Downloading now to give it a try. One my first tests will be with crowd faces, or basically faces at a distance. The Schnell models always looked bad/distorted vs the Dev, so we will see how Shuttle does. UPDATE: After trying it in Forge, it uses so much memory that I can only do 1 image at a time w/my 4090. It has a lot of background blur, so I used the amateur image lora to get a more realistic image. That did improve the results, but still not 100% to my liking. Also it didn't like favorite sampler setup of DEIS and DDIM. Note if you look in the command window you can see this message "Distilled CFG Scale will be ignored for Schnell". So adjusting that really won't make a difference if I understand correctly. Now to test in ComfyUI to see if it works better.
Thanks very much for your videos. I feel the schnell model has less of an AI look. Especially the I Love You example where it looks more photo like. Slightly faded text on the card and more natural skin tone.
Thank you for the headsup. One thing I miss and maybe we can cover is - asymetric tiling. I guess FLUX and SD3 totally cant do it like SDXL is able to...
You don't need the second step thing. You can right click an image in the preview chooser to save the image, but connect the SD ultimate to the image output of the preview chooser to continue the processing if you want.
@@Zenda_Nior I know. I kind of mad too. But ... higher hardware requirements are a natural consequence of technological progress. Models are getting more complex. You can have deviation from that, like something highly optimized, but ...
@@Zenda_Nior After some more testing Yesterday, I sadly have to tell you that you are right with the 50GB. It ran pretty slow for 4 Steps and then I tried the .gguf some other Commenter suggested. That failed and just ate all normal 64GB RAM before crashing. So back to the normal Shuttle. I still had the Task Manager open and so I took a look at the Shuttle RAM usage... yea... 50 - 64GB usage depending on if you generate or idle. It seems that it just so fits into a 24GB VRAM and 64GB RAM System :D I even downgraded the t5xxl_fp16 to the t5xxl_fp8_e4m3fn before checking ^^ But still great Pics! :)
Thanks for the video! Though you hid most of the nodes links by excessively zooming in. You don't understand that in many countries, Patreon is banned, and we cannot pay you to download the complete workflow. Though, the first group worked and the photo realistic images are terribly plastic.
Cheers Olivio. I'll give this a go but as it's trained off Flux Schnell, I can see me sticking with RealFlux and Pixelwave. Still, it's nice for those who want something quicker.
Is this trained on schnell or dev or 1.1 pro? It has big file size. Could this also have a lot of anime characters from say, Dragon ball? I had tested a lot of characters on Dalle 3 in the past, and it perfectly understood a lot of them, including unexpected ones. Is Flux also trained like that? What is best upscaler for photorealistic landscape images? Currently all upscalers(except online AI services that cost a bank) totally ruin landscape images, nothing seems to work.
Create still images with Flux of the first frame of each scene you want in your video. Then use one of the image to video services to generate the clips. Finally, use a video editor to edit and bring it all together with the music.
Someone was talking about a smaller Model already in the Comments. I find it for you. "shuttle-3-diffusion-Q4_K_S.gguf. Memory A: 7.68 GB, R: 8.53 GB, Sys: 9.7/12 GB (80.6%) on GPU 3060 12 GB" For other "refinements", I can't say anything, but I would love some more "Adult Content" with it. So far I am the most impressed I ever have been with any Model!!!!!!!!!
@@fr0zen1isshadowbanned99 Hi ! Thank you ! ^^ I'm more looking forward towards models trained on this but with better images or a specific style. I guess we will see some on civitai soon.
@OlivioSarikas the speed difference between the 2 models you're comparing, and the minimum vram requirements. You say "it's super fast" and half way through the video still didn't make clear how fast. Comparing generations is great for new model, but so is speed and vram requirements.
