I really like your perspective and teaching style! Thank you for making these videos! As a newbie in diffusion model pipelines, I sometimes get lost in terminology in terms of what things do on practice. It would be really valuable to have high-level overview of different pipeline parts: checkpoints, encoders, loras, control-nets, IP adapters, VAE, CLIP, Samplers etc and most importantly what effect they have on the final result
Thank you very much for sharing your information about the new model! From my point of view there is no need to hurry, because you already see some fast developments in improving the schnell and dev-versions .
I thought the girl in the thumbnail was going to be the youtuber talking, I was shocked to realize that it was actually the generated image. It was great for me, considering I'm really into this topic.
Flux is excellent, and to think that it is a base model. it promises a lot. I hope Matteo will develop IPadapter soon. The only problem I see is sometimes not understanding the depth of field, the blurs are very exaggerated and the image does not look organic in some cases.
Enjoyed your Flux review. It is clear that you have expert knowledge as to how all this works together in order to get a desired result. Have you considered creating a sort-of walk-through video of your workflows that identify the parts.. like the checkpoint you choose provides this.. this sort of node does this to the image.. you mention IPadapters and controllers and so on.. I am just a hobbyist and I have used much of this.. but its really just a lit of tinkering and getting lucky.. or not. If any such video, or even a text and graphics source, for this stuff exists on a basic level that would allow someone to knowingly put together workflows like yours while knowing what goes where and why... that would be very helpful.
As things stand now, VRAM is the primary concern, so 3090s and 4090s have a clear edge over any other (consumer) cards. But before pulling the trigger on any specific GPU, if I were you I’d try testing different GPUs on remote services like RunPod, Vast.ai, etc, just to get a sense of what the differences are.
Discussed it on reddit as well, but I find that Flux still has that "MidJourney" smooth effect that makes it less suitable for images in production. still a great model though, waiting for more development, I think the best way to use it now is for fast iteration and then img2img with traditional SD techniques
I don’t like the overtuned “MJ” look too, but that’s usually taken care of by finetunes. And to be quite honest, we never had a open source BASE model with this high of a quality, so I’m way more forgiving of art direction choices in this particular case!
1. Yes, I think flux is awesome, I tried Stable diffusion on Mimicpc, and of course this product also includes popular tools for AI such as RVC, Fooocus, and others. I think it handles detail quite well too, I can't get away from detailing images in my profession and this fulfills exactly what I need for my career.
I am an art director and I work on advertising campaigns. What would be the best AI to generate product images? Or reproduce an image in different ways? thanks
My advertising pipelines usually involve workflows like these using Stable Diffusion: Magnific AI Relight is Worse than Open Source th-cam.com/video/GsJaqesboTo/w-d-xo.html MJ is fine for pre-viz and moodboards, and maybe for background plates, but not for actual production.
I'm pretty terrible at LoRAs and finetunes! When I work on commissioned projects, I have a couple of people who specialize in LoRA training and I'd rather have them take care of that. So I wouldn't feel comfortable lecturing others on how to train LoRAs and finetunes, but there's others on YT who are great at it!
hi! you can't use any IPAdapter as of now, because at the time of writing this there's no Flux IPAdapters. Flux has a completely different architecture than 1.5 or SDXL, so it's not compatible. however, XLAB's developing a IPAdapter for Flux, and I'd be surprised if Matteo (the dev who released IPAdapter Plus and IPAdapter Advanced) wasn't working on his own implementation. same thing for Flux LoRAs and non-Flux models, different architectures, so they're not compatible.
The only exception is that you can still obviously use an SDXL model as a “refiner” or as an image 2 image, or in paint a Flux image with a 1.5 or SDXL workflow, you just can’t directly merge the models themselves. But you can achieve something similar with those techniques. I’ve been generating Flux images for the composition then using 1.5 to inpaint the results.
on the one hand: I haven't shot a single thing since May. on the other: I have never in my life been so submerged with clients. yeaaaah, the direction all this is taking is pretty indicative of the future.
@@risunobushi_ai well, that's good news, really. I still can't really find where to spend my SD skills professionally here. I don't really find so many people interested in a prompt master yet.
I don't know if there's a market for prompters per se, as I'm not in that market. I know for a fact though that there is a flourishing market for workflow devs for optimizing internal processes.
