@isabellatang7753]

no comfyui, no thanks

[@joeterzio7175]

comfyui, no thanks

[@piteshbhanushali1140]

They never fix 3 and dont release 8 gb large model..downgrade of stability..

[@fr0zen1isshadowbanned99]

Fantastic that none of the needed Models are linked on any of the Pages. And even the Search Engines don't know anything about Clip_g and Clip_i... a perfect start again.

[@OlivioSarikas]

Dude, I'm showing the links in my video and the pages are linked under the video. Maybe look first before you complain