Not works for me, using forge and 32gb ram with rtx 3090, im using same textcoders/vae but forge is crashing every generate with this model Extra test note: This model is incredibly insufficient with 32 GB RAM and it locks the system. It also destroyed the virtual HDD memory along with 100% VRAM and 100% RAM usage. It seems that above 32GB should be used with 64GB RAM if possible. "For shuttle 3 diffusion" not other shuttle models.
Downloading now to give it a try. One my first tests will be with crowd faces, or basically faces at a distance. The Schnell models always looked bad/distorted vs the Dev, so we will see how Shuttle does.
UPDATE: After trying it in Forge, it uses so much memory that I can only do 1 image at a time w/my 4090. It has a lot of background blur, so I used the amateur image lora to get a more realistic image. That did improve the results, but still not 100% to my liking. Also it didn't like favorite sampler setup of DEIS and DDIM. Note if you look in the command window you can see this message "Distilled CFG Scale will be ignored for Schnell". So adjusting that really won't make a difference if I understand correctly. Now to test in ComfyUI to see if it works better.
Thanks very much for your videos. I feel the schnell model has less of an AI look. Especially the I Love You example where it looks more photo like. Slightly faded text on the card and more natural skin tone.
Models are from 6.8 GB to 24 GB. It works great on my RTX3060 12GB. Thanks for tip!
witch work best for u?
Thanks for making these vids !
Thank you for the headsup.
One thing I miss and maybe we can cover is - asymetric tiling. I guess FLUX and SD3 totally cant do it like SDXL is able to...
Examples are impressive, will try it out!
You don't need the second step thing. You can right click an image in the preview chooser to save the image, but connect the SD ultimate to the image output of the preview chooser to continue the processing if you want.
How to solve "Tuple index out of range" in comfyui?
Works fine on 8 VRAM / 32 RAM in Comfyui with Loras and gives very interesting results.
Is this for commercial use? Can't find info on the huggingface page or their website or docs?
Forge. Model: shuttle-3-diffusion-Q4_K_S.gguf. Memory A: 7.68 GB, R: 8.53 GB, Sys: 9.7/12 GB (80.6%) on GPU 3060 12 GB
Thank you for your information, it makes me curious what you can do with it 🤓
Amazing, thanks for sharing your knowledge.
50GB model sets to get 4 step renders, requiring beefy hardware. How is this progress?
There is no 50 GB model.
@@JustFor-dq5wc I did say sets, didnt I?
Model + Diffuser + Encoder
Yes 50 is a bit exaggerated but not for long.
@@Zenda_Nior I know. I kind of mad too. But ... higher hardware requirements are a natural consequence of technological progress. Models are getting more complex. You can have deviation from that, like something highly optimized, but ...
@@Zenda_Nior After some more testing Yesterday, I sadly have to tell you that you are right with the 50GB.
It ran pretty slow for 4 Steps and then I tried the .gguf some other Commenter suggested. That failed and just ate all normal 64GB RAM before crashing.
So back to the normal Shuttle. I still had the Task Manager open and so I took a look at the Shuttle RAM usage... yea... 50 - 64GB usage depending on if you generate or idle.
It seems that it just so fits into a 24GB VRAM and 64GB RAM System :D I even downgraded the t5xxl_fp16 to the t5xxl_fp8_e4m3fn before checking ^^ But still great Pics! :)
@@fr0zen1isshadowbanned99interesting as mfs ran it fine on 12gb 3060 with 32gb of vram
Thanks for the video! Though you hid most of the nodes links by excessively zooming in. You don't understand that in many countries, Patreon is banned, and we cannot pay you to download the complete workflow. Though, the first group worked and the photo realistic images are terribly plastic.
Thanks Olivio! Do you know if all flux LoRas work with this model?
Love the shirt Olivio!
Thank you, you can actually get it here and even get a $1 discount by using Olivio as a discount code :) www.qwertee.com/shop/tees/grumpycorn
Cheers Olivio. I'll give this a go but as it's trained off Flux Schnell, I can see me sticking with RealFlux and Pixelwave. Still, it's nice for those who want something quicker.