What type of work you do by using comfyui and how you sell it. I also want to do it , can you tell me where and what i can sell for professional work ?
hi! nowadays I'm mainly a workflow dev. that means I develop and build automated workflows for companies and brands, in order to streamline media creation. I also do genAI R&D with either freelancer or client-provided engineers, to further optimize internal uses of genAI. as for how to get into a job market like mine, developing stuff and releasing it for free as open source definitely helps. but then again I have 10+ years of experience in the fashion and product photography field and 3+ years in the genAI field, so I'm kind of sought after as I have both the technical knowledge and the creative vision needed for these kind of jobs.
it pains me that dev is recieving more sttention than schnell! schnell is certainly capable, fast, commercial, and overtime im sure it can be easily tuned.
Sorry, I wanted to put the link in the description but it got lost amidst the heaps of links. Here it is: www.reddit.com/r/StableDiffusion/s/f0MfliDd59
@@risunobushi_ai I've tested it a bit with comfyui, there is tutorial by cgtoptips but my tests show that it's not that great. I've tested with both dev and schnell and dev fails miserably and messes up, no understanding of surrounding image. I used the workflow file by same channel I mentioned. It works for simple stuff he has in video but quite limited. I'm not sure about this differential diffusion node. Maybe you could do inpainting tutorial with flux? 😊
what you'd basically want to do is use the inpaintmodelconditioning node to prep the latent, and the differential diffusion node between the checkpoint and the ksampler, using a grow mask with blur node with a high blur value to better prep the area for inpainting. It'd be a 1 minute tutorial lol, but I'll add that to the list!
I've had no problems with many different input sizes, I use 1280 a lot. Dev does poor initial generations txt 2 img at higer resolutions but it does img to img up to 3000 picels without issues.
Are you using it as a base text 2 image model / workflow or as a complex img2img workflow with controlnets and LoRAs? I’m having no issues with simple text2image workflows, the issues arise in the img2img workflows (because resizing a image to the correct latent size accepted by Flux is not as straightforward) with the XLAB’s ControlNet (because adding the XLAB’s ControlNet breaks some resolutions).
@@risunobushi_ai Both, works fine either way. The problem with Flux is that seeds are so varied. To be sure of something you need to test it on 10 or so seeds because one seed might be garbage and the next perfect.
Sorry, I didn't realize who you were, otherwise I wouldn't have asked such a basic question. I agree, in my experience the seeds vary a ton and have a sort of "pre-envisioning" of what a given image should look like. I don't know yet if that's a seed thing only, or a mix of seeds and t5 encoders, but regardless I'm not sure I like it as an approach, because it hard locks some generations way more than what I'm used to with 1.5 or XL. It might also be an issue for me specifically because I'm used to working with fixed seeds rather than random seeds in order to streamline processes. If I'm building a multi stage, complex workflow, I don't want to lose time regenerating things from scratch every time I change a node, and that's why I keep encountering "bad" seeds.
@@risunobushi_ai I usually use fixed seeds it's the only way if you have a fixed intent rather than random nonsense. One thing you might find useful is that you can use the guidence value to get variations. It seems it isn't the same as CFG I went up to 80 and kept getting good consistent generations on the same seed, Flux is very different, I am struggling with it!
Forge is really good as a UI, but the things I do with comfy (complex pipelines for clients) would either straight up be impossible with it, or would require bending over backwards. But for casual generations it’s amazing!
NF4 is fine if you’re hardware bound, but IMO I’d rather use FP8 as a quantized version. And then there’s the compatibility issues with controlnets developed for FP16, it’s a bit of a mess having so many different variations at this stage :/
i guess it was the mega image link... My question: Im getting very poor and blury images in the bottom workflow. How can i get good background images, what settings do i need? thanks
hi! yeah, links are not accepted by the auto moderation. are you using a fixed or random seed? what's your controlnet strength? what does your prompt looks like?
hi! yeah, links are not accepted by the auto moderation. are you using a fixed or random seed? what's your controlnet strength? what does your prompt looks like?
@@risunobushi_ai hey, Ive tried lots of different settings but right now: im using a random seed, controlnet strength is 0.66 (havnt changed it), I did change to the xlabs_flux_controlnet, prompts have changed a fair bit, currently: advertising photo of a bottle of spirits on a bench, at a dinner gathering. I will try put the link to workflow image below
I don't see anything wrong with the Dev license. It is good that people cannot make derivative model and ask money for it. It is dishonest in the first place,
Hmmm.. but one wonders if FLUX is derivative in any way... the FLUX folk all brought knowledge from elsewhere. Did they bring anything else? (PS, I am glad they did.. Flux helped this noob a lot already!)
I don’t see anything wrong either, a license is a license and as long as it’s clear it’s fine by me. I’m laying out the different licenses as something to be considered in choosing one dev path or the other, be it for users, for finetuners, or for devs.
