Developing Complex Flux Workflows Kinda Sucks

[@agusdor1044]

finally! Now I don't feel like the only one with this issues. We are in the same pague, Andrea. Thanks for sharing!

[@risunobushi_ai]

I was actually worried I was the only one, I was thinking “wait, maybe I shouldn’t post it, what if I’m stupid” lol

[@agusdor1044]

@@risunobushi_ai i was struggling with my fine tunned loras and ControlNets by Xlabs lmao FLUXSTRATIONS

[@SasquatchBC]

And I was worried that it was a skill issue for me too. You did the right thing by sharing your thoughts. Thanks for this!

[@antiplouc]

I agree 100% with you. So many nodes are half working. You have to change your entire workflows for a result that doesn't work.

[@robadams2451]

I agree completely, I don't think either the controlnets or the attempts at an IP adapter are well founded. As far as I know Matteo is holding off with IP adapter for Flux until some of the technical obstacles are sorted. I have found some workflows transfer well and others have unexpected results. Things like tiled upscale work surprisingly well. Ordinary IMG to IMG is quite different with high levels of denoise required. Loras work very well in some cases and less so in others. I feel we are a way from being able to produce specific imagery in a reliable way, it will come though.

[@risunobushi_ai]

yeah, in my experience as well some thing work exceptionally well - like one image LoRAs, for example, which would have been borderline impossible (at least with these results) on other models. but the overall workflow building experience feels only halfway there, and I guess it's mainly because the tools we have now have the same name (and theoretical purpose) as the tools we have for older models, while not providing the same level of finesse. also I personally don't like natural language because I'm more drawn to visual based pipelines, but that's a me problem, that much I can admit.

[@robadams2451]

@@risunobushi_ai I've been testing natural language prompts... I think people are perhaps imagining many of the benefits. Flux can undestand relatives, beside, above etc... but it doesn't understand poetry! I took a few of the more extravagent prompts and slowly removed the words... to my eye in most cases the image improved and I felt consistency was better.

[@hempsack]

I think it's funny that it plays out like sd3 when you ask it to just make feet, or hands. Now they might come out better if you give it a more coherent prompt. I haven't gone that far into testing. But it is funny how it can make beautiful images of people with great looking hands and feet. But when asked to make them alone, the training lacks in that perspective. Just my opinion.

[@svenvarg6913]

A hundred % agree, the custom nodes, non-standard conditioning are odd choices.

[@crypt_exe8709]

Can't agree more with this XLabs sampler, I tried it on day one and really disliked the "proprietary" solution that they proposed, and the nonsens of this Controlnet Input, and why having both denoise and image to image paramters? I won't use this or spend more time on it, better using SDXL controlnet and Ipadapter workflow and using Flux as a img2img detailer on another chrome tab

[@expaintz]

great video! i was loosing my mind over this latly - so its not only me. great and informative

[@andreizdetovetchi]

Yes! You "read" my mind, I'm so 100% with you. While Flux is awesome, nothing really works as expected or as before, with SDXL for example. So many errors, the controlnets don't actually work as expected, nor the IPAdapter, and having troubles with LoRAs too.

[@kamruzzamanuzzal3764]

I 100% agree with the title, like usually do a few layers of pass, like first pony or sdxl and then again sdxl lighting and then adetailers. Now it feels like flux alone gives much more aesthetically pleasing images in most cases, no longer need any 2nd passes. Instead just save the images and then later upscale it with a lil bit of face detailer

[@svenhinrichs4072]

Just wasted 2 days experimenting with the new malfunctioning-controlnets and then laughed a lot seeing your video afterwards :))))

[@saumyadubey8255]

Wow so everyone is equally confused. Hoping you hear back from black forrest soon!

[@risunobushi_ai]

Yeah, at least I am. Flux is a fantastic model, but apart from its base fussiness, the ecosystem is just plain rushed imo.

[@denisquarte7177]

This is the expierence I was expecting. Similar to the SDXL release (remember the refiner hassle?) I just wait it out. We had to wait a while before controlnets emerged and even then it took ages especially for open pose to even become decent. That being said there is more frustration over this maybe because flux looks more promising than sdxl base at release.

[@risunobushi_ai]

oh god the refiner days. I think my frustration comes from the fact that, although I was working with SD1.5 and SDXL at release, it wasn't my full time job yet. Now that it is, and clients are picking up on new releases and tech much faster, I get requests that are simply not doable in a full Flux pipeline (and even if they were, I still have no definitive reply from BFL on their commercial license).

