these are the mad scientists I remember, and no wonder SD felt like you were doing something interesting, these guys were running the show. I guess after a bit they left and then the other people took over and made SD boring.
@@allyourtechai I got in on beta wave 2, and it was like we were doing something special and it was fun. Then they started taking away the uniqueness of the experience. Yeah you're right.
Cool, I think the author explains it so well, according to the author's procedure, I also run Flux online on Mimicpc, and the effect is not much different!
Can this model be used on the same way as other SD models work? IE Loras, Controlnet, etc. In other words, is this the same architecture, or is completely different? And thanks for this nice video.
I believe those will all be possible very soon. The architecture is different enough that you can’t just drop this in as a replacement, but I think this model is impressive enough that development will happen quickly
Would like to see an apples to apples comparison between Flux & SD3 on similarly specified prompts, to see whether it in fact "completely destroys" SD3, or whether they're reasonably comparable...
Love how people are hyping up Flux so much because its what SD3 was supposed to be lol. I think SDXL and ponydiffusion have developed enough to a point where they are already incredible by themselves, especially because of the community, to the point where i'm not really impressed by new AI image models. I love that it adheres the prompts better and being able to generate text is also awesome, but those are things you could already add in tools like Canva anyway. What i'm excited to see more of, is the workflows for fully generated/animated high quality scenes.
Can this be locally installed on MacBook Pro Max M3, 96 GB? What is pricing/usage/credits etc of Pixel Dojo? Looks pretty cool. I’d love to support your product if I can swing it
You should be able to run the schnell model locally. It’s still very impressive. The pro model is not open source, but that’s what’s running on Pixel Dojo. It’s 25 per month and includes over 20 image tools
Great video, but you didn’t do the most important test hands and limbs 😅. Also want to ask if Flux can do Anime and can people create their own checkpoints or Lora equivalents with it?
Haha good point. It’s actually quite good at hands and feet which is also impressive. I don’t think you can do a Lora or equivalent just yet, but soon!
There is a resemblance slider in the options. Turn that up, and turn creativity down and you get the original image but upscaler with very subtle changes. I’ll make that a preset (subtle upscale)
But how well can it do the following things: 1. Different art styles e.g. paintings of various eras. 2. Novel concepts (e.g. "Stirrup-shaped bottle) 3. Selective details (e.g. Group photo of 50 people) or incarnate latticework 4. Photos that look like they're taken casually 5. Stylistically bad drawings (4-fingered scribbled hand that looks like it's drawn by a child)
I've been messing with it this evening locally in comfyui with the schnell version. for realism its definitely top notch and if thats your thing then this is really good, especially at various elements in an image. because of your question i went ahead and tried your #2 question as i hadnt tried anything like that, and it blew it out of the water. it gave me //exactly// this: "extreme close up of a beer can on a kitchen countertop, the beer can has an anthropomorphic image of a half penguin half fish smiling holding a can of the same beer he is printed on drinking it. extreme depth of field with focus on the beer can, volumetric lighting, masterpiece, ultra realistic" however in regards to your other questions, i've been experimenting with specific anime styles and it really struggles at hitting them even when i've been super descriptive with references to the specific artist or art style, as well as also trying to improve the prompt by running it through a prompt helper. it also struggles with characters as I'm assuming it has no idea who popular characters are. i tried very common anime characters like rei ayanami which failed to do her hairstyle completely, and also others like spike spiegel and it just completely failed. i didnt try goku but yeah if it failed at that i wouldnt be surprised. outside of anime, it also struggled with other painting styles. i asked it to do a remix of van goghs starry night and got nothing even remotely close to it. so to answer your question, 1 - 5/10, 2 - 9/10, 3 - 8/10 main beef i have with it is how slow it is. a 3070ti isnt enough unless you want to queue up a huge batch overnight. sdxl is about 15-20 times faster at generating images, and thats even with including latent upscaling on those (i haven't tried a latent upscale with this model yet). im doing 1024x1024 images and its roughly 680 seconds wait time for an image while as on sdxl i can spit those out in about 30-45 seconds if the models are loaded in memory before generating.
@@aylictal I've heard others say similar things about the speed. I'm running it on Nvidia H100's on pixelDojo ($25,000) gpu's, so there is a clear advantage in terms of speed. I'm betting the model will improve over time though.
UI Ux: Figma Frontend: HTML, CSS, Javascript, Backend: Javascript(NodeJS) Starting point: 1. Draw a red rectancle in the browser, if you click on it it get blue. 2. Draw a red rectancle in the browser, if you click on it it displays a picture 3. Draw a red rectancle in the browser, if you click on it it displays multiple pictures 4. Write a Script that takes a text as a input and throws a picture out (use local ai, f.e. comfy ui) 5. Write a Script that takes multiple text as a input and throws a pictures out (use local ai, f.e. comfy ui) 6. Stich carefully 1-3 + 4-5 together. Thats how I would do it. Maybe it helps^^
So, AMD does not auto work for Windows? And to make AMD gpu work, it must be in a linux environment? Help us detail a tutorial to use comfyui work specifically with AMD gpu. I have msi armor radeon rx580 8g on win 10. Thanks. Looking forward for your video.
@@allyourtechai The schnell and dev model are the same size. Dev is distilled for quality and prompt adherence, schnell is distilled for 4 step generation. Pro is not local.
If you don't want the nodes of ComfyUI, it works in SwarmUI, which is still Comfy but with a UI similar to A1111 and the others. Both the dev and schnell work.
@@allyourtechai Only the smallest model, or the dev model too? Seems those models are about the same size (while the dev being more capable judging by the graph in your video), so I guess the same Vram requirements too? Thanks.
@@fossil98 Thanks. I have 64GB installed, so FP16 works ok (when nothing is running in the background = more free ram). The question was more about Vram requirement. Meanwhile, I hear people with only 8GB Vram can run it, so good to see that. Dev model should indeed be more capable, but it takes 28s on my 4080 to generate an image (20 passes), while Schnell takes only 8 seconds (4 passes) without obvious image quality loss. So I think I will use that one and FP8 because I need more ram for other stuff I use on my PC.