Wait. Maybe I'm missing something here... so, the license on Dev variation only preferent the finetune models based on it but not the output images themselves? Cause, the first thing kind of name sense if a person try to sell a model based on that but put it behind a paywall. But what about the generated images? On Dev they are free for commercial purposes?
@@AlexUnder_BRGenerated images are free to be used by the individual who creates them, they say specifically that output of the model has nothing to do with them. They only own the model itself and any derivative of that Dev model.
what @ghost-user559 is saying: the dev non-commercial license covers only the model itself and its derivatives (i.e.: finetunes). a classic example in the space would be Pony's business model: they're developing and releasing the Pony checkpoints for free, but they're also giving the possibility to users to generate on their discord (iirc) via a bot that works à la MidJourney, so by paying. that would not be possible under the current non-commercial license without applying and obtaining a commercial license first. at least this is what I understand (I have a law degree but I'm not a lawyer) from the license. generated images are a completely different matter, and that's due to their nature: as far as some jurisprudence and informed opinions go, they would fall under public domain because of their nature, or for other interpretations they would fall under the rights of the people who generate them because of their "artistic" nature. be it as it may, regardless of which one interpretation is the correct one (and which one down the line will be held by Courts throughout the world), it'd be pretty hard to lock generated images under a non-commercial license.
New ControlNet Collection by XLAB dropped today, now we wait for a IPAdapter: huggingface.co/XLabs-AI/flux-controlnet-collections
This is an awesome review! I love that you're putting all the aspects in perspective. Thanks for sharing your honest opinion.
I really like your perspective and teaching style! Thank you for making these videos!
As a newbie in diffusion model pipelines, I sometimes get lost in terminology in terms of what things do on practice. It would be really valuable to have high-level overview of different pipeline parts: checkpoints, encoders, loras, control-nets, IP adapters, VAE, CLIP, Samplers etc and most importantly what effect they have on the final result
Thank you very much for sharing your information about the new model! From my point of view there is no need to hurry, because you already see some fast developments in improving the schnell and dev-versions .
I thought the girl in the thumbnail was going to be the youtuber talking, I was shocked to realize that it was actually the generated image. It was great for me, considering I'm really into this topic.
wow you're amazing, i'll use it for my bussiness
Flux is excellent, and to think that it is a base model. it promises a lot. I hope Matteo will develop IPadapter soon. The only problem I see is sometimes not understanding the depth of field, the blurs are very exaggerated and the image does not look organic in some cases.
DOF sometimes is way more exaggerated than I'd like, but yeah, as a base model it's fantastic image quality wise.
Enjoyed your Flux review. It is clear that you have expert knowledge as to how all this works together in order to get a desired result. Have you considered creating a sort-of walk-through video of your workflows that identify the parts.. like the checkpoint you choose provides this.. this sort of node does this to the image.. you mention IPadapters and controllers and so on.. I am just a hobbyist and I have used much of this.. but its really just a lit of tinkering and getting lucky.. or not. If any such video, or even a text and graphics source, for this stuff exists on a basic level that would allow someone to knowingly put together workflows like yours while knowing what goes where and why... that would be very helpful.
I’ve got a ton of videos on my channel ranging from basic stuff (the older videos) up till very specific, in depth stuff!
Thanks very good video very clear and precise instructions !
Hey! What GPU do you recommend? I want to run Flux locally but currently have a 1070 ti :/
As things stand now, VRAM is the primary concern, so 3090s and 4090s have a clear edge over any other (consumer) cards. But before pulling the trigger on any specific GPU, if I were you I’d try testing different GPUs on remote services like RunPod, Vast.ai, etc, just to get a sense of what the differences are.
Discussed it on reddit as well, but I find that Flux still has that "MidJourney" smooth effect that makes it less suitable for images in production. still a great model though, waiting for more development, I think the best way to use it now is for fast iteration and then img2img with traditional SD techniques
I don’t like the overtuned “MJ” look too, but that’s usually taken care of by finetunes. And to be quite honest, we never had a open source BASE model with this high of a quality, so I’m way more forgiving of art direction choices in this particular case!
1. Yes, I think flux is awesome, I tried Stable diffusion on Mimicpc, and of course this product also includes popular tools for AI such as RVC, Fooocus, and others. I think it handles detail quite well too, I can't get away from detailing images in my profession and this fulfills exactly what I need for my career.
I am an art director and I work on advertising campaigns. What would be the best AI to generate product images? Or reproduce an image in different ways? thanks
MidJourney
My advertising pipelines usually involve workflows like these using Stable Diffusion: Magnific AI Relight is Worse than Open Source
th-cam.com/video/GsJaqesboTo/w-d-xo.html
MJ is fine for pre-viz and moodboards, and maybe for background plates, but not for actual production.