I wish we had seen more examples
hey Olivio what hardware/desktop do you recommend to use FLUX with that generates fast?
Hi, thanks Olivio for the video. I saw the size of the file is nearly 24 GB. I guess that means with a 16 GB VRAM running it will not be possible.
i have 16GB and it run fine. Someone on the comments said even 8 VRAM work with 32 RAM
@@OlivioSarikas thanks I'll give at a whirl 🙂
Could you make a new review on Invoke AI 5.3? I think they have FLUX as well as all the other stuff from before.
Is this trained on schnell or dev or 1.1 pro? It has big file size. Could this also have a lot of anime characters from say, Dragon ball? I had tested a lot of characters on Dalle 3 in the past, and it perfectly understood a lot of them, including unexpected ones. Is Flux also trained like that? What is best upscaler for photorealistic landscape images? Currently all upscalers(except online AI services that cost a bank) totally ruin landscape images, nothing seems to work.
Its schnell.
nice to know❤
thank you for these videos !! I would really like to make a music video from Flux how would you make a Music video using flux ?
Create still images with Flux of the first frame of each scene you want in your video. Then use one of the image to video services to generate the clips. Finally, use a video editor to edit and bring it all together with the music.
@@Elwaves2925 I need to be able to create the videos for free open source like flux in forge
I will post tutorials on that in my channel
Thanks . This is always so helpful and top news. Just discussion patreon support with my firm. I think this is a done deal.
What is the difference in speed between Shuttle and Schnell?
wicked
Hi !
Is it a refined model already or can we expect a fine tuned version on CivitAI of it soon ?
Someone was talking about a smaller Model already in the Comments. I find it for you.
"shuttle-3-diffusion-Q4_K_S.gguf. Memory A: 7.68 GB, R: 8.53 GB, Sys: 9.7/12 GB (80.6%) on GPU 3060 12 GB"
For other "refinements", I can't say anything, but I would love some more "Adult Content" with it. So far I am the most impressed I ever have been with any Model!!!!!!!!!
@@fr0zen1isshadowbanned99 Hi ! Thank you ! ^^
I'm more looking forward towards models trained on this but with better images or a specific style. I guess we will see some on civitai soon.
what make shuttle 3 diffusion model amazing is its license, it Apache!!!
I woul be more interested in quality then speed. I would love to get 1.1 pro quality local. Thats a video i would be interested in. Just in case ...
You are a great communicator and educator :) Thanks for the work you put into these
I prefer the look of the FLUX images, as shuttle ones look overtrained(?)
yes.
Disagree. It's better than flux schnell imo. But then again, who uses schnell anyway? Schnell is also a couple of months old now.
@@TheBann90”who uses schnell” i could see a lot in different AI servers 👍
How to turn on gpu weight
Why does every UA-cam ignore vram and speed as if that's not the most important thing we'd want to know first? Because of watch time 😑
because I don't have 10 computers to benchmark it on different configurations. Where would I find the Vram and Speed for each new model on release?
@OlivioSarikas the speed difference between the 2 models you're comparing, and the minimum vram requirements. You say "it's super fast" and half way through the video still didn't make clear how fast. Comparing generations is great for new model, but so is speed and vram requirements.
@@Tokaint it uses 4 steps - that's what i mean by fast. especially for that quality. how long 4 steps take depends on your system
0:05 I was told if I can't say something nice, don't say anything at all. I am doing _______
Time to discard Schnell ultimately, I see.
Don’t need fast need 6 vram flux not gguf
same plastic
Not works for me, using forge and 32gb ram with rtx 3090, im using same textcoders/vae but forge is crashing every generate with this model
Extra test note: This model is incredibly insufficient with 32 GB RAM and it locks the system. It also destroyed the virtual HDD memory along with 100% VRAM and 100% RAM usage. It seems that above 32GB should be used with 64GB RAM if possible.
"For shuttle 3 diffusion" not other shuttle models.