Fight scenes and Complex INTERACTIONS with MULTIPLE characters
Vložit
- čas přidán 31. 05. 2024
- Create interactions between multiple trained characters in text to image. Multiple characters with regional prompter/latent couple + composable lora.
Including creating fight scenes for your coming creation, and improving your Story Telling with Stable Diffusion!
Join our discrod server: / discord to learn about this and more!
[ _____lora_____ #cmd\(cooldown\(0.5\)\)]
------------ Links used in the VIDEO ---------
Videos you should check out:
-After Detailer by @KeyboardAlchemist: • ADetailer in A1111: Ho...
+Latent Liminality Channel: / @latentliminality
-Composable LORA: • [Stable Diffusion AI畫圖...
-Fix Hands: • Ultimate Guide to HAND...
-Create Easy Controlnet References: • Best Tools and extensi...
Extensions used in the video:
Composable LORA: github.com/ashen-sensored/sta...
Latent Couple: github.com/ashen-sensored/sta...
Regional Prompter: github.com/hako-mikan/sd-webu...
Photopea Embed: github.com/yankooliveira/sd-w...
After Detailer: github.com/Bing-su/adetailer
Controlnet: github.com/Mikubill/sd-webui-...
------------ Useful stuff for this ---------
Keep in mind that some checkpoints might have a better time than others when using things like Regional Prompter or Latent Couple!
Recommended are:
ANYLORA by Lykon for anime generations
SDXL or Deliberate for others
------------ Social Media ---------
-Instagram: / not4talent_ai
-Twitter: / not4talent
Make sure to subscribe if you want to learn about AI and grow with the community as we surf the AI wave :3
#aiairt #digitalart #automatic1111#stablediffusion #ai #free #tutorial #betterart #goodimages #sd #digitalart #artificialintelligence #latentcouple #couple #composableLora #posing #controlnet #SD15 #inpainting #openpose #depthlibrary #AI #midjourney #interaction #relation #comic #storytelling #manga #anime
0:00 intro
0:12 Requirements
0:28 Installing extensions
0:40 Issue at hand
1:18 Start
1:35 LC vs RP
2:25 Main method
5:05 Composable LORA
7:25 CL results
7:40 RP workaround
8:20 After Detailer
9:00 Big thanks!
9:20 Upscaling
9:55 Base Method done
10:20 Levels of difficulty
10:55 Compositional
11:45 Clean up
12:45 Divided Interaction
13:33 Fixing hands
13:45 Overlapping
14:45 Using DEPTH
15:25 Inpainting
16:25 Thanks for watching
![[柴犬ASMR]曼玉Manyu&小白Bai 毛发护理Spa asmr](http://i.ytimg.com/vi/0TsXQ7z2Dh4/mqdefault.jpg)
Wow, TIL, another amazing video! It is crazy how you are able to make perfect use of all these extensions and generation techniques to bring the perfect composition to life! My hats off to you, good sir! And thank you very much for the shout out. I appreciate it!
Thanks!!! Would have gotten stuck if it werent for your great Adetailer video. So ty for that too 😄
This is where one moves from regular AI enthusiast to artist
sure hope so!
When people ask me if AI is going to take away artist's job. I'd show them this video
Hahhaha kind of a shortsighted take, but yeah atm it wont. Artists are way better than AI rn
always bringing the good stuff!!! thank you my dude
Hahaha ty for watching!
Thanks for the shout out! Nice to see how the video turned out in the end!
Thanks!!!!
This was exactly what I needed to watch, because yesterday was struggling SO much with a composition of 4 characters and I couldn't make it work (using CNet openpose + ADetailer + Regional Prompter, all I was getting was pure noise no matter what)
Yeah, as you add more characters AI goes crazy 😂😂😂
Extremely useful stuff. I did not know about latent couple extension, will try it out today!
You video addresses one of the pain points of AI generation - it's quite easy to get a simple image (like a portrait) but anything dynamic (movement, emotion) is a pain the ass.
Great job mate, many thanks!
Thank you for the kind feedback! Hope it helps and, if anything, ask on discord!
Dude I just found your channel and wow, very helpful, very informative. On my way to try this out now! Thanks
Happy to hear that!!! Hope it helps. If you have any question Ill try to aid on discord (Ill go to sleep first tho 😂)
Thanks btw!
Brother back with a banger, thankyou!
Thank YOU!!
the goat is back, great vid! I think as a community we should continue to develop a pack of poses for different genres/scenes
Thanks!! And totally, I think that's a great fking idea ngl
Amazing Work.
I'm excited to try it.
Thx.
Thank you!! Hope it helpss
Brother, incredible video. Thank you!
