Hey folks, I totally forgot to mention that along with Comfyui you can also run SD3.5 large and medium on SDNext. SDNext is just like A1111 and Forge just with a slightly different UI.
I'm pretty sure we'll get some fine tunes from this model and the large model because there are some technical additions that makes SD3.5 model easier and more stable to train. Give it a couple weeks I'm sure we'll see some fine tune models soon!
Very helpful video, i noticed some stuff for example i was trying illustration stuff hand-made and dint get it that good on large but i got it at first try on medium, interesting indeed
I think it's very promising, quite better than sd3.0 medium and way better than base sdxl 1.0, with only 2.5B parameters it seems easier and there's more development support from the community
Agree 100% I'm sure once the model devs get going on training this will gain SD some ground on Flux. Too many people are forgetting how it was when SDXL first came out and the same people complained then too. 😬
They will start working on the updates to Forge this week and I did see a pull request on the Github regarding SD3.5 so it seems they are already working on it. Let's hope sooner than later.
IMO, SD3.5 medium doesn't compare very well to SD3.5 large or Flux but that might not be it's intention. It's closer to SDXL but with a few improvements and no real noticeable improvements to hands and anatomy (not counting your workaround). So far, I still prefer SDXL to SD3.5 medium, by a long way but we'll see what fine tunes appear. FYI, if you want to deal with that plastic look in Flux, I recommend trying the Flux Realistic Slow sampler (Forge only) with Beta and the Real Flux Dev model, or the Pixelwave model for Flux. It either removes it, or lessens it enough so it's not an issue.
@@MonzonMedia Of course, that's why I mentioned SDXL. I see people doing the whole "Is it a Flux killer?" cliched headline and it makes me roll my eyes. So I like to point out it's not comparable to anyone reading. I tried a Google Flan version of medium earlier and (hands aside) the results were impressive. Much better than just the base.
@@Elwaves2925 hahaha yeah they do it for views and I can't blame them but yeah those titles make me smh. When it comes down to it, it's great to have options. When it was just SDXL it was kind of stagnant for a while and even the fine tune models are all kind of the same thing. BTW I did a video recently on PixelWave, it's what I use mainly now since I found out about it. Great model.
@@MonzonMedia I found Pixelwave through you and it's fantastic. I find it has some minor blurry and pixelating issues with character loras (for photorealistic) but aside from that I love it. For loras like that I can still use RealFlux and Pixelwave for everything else.
How much does your cpu affect performance? I have a ryzen 5600 x3d, a 3080 ti with 128gb ram on an x570 dark hero motherboard, I was thinking of putting my 5950 x back in and seeing if that helps, but if not i dont want to tear down my loop. Thanks for the info bother.
It probably won't make too much of a difference since most things are mainly dependent on GPU when it comes to image generation. With that being said the 5950 would help you with other tasks and balance out your rig which would help your overall performance in general.
I’m wondering where to go, forge seems to have stopped (no innovations at all), a1111 is frozen, the choice is to wait for Comfy Desktop, or switch to InvokeAi or SD Next
The latest on forge is updates started yesterday. Invoke is good but the memory management isn’t the greatest. So cards that have 16gb and up suit invoke Ai for now until they improve memory management.
I can render 1440x1440 on SDXL but until I have access to a SwarmUI or A1111 out of the box edition of SD3.5 I don't see me being more than a spectator. I'm just too new to the whole image rendering ecosystem.
True but SDXL native resolution is 1024x1024, with SD3.5 Medium you can generate over 1440 without deformities. Swarm should be able to run it since the backend is comfyui. I haven’t checked but should work.
@@MonzonMedia It kind of runs it but I get VRam errors with a 7900XTX even at 1024X1024 so I am guessing some node or back end setting in Swarm isn't firing off. Or AMD is causing me issues.
Yeah I see people are having issues with Swarm so I guess the devs need to make some adjustments. They are pretty quick so I suspect within the next couple days it should be good to go.
@@OcihEvE Did you try the AIO model with text encoders included? huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/tree/main Should work according to someone on my Discord. I haven't had the chance to test myself yet.