Hey, Can you make a video on how abouts of finetuning (LoRA) flux models ? I came across SimpleTuner a while ago
I'm pretty terrible at LoRAs and finetunes! When I work on commissioned projects, I have a couple of people who specialize in LoRA training and I'd rather have them take care of that. So I wouldn't feel comfortable lecturing others on how to train LoRAs and finetunes, but there's others on YT who are great at it!
Can we use Flux with IPAdapter? How about combining Flux Lora realism with sdxl model checkpoint?
hi! you can't use any IPAdapter as of now, because at the time of writing this there's no Flux IPAdapters. Flux has a completely different architecture than 1.5 or SDXL, so it's not compatible. however, XLAB's developing a IPAdapter for Flux, and I'd be surprised if Matteo (the dev who released IPAdapter Plus and IPAdapter Advanced) wasn't working on his own implementation.
same thing for Flux LoRAs and non-Flux models, different architectures, so they're not compatible.
The only exception is that you can still obviously use an SDXL model as a “refiner” or as an image 2 image, or in paint a Flux image with a 1.5 or SDXL workflow, you just can’t directly merge the models themselves. But you can achieve something similar with those techniques. I’ve been generating Flux images for the composition then using 1.5 to inpaint the results.
I am, or better I was a fashion photographer 😂😂😂😂
As time passes by, I'm afraid the more and more of us WERE photographers here.
on the one hand: I haven't shot a single thing since May.
on the other: I have never in my life been so submerged with clients.
yeaaaah, the direction all this is taking is pretty indicative of the future.
@@risunobushi_ai well, that's good news, really. I still can't really find where to spend my SD skills professionally here. I don't really find so many people interested in a prompt master yet.
I don't know if there's a market for prompters per se, as I'm not in that market. I know for a fact though that there is a flourishing market for workflow devs for optimizing internal processes.
What type of work you do by using comfyui and how you sell it. I also want to do it , can you tell me where and what i can sell for professional work ?
hi! nowadays I'm mainly a workflow dev. that means I develop and build automated workflows for companies and brands, in order to streamline media creation. I also do genAI R&D with either freelancer or client-provided engineers, to further optimize internal uses of genAI.
as for how to get into a job market like mine, developing stuff and releasing it for free as open source definitely helps. but then again I have 10+ years of experience in the fashion and product photography field and 3+ years in the genAI field, so I'm kind of sought after as I have both the technical knowledge and the creative vision needed for these kind of jobs.
it pains me that dev is recieving more sttention than schnell! schnell is certainly capable, fast, commercial, and overtime im sure it can be easily tuned.
what's the prompt for 16:20? or maybe link to the source 🙏
Sorry, I wanted to put the link in the description but it got lost amidst the heaps of links.
Here it is: www.reddit.com/r/StableDiffusion/s/f0MfliDd59
@@risunobushi_ai many thanks! You're the legend
Will flex be able to be trained by community to do inpainting properly?
from what I've seen and tested, Flux is already pretty great at inpainting, moreover if you pair it with the differential diffusion node
@@risunobushi_ai I've tested it a bit with comfyui, there is tutorial by cgtoptips but my tests show that it's not that great. I've tested with both dev and schnell and dev fails miserably and messes up, no understanding of surrounding image. I used the workflow file by same channel I mentioned. It works for simple stuff he has in video but quite limited. I'm not sure about this differential diffusion node. Maybe you could do inpainting tutorial with flux? 😊
what you'd basically want to do is use the inpaintmodelconditioning node to prep the latent, and the differential diffusion node between the checkpoint and the ksampler, using a grow mask with blur node with a high blur value to better prep the area for inpainting. It'd be a 1 minute tutorial lol, but I'll add that to the list!
@@risunobushi_ai +1 also interested in an inpainting tutorial of Flux. Keep up the good work
I've had no problems with many different input sizes, I use 1280 a lot. Dev does poor initial generations txt 2 img at higer resolutions but it does img to img up to 3000 picels without issues.
Are you using it as a base text 2 image model / workflow or as a complex img2img workflow with controlnets and LoRAs? I’m having no issues with simple text2image workflows, the issues arise in the img2img workflows (because resizing a image to the correct latent size accepted by Flux is not as straightforward) with the XLAB’s ControlNet (because adding the XLAB’s ControlNet breaks some resolutions).
@@risunobushi_ai Both, works fine either way. The problem with Flux is that seeds are so varied. To be sure of something you need to test it on 10 or so seeds because one seed might be garbage and the next perfect.