Ty so much!
that amount of example you made is incredible !
latent couple seems to be back and working so i'll consider that over reginal prompter now.
for lora control in regionnal prompter i use the extension "LORActl" extension, it have the same dynamic role as composable lora with this typing format :
(= 100% weight at 10% of the steps / @ / 0% weight at 100% of the steps)
i also know that you can paint your prompt area in regionnal prompter but honestly overall LC seems easier to use.
Thanks!!!
Had no idea about loractl, should have checked it out. Id pin the comment but I reserved the spot for keyboardalchemist.
And yeah, to me LC is a littlr simpler to use for this, but for some reason it gives people errors from timw to time
Thank you !!!
ty for watching! Hope it helps :3
The alternative to this is rendering separate for each character and composing them in photoshop, but having a workflow like this means in the long run you can save a lot of time. especially if you are working on something like a comic
Yess! Some times composing then will be necesary but it is what it is 😂😂 at leas I hope this saves some time. Mainly if you have a nicer pc than mine, cut this generations took like 5 min each hahhaha
@@Not4Talent_AI what generation GPU? is it 3000 series?
nope GTX 1080 TI@@highcommander2007
You are the best!!
thank you so much! hope it helps!!
@@Not4Talent_AI helps! It was a life saver, thank you for making these amazing videos!
happy to hear that. Thank you for the kind words!! @@gauravranjan7593
dude definitely deserves
praises
thank you so much!
Thx for the help nice video
no problem, thank you!! Hope you get the results you were looking for. dont mind asking again if you need to
Great video
thank you!!!
Thanks!
Wooow, thank YOU!!! hope it helped!
Im so happy I came to your channel! Your really charismatic and do real good infomative content!
hi!! Thank you so much for the kind words!!
I have a video about inpainting in the channel, even though it is probably a little bit too fast paced. It touches pretty much everything there is about inpainting tho.
I'll eventually come back to it. And it is likely that in the next vid I'm planning I'll also talk about it a littlebit, not in too much detail tho.
czcams.com/video/hDfJajYxOc4/video.htmlsi=CVqYH1AaNsV5o-sV
@@Not4Talent_AI Thank you so much c:
I look at it ASAP :>
And I bet it wont be too fast paced ;p
hahahaha I sure hope it isnt xD @@michellebecker4773
Neophyte here but into 1 minute of video, the hat and color issue is an easy prompt syntax fix not extensions needed. Was able to set a t-shirt for man and an apron for woman which without prompter prompt both would either wear a t-shor or apron or fused version of the two. But yes simple prompt fix.
Again, yep, its possible to fix with prompting when it comes to simple things like that. But it wont always be the case.
[SPOT THE LORA CHARACTER] I spat my drink... Great tutorial, as usual.
Hahahaaa tyty!
Amazing job! I wonder if making something similar on comfy would be complicated. I know it has nodes for that sort of thing but I never really managed to make them work properly.
Thank you!!
Tbh Im not sure. Ill probably find out when I start using comfyui (no idea when, but I probably will eventually 😂)
I’m thinking how well regional prompter will work with a ControlNet Lineart base instead of a pose, without LoRA. Have to experiment more with interactions, I usually just generate the characters separately, and then photobash them, clean and give them a 0.35 tile upscaling pass for cleaning, but I haven’t tried that with direct contact interaction.
The colors will probably still get fked. But It will surely help with the over all shape if you have a matching depth map
@@Not4Talent_AI Well, colors always get messy with CN Lineart unless your prompt is very good or you provide additional guidance. For my use case, I use img2img with a flat colored version of the lineart sketch for aiding with that. Im my tests (done with Mistoon Anime v2 and Meina, other models could have different thresholds), At 0.5 denoise, colors tend to be preserved but the details added are meh. At 0.7 the colors are only preserved in general, but the details are better. Anything beyond 0.7, and SD starts to ignore the colors completely.
If you have the gidance ready than it's really nice in that way yeah. Probably better and easier too haahaha
@@Nik.leonard
Will check your video later. Hope I can finally get some good results with my compositions.
Also, I don't know why, I can't get Latent Couple to work properly, most of the time it won't work at all because of weird errors I can't really understand to find the problem. Whilst Regional Prompter works pretty fine once you understood how it works. Although it also has some problems sometimes.
Nice!
Unlucky. In any case you can 100% use regional prompter. It also has a paninted mask option so It would be using that for complex poses.
Wish you luck in your generations!
Very nice video !! Thx !
I have a question : how do you get/generate openpose skeleton and depthmap images ?
Thank you!!
I usually use posemy.art for it
Amazing vid! I'm inspired to start experimenting.
Say you want to train a style LoRA on images that are compositionally complex (e.g., fight scenes with overlapping limbs). I can only source complex compositions from this style and the models I train come out with extra, missing, or misplaced limbs. Worse than normal.
Is it impossible to style train on this complex image set? I've nearly given up entirely.
I'd ask in the discord you mentioned, but I can't find an invite.
Thanks!!!