[@JeremyOpio-000]

Great video! Thank you!

[@heiferTV]

NO, You are not alone. I had experienced the same frustration with this. 100x100 agree. Tks. for the video, it served as therapy of this awful experience for me.XD

[@svenvarg6913]

I posted in the stable diffusion subreddit that we've accepted a model more restrictive than SD3 because we've been seduced by a good model.

[@risunobushi_ai]

the stable diffusion subreddit is an odd one - it seems to me like the majority of users are not power users (well, of course, that's to be expected and it's not their fault), so they completely bypass the legal implications of a license in a commercial setting, and how much companies and brands actually care for legal liabilities.

[@ArrowKnow]

Preach that truth Brother!!!

[@saurabhsswami]

yeah, just give it some time to cook. Right now, the flux model for me is a good starting image to modify with sdxl or sd 1.5 later

[@risunobushi_ai]

for sure, and I'm certain it will eventually get there - it's just a bit of a pain when clients ask for fully modular Flux workflows and you just know it's not there yet.

[@flyingdogsoft]

I am waiting for the Flux inpaiting model before I dive deeper into more complex Flux workflows.

[@serasmartagne]

I've had some success using the XLabs models with the standard "Apply Controlnet (Advanced)" node, with strength 0.43, start percent 0 and end percent 0.5. I use a standard KSampler with a Flux Guidance node for the positive conditioning, set to 4.0. I've daisychained depth and canny conditioning this way. I've had no success at all using the custom XLabs Sampler node.

[@risunobushi_ai]

Most of the fine precision stuff I do require controlnets at almost 100% strength, so 0.8+. I had some success too trying weird combos, but for the life of me I cannot fathom why someone would release something as non-experimental if it only works under very specific circumstances. SDXL controlnets before Xinsir were worse than 1.5’s, but they were nowhere this half baked. In my experience this situation ends up leading to complex multi-workflow setups, where in order not to run out of memory or having basic modularity I have to segment going from point A to Z in 3 or 4 different workflows :/

[@serasmartagne]

@@risunobushi_ai I haven't tried the InstantX stuff yet, but I expect to have a similar frustrating experience. I agree these extra tools on top of Flux are half-baked at the moment, I think we need to exercise some patience before workable additions are available. In my experience I've noticed that Flux workflows tend to break down when going above around 1.5 megapixels.

[@MilesBellas]

Bycloud's Flux video mentions LORA training on distilled models such as as Schnell and Dev creates stability issues.

[@nholmes86]

Flux is just Flux...no workings around, no mid ground lol

[@UnlimitedGRenemy]

had the exact same questions myself

[@DanDanTheAiMan]

10000% agree

[@ChipRiggs]

💯

[@fontenbleau]

you used a bad lawyer. 1. if you write truth that art made by Ai=have no value, because everyone can copy it, no machine rights, can't be co-author with machine(USPTO require exactly mention every pixel edited by human in Ai works). 2. if you hide it's Ai - thats a fraud of customer and serious future troubles because see #1, no one need no-value=no property in fact. So commercialization impossible anyway without any licenses, it's a trap from any side of view, only tools there can be sold but not Ai works, esp they would be able to find out it's Ai in future much better or there will be global presumption of Ai-made for everything created after 2022 to not overload courts (you would need prove USPTO you drew this by hand also recorded on VHS tape(bye-bye any digital video evidence after Sora in every phone-only analogue required) to award copyright-which will be on-claim basis for very elite group globally of the few). I see Ai works will be culture of lower very poor classes for themselves, rich won't even touch this, costs basically nothing but also no need to educate or master any skill to make like ancient artists, almost like quantum-something&nothing.

[@23rix]

totally agree. everybody understand this, but this guy told that it's necessary. so.. what now. after 1.5, sdxl were not always the best, depends on use, in my projects, improving 3d architectural renderings , 1.5 is far ahead than the others, dindt checked kolors though yet... but it's starting to be annoing. to spend much time to test new stuff, constantly test new stuff.... and again. and then appeared, f%ck you, old stuff is better.

[@Dalin_B]

Holy... Moly...

[@breakl7010]

I feel ya pain, im a developer but new to the scene and this is a shit show