Sorry, I didn't realize who you were, otherwise I wouldn't have asked such a basic question. I agree, in my experience the seeds vary a ton and have a sort of "pre-envisioning" of what a given image should look like. I don't know yet if that's a seed thing only, or a mix of seeds and t5 encoders, but regardless I'm not sure I like it as an approach, because it hard locks some generations way more than what I'm used to with 1.5 or XL.
It might also be an issue for me specifically because I'm used to working with fixed seeds rather than random seeds in order to streamline processes. If I'm building a multi stage, complex workflow, I don't want to lose time regenerating things from scratch every time I change a node, and that's why I keep encountering "bad" seeds.
btw, if anyone's reading this and is interested in exploring latent spaces and noise, @robadams2451 's videos are absolutely brilliant.
@@risunobushi_ai I usually use fixed seeds it's the only way if you have a fixed intent rather than random nonsense. One thing you might find useful is that you can use the guidence value to get variations. It seems it isn't the same as CFG I went up to 80 and kept getting good consistent generations on the same seed, Flux is very different, I am struggling with it!
I find it more convenient to use Forge for FLUX
Forge is really good as a UI, but the things I do with comfy (complex pipelines for clients) would either straight up be impossible with it, or would require bending over backwards. But for casual generations it’s amazing!
@@risunobushi_ai look at the mode___ l flux1-dev-bnb-nf4.safetensors ___the results will surprise you.
NF4 is fine if you’re hardware bound, but IMO I’d rather use FP8 as a quantized version. And then there’s the compatibility issues with controlnets developed for FP16, it’s a bit of a mess having so many different variations at this stage :/
@@risunobushi_airtx3060 - 12vram is the most common card, open source should suit most people's needs
background music is very interesting and misterious but distracting
I'm sorry, I'm not great at audio mixing and I thought the audio levels were alright. I'll lower it next time.
@@risunobushi_ai i think the choicd of the music is the issue there..
why are my comments being deleted??
i guess it was the mega image link...
My question: Im getting very poor and blury images in the bottom workflow. How can i get good background images, what settings do i need? thanks
hi! yeah, links are not accepted by the auto moderation. are you using a fixed or random seed? what's your controlnet strength? what does your prompt looks like?
hi! yeah, links are not accepted by the auto moderation. are you using a fixed or random seed? what's your controlnet strength? what does your prompt looks like?
@@risunobushi_ai hey, Ive tried lots of different settings but right now: im using a random seed, controlnet strength is 0.66 (havnt changed it), I did change to the xlabs_flux_controlnet, prompts have changed a fair bit, currently: advertising photo of a bottle of spirits on a bench, at a dinner gathering.
I will try put the link to workflow image below
I don't see anything wrong with the Dev license. It is good that people cannot make derivative model and ask money for it. It is dishonest in the first place,
Hmmm.. but one wonders if FLUX is derivative in any way... the FLUX folk all brought knowledge from elsewhere. Did they bring anything else? (PS, I am glad they did.. Flux helped this noob a lot already!)
I don’t see anything wrong either, a license is a license and as long as it’s clear it’s fine by me. I’m laying out the different licenses as something to be considered in choosing one dev path or the other, be it for users, for finetuners, or for devs.
Wait. Maybe I'm missing something here... so, the license on Dev variation only preferent the finetune models based on it but not the output images themselves? Cause, the first thing kind of name sense if a person try to sell a model based on that but put it behind a paywall. But what about the generated images? On Dev they are free for commercial purposes?
@@AlexUnder_BRGenerated images are free to be used by the individual who creates them, they say specifically that output of the model has nothing to do with them. They only own the model itself and any derivative of that Dev model.
what @ghost-user559 is saying: the dev non-commercial license covers only the model itself and its derivatives (i.e.: finetunes). a classic example in the space would be Pony's business model: they're developing and releasing the Pony checkpoints for free, but they're also giving the possibility to users to generate on their discord (iirc) via a bot that works à la MidJourney, so by paying. that would not be possible under the current non-commercial license without applying and obtaining a commercial license first. at least this is what I understand (I have a law degree but I'm not a lawyer) from the license.
generated images are a completely different matter, and that's due to their nature: as far as some jurisprudence and informed opinions go, they would fall under public domain because of their nature, or for other interpretations they would fall under the rights of the people who generate them because of their "artistic" nature. be it as it may, regardless of which one interpretation is the correct one (and which one down the line will be held by Courts throughout the world), it'd be pretty hard to lock generated images under a non-commercial license.