I'm not entirely sure since it isnt something I tried. But I get where you are coming from. Idk if the idea you have is to train only the style or the style + complex positions. but if you are only going for style, maybe crop the images in a way that arent as complex? if possible ofc.
Link to the disc should be in the description but, in any case here you go:
discord.gg/SXNncPtgT5
good
The Goat! behehe 🐐🐐
Hahahahha ty!!!😂😂😂
Yeah, it's also often bugging me, that even if you combine Depth and Open Pose, the AI still cant differentiate front from back. This had me create so so many weird images.
Hahahhahaha I feel ya😂 the ammount of cursed stuff I made while experimenting is pretty high hahahhaah
Love your video as always :)
Just wanted to add, that you made a tiny mistake in the video description. The link for Regional Prompter is calling Photopea instead. :/ Just so you know :) Probably a copy paste mistake xD Happens to the best of us
oh, sht, I'll change it up rn, thank you so much for pointing it out and for the compliment!! tytyty (edited)
@@Not4Talent_AI Glad to help c:
I have none of that you listed... also my computer sucks so installing sd is a problem also with small amount or ram don't have money for a new comp... im using a service of SD and its working out... not sure what other options are there...
Hmmmm in that case, I hope you can at least use a service that lets you instal CN and either LC or Regional prompter. (Or that have them installed already).
You can try to go for a 2 step approach. Where you dont use LORA in text to image but instead just describe the character you will want.
This will hopefully give you a close base. Then just put that image to inpainting and inpaint 1 by 1 with the corresponding loras.
Hopefully you get it to work!
Thanks for the video. What software did you use to generate the depth maps? When I export a depth map in PoseMyArt with two characters the front most character covers the back character. The back character doesn't show through like in your depth masks.
Happy to help!
I use pose my art, but you need to play around with the blue slider under "toggle depth map". If you toggle depth map you'll be able to see what that slider does. And adjust the map in a way that both characters are visible and have good contrast.
I can't be the only person that sees Brandon Routh.
ahhahhahahhaha actually flattering ngl xD
i don't understand why your latent couple jsut so different to mine, I cannot draw in my latent couple
did you get the one in the description? there are 2 different latent couples. one is super old and doesnt have a mask capability
Wow!
ty!!!
Great tutorial! But I still can't figure out how to use same LoRA in Latent Couple (or Regional Prompter) without overweighting. And placing LoRA at base prompt gives underweighted result.
hi!! tyty, I'm unsure of what you mean, could you give me an example ?
So I’m having some trouble with latent couple. As a basic example of the problem, let’s say I mask 2 different zones on the canvas, right? That’s 3 zones total: 1 for the background, and 1 for each of the 2 masks. Now let’s say I want a dog in one masked zone and a cat in the other. Well what happens is instead of generating a cat and a dog, it fuses them together into 1 amalgamation. The same thing happens with 2 different Lora. If I do a lora for 2 different characters for instance, it tries to combine them into 1 singular character. On very rare occasions it will actually do 2 of the same character. Does anyone know why this issue is happening and how to fix it?
hi! That shouldnt be happening. Please make sure that the latent couple extension is activated and check the cmd for more info in case it is but isnt working. Could also be that you have the "separation sliders" too low. there is a preview to see if the masks are clearly separated.
I can't seem to get latent to work. It also looks like it is missing features such a creating and uploading masks
weird. You could also try to use "regional prompter" then. It has pretty much the same stuff
I'm not getting any of the commands, like this [#cmd\(cooldown\(0.5\)\)] to work. I always get a flat horizontal line at LoRA weight 0.00 or 1.00 through all the steps on the graph. Changing between # and \u0023 doesn't help
Do you have the "dynamic prompts extension" installed?
If not, send me please the prompt (shown in the metadata of the image, not the one you have typed out). That way we can see if it is an extension issue
@@Not4Talent_AI Just did a simple test with just Composable Lora (Enabled checked, Composable LoRA with step checked and Plot LoRA weight in all steps checked). Again it gave LoRA weight 1.00 through all 20 steps. This is the meta: katheryn winnick female elf, [#cmd\(cooldown\(0.5\)\)]
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 123212613, Size: 1344x768, Model hash: 31e35c80fc, Model: sd_xl_base_1.0, Denoising strength: 0.7, Hires upscale: 1, Hires steps: 15, Hires upscaler: Latent, Lora hashes: "testkira-000001: 71be084720ad", Refiner: sd_xl_refiner_1.0 [7440042bbd], Refiner switch at: 0.8, Version: v1.6.1
What are you using to create your 3D poses these days that you put into Controlnet later?
posemy.art
Thanks mate. Was wondering if you used something else nowadays. Also I would like to make a suggestion. One thing I am still struggling with a lot are very specific camera angles. For example a shot from the feet upwards but midair, or a shot straight down above the head. Even with Poses and depths and canneys SD seems to refuse to do these extreme angles. Maybe as the god of SD you know how to do that? haha @@Not4Talent_AI
Nope, still using the same for now.
