Right now Sana looks slightly better than SD 1.x that has been taught to work with larger resolutions and captions. But worse than vanilla SDXL. Most faces look ugly, objects morphing into each other, messiness on the edges of objects, irregularly shaped circles (pupils, lenses, buttons). Parallel lines (grids, parts of the ship) are distorted. But the text looks better than in SDXL. It would be fun to compare SD1 (with upscaler) models like Juggernaut, Dreamshaper, Deliberate, Colorful with Sana
Thanks for the update! The SANA model: Doesn't do hands or feet - realism is bad - does not follow instructions - no I don't like it. None of the tests gave satisfactory results. Disappointed? no. There are some good models out there one could use instead. - And as always.
Flux is my main. It is a beast. Flux pro is incredible but my Flux Dev is better cuz i tun it locally with loras. Midjourney gives the best image compositions but Flux gives tbe best quality. No one else is close.
What does flux do better than MJ? I love MJ and the ability to edit images and fine tune them is better than any other model. The biggest issue with MJ is that it's not that good with understanding prompts. Ideogram is better. How is flux in these regards? What is its biggest advantages over MJ?
@@adarwinterdror7245 MJ is not even close to the personalization and edits we have with local models lol. Have you ever looked at what local users are doing for years? Loras, inpainting, outpainting, creative upscaling... Everything was here years before MJ and is better. MJ is better to get an initial aesthetically pleasing image (composition wise and lighting mainly). But youre very limited compared to what you can do with a local finetuned model.
@@GCT_777 I haven't delved into local AI image generation. No. My GPU is 3070 so i dont know if it's good enough. I guess it is. But installing looked kind of complicated and got me into dependency-hell every time I looked into it in the past, like a year ago. Maybe today things are better. Does it use Pinokio? And ComfiUI looks kind of intimidating, but it might look like it from the outside... In any case, MJ is great at transferring styles across images, out and inpainting from one image to others, etc. and the new Edit mode with the retexture seems like it will be useful for me. I don't care who did it first. What I care about is what is available NOW and what is better NOW. If there is good reason to go Local, I might indeed try it. What I'm looking for the most is AI that understands my prompts. Nothing compares to Dall-E in this regard, but Dall-E's images suck. Can you help point me into understanding better what is available Locally, and how to gets started in installing it? Also what is better in your opinion - SD or Flux?
@@adarwinterdror7245 MJ is garbage compared to what Flux can do. MJ is so far behind right now. At least 6-12 months behind, and all their future projects are things that are already available with Flux/SDXL on comfyUI/invoke
I've been playing with SD3.5-Large and finding it good for a base gen as I like the compositions (seems to do better with fanatsy illustrations at least), and then Flux Dev for upscaling and inpaint.
It's definitely exciting to see NVIDIA bringing Sana into the space, but I agree that its early results show room for improvement compared to models like Flux and even SDXL. The main attraction seems to be its speed and lower hardware requirements, yet "4K garbage is still garbage" is totally warranted. It’s clear that resolution alone isn’t enough if the quality and realism fall short, especially with faces and complex details. But with NVIDIA’s track record, it’ll be interesting to see how they refine this and whether it becomes a true alternative for both speed and quality.
I've said this several time on several videos, but I keep brining it up because creators keep liking to push the sensationalism of "midjourney killer". Options in models and services is great. I use many different ones myself, including both Midjourney and Flux locally depending on my needs. But there is much more to MJ than the model, so thinking every new model might be an MJ killer is absurd. There is a lot more going on under the hood in MJ, including the fact that your prompt itself is modified before it is passed to their model. For the average user, none of these models, no matter how good they are, are going to be an MJ killer in the broad landscape. For people who are extremely good with prompting, then yeah some of these models may come to the same level or slightly better - and being offline makes them better by default - but for average users, it isn't something that can be achieved with a new model alone.
In images that I generated on MIT's test site, I saw problems with fingers (7 on a hand) and with text ("chicken scratches" that only vaguely resembled the specified letters and words). They've got a ways to go yet.