@@MonzonMedia The 'all in one' at the bottom. It does work. Both with default 1024x1024 baked in VAE than comes with Swarm and the stableDiffusion35VAE_official VAE that changes the default to 1440x1440. Renders are taking me roughly 12 minutes but it is rendering. Even though the estimates are saying 25 minutes. Only other thing to note is, The prompt token changed. Identical prompt on an SDXL had a token value of 74. In 3.5 is shows a token value of 52. Not quite sure how that works but it's something I noticed.
@@MonzonMedia the 3.5 large model is no better than a fp8 flux model....I'm talking about the same size...I can't even imagine this one, which must be at GGUF4 level
@@ragemax8852 There's no point in being a groupie for an AI model, just get your head working ... SD3.5 is terrible for hands, for example, and Flux is far ahead of the rest. SD3.5 fanboy?
What about the comparison to the turbo version? Is there any point in starting to use the medium version if you have the turbo version? I have basically the same hardware as you, only instead of a Ryzen I have an Intel i5, the same amount of memory and the same graphics card. In the turbo version, a 1024X1024 image takes 25.33 seconds, so?
It's a good question and honestly it's still too early for me to say but one big thing is the native resolution being 1440x1440 so you can generate bigger images. But I would think it will be more better when fine tune models are out. If speed is important to you then yeah turbo makes sense. Medium also tends to have more of a creative artistic feel to it where Large leans towards photorealism.
Hey folks, I totally forgot to mention that along with Comfyui you can also run SD3.5 large and medium on SDNext. SDNext is just like A1111 and Forge just with a slightly different UI.
Nice, it's really fast and not really a vram hungry model. hope the community starts from here and make amazing models.
I'm pretty sure we'll get some fine tunes from this model and the large model because there are some technical additions that makes SD3.5 model easier and more stable to train. Give it a couple weeks I'm sure we'll see some fine tune models soon!
@@MonzonMedia Great 🎉
Very helpful video, i noticed some stuff for example i was trying illustration stuff hand-made and dint get it that good on large but i got it at first try on medium, interesting indeed
Yeah it's kind of like Schnell where for more artistic things, it seems to adhere to those types of prompts better.
Great video and as always packed with valuable information! Thanks for sharing!
Appreciate it bro! 🙌🏼
I think it's very promising, quite better than sd3.0 medium and way better than base sdxl 1.0, with only 2.5B parameters it seems easier and there's more development support from the community
Agree 100% I'm sure once the model devs get going on training this will gain SD some ground on Flux. Too many people are forgetting how it was when SDXL first came out and the same people complained then too. 😬
Andy idea when sd 3.5 will work on Forge ?
They will start working on the updates to Forge this week and I did see a pull request on the Github regarding SD3.5 so it seems they are already working on it. Let's hope sooner than later.
@@MonzonMedia Thank you - great !
Hi Ermin, PicLumen will integrate Pony Diffussion soon. Would you like to make a video on this?
Bro 1440x1440 native is very exciting stuff😍
Definitely! When the fine tunes come out it will be even better!
New smart t5 models, if exceeded, will probably take all the most important things, and will not affect what is of less meaning.
I'm sure we will see that soon.
IMO, SD3.5 medium doesn't compare very well to SD3.5 large or Flux but that might not be it's intention. It's closer to SDXL but with a few improvements and no real noticeable improvements to hands and anatomy (not counting your workaround). So far, I still prefer SDXL to SD3.5 medium, by a long way but we'll see what fine tunes appear.
FYI, if you want to deal with that plastic look in Flux, I recommend trying the Flux Realistic Slow sampler (Forge only) with Beta and the Real Flux Dev model, or the Pixelwave model for Flux. It either removes it, or lessens it enough so it's not an issue.
Well it is a 2.5B model so we need to consider for it's size it's not too bad. Flux is 12B.
@@MonzonMedia Of course, that's why I mentioned SDXL. I see people doing the whole "Is it a Flux killer?" cliched headline and it makes me roll my eyes. So I like to point out it's not comparable to anyone reading.
I tried a Google Flan version of medium earlier and (hands aside) the results were impressive. Much better than just the base.