I am most unimpressed. 1st comparison: plain grey background; flat, simple details; a woman with horns on overlaid with fire clipart. 2nd: you seem to have spelled "halfling" wrong; minimal background; I cannot complain much otherwise, given the minimal prompt. 3rd: wrong prompt is shown, no idea. Grass: SANA looks about 2 years behind. It is new and I am in a bad mood though, I'm sure it'll improve.
@@jibcot8541 Yeah but if the images you get out of it are 30x worse? Then it's pointless. Let's hope they keep training it because it seems under-cooked to me.
I'm sorry to tell you that Nvidia has a habit of launching such projects and then abandoning them, they were the first to produce AI black and white to colour images and that project was abandoned way back, every Nvidia project is just a demo.
4K garbage is still garbage. Very pathetic so far even SD1.5 with a 8GB GPU can do better (not faster of course), hope Sana gets better otherwise we can safely forget about it. And I wouldn't call a computer with 16GB GPU a "potato", that's the dumbest thing I ever heard. Did I heard that wrong ?
Right now Sana looks slightly better than SD 1.x that has been taught to work with larger resolutions and captions. But worse than vanilla SDXL. Most faces look ugly, objects morphing into each other, messiness on the edges of objects, irregularly shaped circles (pupils, lenses, buttons). Parallel lines (grids, parts of the ship) are distorted. But the text looks better than in SDXL. It would be fun to compare SD1 (with upscaler) models like Juggernaut, Dreamshaper, Deliberate, Colorful with Sana
In the beginning you mentioned flux commercial use issue but what is sanas commercial use?
👋 hi
Hey 👋
@@RenzoAlba
👋
Thanks for the update! The SANA model: Doesn't do hands or feet - realism is bad - does not follow instructions - no I don't like it. None of the tests gave satisfactory results. Disappointed? no. There are some good models out there one could use instead. - And as always.
Flux is my main. It is a beast. Flux pro is incredible but my Flux Dev is better cuz i tun it locally with loras. Midjourney gives the best image compositions but Flux gives tbe best quality. No one else is close.
What does flux do better than MJ?
I love MJ and the ability to edit images and fine tune them is better than any other model.
The biggest issue with MJ is that it's not that good with understanding prompts.
Ideogram is better.
How is flux in these regards? What is its biggest advantages over MJ?
@@adarwinterdror7245 MJ is not even close to the personalization and edits we have with local models lol.
Have you ever looked at what local users are doing for years? Loras, inpainting, outpainting, creative upscaling... Everything was here years before MJ and is better.
MJ is better to get an initial aesthetically pleasing image (composition wise and lighting mainly). But youre very limited compared to what you can do with a local finetuned model.
@@GCT_777 I haven't delved into local AI image generation. No. My GPU is 3070 so i dont know if it's good enough. I guess it is.
But installing looked kind of complicated and got me into dependency-hell every time I looked into it in the past, like a year ago.
Maybe today things are better. Does it use Pinokio?
And ComfiUI looks kind of intimidating, but it might look like it from the outside...
In any case, MJ is great at transferring styles across images, out and inpainting from one image to others, etc. and the new Edit mode with the retexture seems like it will be useful for me.
I don't care who did it first. What I care about is what is available NOW and what is better NOW. If there is good reason to go Local, I might indeed try it.
What I'm looking for the most is AI that understands my prompts. Nothing compares to Dall-E in this regard, but Dall-E's images suck.
Can you help point me into understanding better what is available Locally, and how to gets started in installing it?
Also what is better in your opinion - SD or Flux?
@@adarwinterdror7245 MJ is garbage compared to what Flux can do. MJ is so far behind right now. At least 6-12 months behind, and all their future projects are things that are already available with Flux/SDXL on comfyUI/invoke
I've been playing with SD3.5-Large and finding it good for a base gen as I like the compositions (seems to do better with fanatsy illustrations at least), and then Flux Dev for upscaling and inpaint.
Could be they didn't give you halflings because you spelled it "handling", at least in the video.
I just tried it. Feels like I went 2 years back in the past. Have no clue who might need it.
It's definitely exciting to see NVIDIA bringing Sana into the space, but I agree that its early results show room for improvement compared to models like Flux and even SDXL. The main attraction seems to be its speed and lower hardware requirements, yet "4K garbage is still garbage" is totally warranted. It’s clear that resolution alone isn’t enough if the quality and realism fall short, especially with faces and complex details. But with NVIDIA’s track record, it’ll be interesting to see how they refine this and whether it becomes a true alternative for both speed and quality.
I'm pretty excited about this given I only have a 6gb gpu
It’s says 16 gb minimum not 6 gb
@@ApexArtistX I saw this but it doesn't make a lot of sense for a .6B model. We'll see when it comes out :)
I've said this several time on several videos, but I keep brining it up because creators keep liking to push the sensationalism of "midjourney killer". Options in models and services is great. I use many different ones myself, including both Midjourney and Flux locally depending on my needs. But there is much more to MJ than the model, so thinking every new model might be an MJ killer is absurd. There is a lot more going on under the hood in MJ, including the fact that your prompt itself is modified before it is passed to their model.
For the average user, none of these models, no matter how good they are, are going to be an MJ killer in the broad landscape. For people who are extremely good with prompting, then yeah some of these models may come to the same level or slightly better - and being offline makes them better by default - but for average users, it isn't something that can be achieved with a new model alone.
In images that I generated on MIT's test site, I saw problems with fingers (7 on a hand) and with text ("chicken scratches" that only vaguely resembled the specified letters and words). They've got a ways to go yet.
It’s going to be censored
can it be run locally, are the weights released to the public? if not, this can join the stack of models i absolutely don't care about.
Exactly.
Sana is clearly based upon what they've learned/built for DLSS
SANA sounds like a beast!
No it sounds like a girl
What was the name of the LLM? I can't find it
SANA looks awesome! Can't wait to try it out.
❤🔥🔥🔥🔥
I am most unimpressed. 1st comparison: plain grey background; flat, simple details; a woman with horns on overlaid with fire clipart. 2nd: you seem to have spelled "halfling" wrong; minimal background; I cannot complain much otherwise, given the minimal prompt. 3rd: wrong prompt is shown, no idea. Grass: SANA looks about 2 years behind. It is new and I am in a bad mood though, I'm sure it'll improve.
If it's 30x faster than Flux Dev that is still impressive.
@@jibcot8541 Yeah but if the images you get out of it are 30x worse? Then it's pointless.
Let's hope they keep training it because it seems under-cooked to me.
I'm sorry to tell you that Nvidia has a habit of launching such projects and then abandoning them, they were the first to produce AI black and white to colour images and that project was abandoned way back, every Nvidia project is just a demo.
@@W-meme Sounds like every new tv show on Netflix hahaha
4K garbage is still garbage. Very pathetic so far even SD1.5 with a 8GB GPU can do better (not faster of course), hope Sana gets better otherwise we can safely forget about it.
And I wouldn't call a computer with 16GB GPU a "potato", that's the dumbest thing I ever heard. Did I heard that wrong ?
not only they need to use Nvidia cards but also the most expensive on the market :D
If it's not local it's trash, if it won't work with low v-ram (even high end cards nvidia released are low vram) then it's trash. Flux was awful.
How was flux awful? You're basically saying that anything you can't get for free with zero effort and shitty hardware is awful... maybe you're awful.
The examples makes sana looks bad. It may be fast but it's not up to today's standard.
Tried the online demo. Got the ugliest image I have ever generated.
RIP 8 to 20 vram GPUs
CCP involved project? No, thanks.
Flux, Stable Diffusion... now SANA? Loving this competition! 🤩🤩🤩🤩🤩🤩