@@Elwaves2925 hahaha yeah they do it for views and I can't blame them but yeah those titles make me smh. When it comes down to it, it's great to have options. When it was just SDXL it was kind of stagnant for a while and even the fine tune models are all kind of the same thing. BTW I did a video recently on PixelWave, it's what I use mainly now since I found out about it. Great model.
@@MonzonMedia I found Pixelwave through you and it's fantastic. I find it has some minor blurry and pixelating issues with character loras (for photorealistic) but aside from that I love it. For loras like that I can still use RealFlux and Pixelwave for everything else.
How much does your cpu affect performance? I have a ryzen 5600 x3d, a 3080 ti with 128gb ram on an x570 dark hero motherboard, I was thinking of putting my 5950 x back in and seeing if that helps, but if not i dont want to tear down my loop. Thanks for the info bother.
It probably won't make too much of a difference since most things are mainly dependent on GPU when it comes to image generation. With that being said the 5950 would help you with other tasks and balance out your rig which would help your overall performance in general.
I’m wondering where to go, forge seems to have stopped (no innovations at all), a1111 is frozen, the choice is to wait for Comfy Desktop, or switch to InvokeAi or SD Next
The latest on forge is updates started yesterday. Invoke is good but the memory management isn’t the greatest. So cards that have 16gb and up suit invoke Ai for now until they improve memory management.
@@MonzonMedia wait )
There has been no new content for a very long time (
Been preoccupied with some personal projects but I have some content coming soon! Appreciate you checking in 👍🏼
I can render 1440x1440 on SDXL but until I have access to a SwarmUI or A1111 out of the box edition of SD3.5 I don't see me being more than a spectator. I'm just too new to the whole image rendering ecosystem.
True but SDXL native resolution is 1024x1024, with SD3.5 Medium you can generate over 1440 without deformities. Swarm should be able to run it since the backend is comfyui. I haven’t checked but should work.
@@MonzonMedia It kind of runs it but I get VRam errors with a 7900XTX even at 1024X1024 so I am guessing some node or back end setting in Swarm isn't firing off. Or AMD is causing me issues.
Yeah I see people are having issues with Swarm so I guess the devs need to make some adjustments. They are pretty quick so I suspect within the next couple days it should be good to go.
@@OcihEvE Did you try the AIO model with text encoders included? huggingface.co/Comfy-Org/stable-diffusion-3.5-fp8/tree/main Should work according to someone on my Discord. I haven't had the chance to test myself yet.
@@MonzonMedia The 'all in one' at the bottom. It does work. Both with default 1024x1024 baked in VAE than comes with Swarm and the stableDiffusion35VAE_official VAE that changes the default to 1440x1440. Renders are taking me roughly 12 minutes but it is rendering. Even though the estimates are saying 25 minutes. Only other thing to note is, The prompt token changed. Identical prompt on an SDXL had a token value of 74. In 3.5 is shows a token value of 52. Not quite sure how that works but it's something I noticed.
I think the Flux developers will take some kind of step forward, open a model or provide something new.
I hope so but likely not fully open source. Who knows though what they have in store?
But quality is so far of flux
You are comparing a 12b model to a 2.5b model. What do you expect?
@@MonzonMedia the 3.5 large model is no better than a fp8 flux model....I'm talking about the same size...I can't even imagine this one, which must be at GGUF4 level
@@ronbere Flux is not even better than SD 3.5 large, so why you lying, Flux fanboy?
@@ragemax8852 There's no point in being a groupie for an AI model, just get your head working ... SD3.5 is terrible for hands, for example, and Flux is far ahead of the rest. SD3.5 fanboy?
@@ronbere How long, though, Flux fanboy slash 3.5 hater.
What about the comparison to the turbo version? Is there any point in starting to use the medium version if you have the turbo version? I have basically the same hardware as you, only instead of a Ryzen I have an Intel i5, the same amount of memory and the same graphics card. In the turbo version, a 1024X1024 image takes 25.33 seconds, so?
It's a good question and honestly it's still too early for me to say but one big thing is the native resolution being 1440x1440 so you can generate bigger images. But I would think it will be more better when fine tune models are out. If speed is important to you then yeah turbo makes sense. Medium also tends to have more of a creative artistic feel to it where Large leans towards photorealism.