And the camera angles thing is on my list of stuff to test out and find a way. Atm I dont really know how to get that consistently.
And thank you hahhahahhaha but I'm no god at all tbh xD I just experiment until I get something good enough hahahhaa@@S4SA93
Love to hear it man. Don't be so humble, you're insane! Much love
When I tried regional prompter I found it ruins the models ' style more than addin 6 loras at 1... I suspect the size of the regions might be too small for the model, but I couldn't find a good combination of regions and image size :-/
hmmm That's weird. I don't think that should happen tbh. Maybe it was a model that didnt support the "AND" keyword?
@@Not4Talent_AI I can try that tomorrow, thanks for the suggestion 😊
your results are amazing but after a lot of trying I am just not able to get what I want but awesome tutorial
Thanks! What issues are you encountering atm?
sooo, I need to figure out how and which lora of mine to input here.
[#cmd\(cooldown\(0.5\)\)]
yep. Just click one lora and change it for "". Then just play arround with the numbers a bit.
Like: [#cmd\(cooldown\(0.45\)\)]
Do you know a method for two character Loras working in Forge?
hmmm dont use forge so I dont really know. I guess it'd be the same but with extensions that run properly on there
@@Not4Talent_AI thanks, no. I tried every Extension you and others talked about. Forge is awesome other than that. Way faster
@@NicoSeymore Yeah, it's sad that it's not a 1 to 1 mimic of a1111. If everything that works in a1111 worked there it'd be absolutely incredible
I cannot put mask or my latent couple just doesn't have mask funciton.....
if it continues to fail, you could always just use regional prompter
i use jaggernaut xl, I tried the balls example but never managed to get 3 different balls. Where I make a mistake, I leave everything as default.
I think with sdxl it should work just fine tbh, if you are on discord, an image of the ui would help tbh. I'll try myself too and see what it could be
@@Not4Talent_AI i have discord yu want see?
sure!
@@artificialbeat
Results are good but I feel like someone will make a better program one day that my grandma could even use ( she thinks anime girls are dolls when she sees them ). Its getting so complicated that it reminds me of the bad times I had with Blender. I remember your other video where you said its kind of hard to make money with AI so I wonder is it even worth to become an expert in using all this :/ I feel like it must be like a full time job to keep at the cutting edge of it.
Ofc. I hope it becomes easier one day 😂😂 atm I just try to find solutions for the current state of AI, since there is a lot of people who cant draw but have stories they want to tell.
Hopefully in a few months it becomes way easier, but for now I havent found a better way.
And as far as expertese goes I couldnt tell you. If you just want to learn it to make money, then probably not. For that you should learn how to train models and checkpoints for other people. Thats where money is atm.
Is latent couple broke? Is there something that can replace it?
it isnt for me, but you can also use regional prompter in case lc isnt working
@@Not4Talent_AI Thanks. Believe it or not, I feel better. At least I know it's not me. Um, maybe do a follow on Regional Prompter as a replacement then? And yes, I subscribed.
@@rwarren58 hahaha tyty!
It's pretty much the same thing using regional prompter. If you feel like you need more info. In the discord we have a FAQ section where Lefourbe made a detailed regional prompter guide on masking and using it for this purpose
Does this work with SDXL?
yes, mostly. Even though parameters might be different as sdxl controlnet models work a bit worse
8:36 holy shit thats huge
Right?? You can do that with the full body too. Im so glad I saw Keyboard alchemist's video on it 😂😂 totally worth a watch
I prefer just sketching random characters then inpainting each one of them with loras , no need to deal with these mega complex addons and results are all the same
What about the interactions?
Thats 100% valid yeah. The more skilled an artist you are, the easier it will be ofc.
Looking forward to Nightshade.
Hahahahhahaaa hope it goes well for what it needs to. Sounds like an interesting tool to avoid illegal training
It won't really do much. Especially since you can use your own references and remove stuff put in the images.
Bingo: fully body , Street photography, man in a blue suit hat on man and woman wearing red full length slip dress no coat bare shoulders no hat on woman, walking through a ((busy street)), 4k, daylight, wallpapper. - so no extension or Loras .. girl in red dress no hat boy in blue outfit with hat. Did loose crowd but able to bring that back. LOL able to change to baseball hat too.
Hahaha yeah, and using keys like BREAK and such it will be easier. But the point of the example is that even simple things are hard to get with prompting. And more complex scenes will just be impossible to get consistently.
Latent couple isnt working anymore if im not wrong
Really? Whats the error?
@@Not4Talent_AI I will join the discord to share the error that I'm facing. Basically after i mask the image & hit I have done masking, it's generating blank sheet
@@erzascarlet2577 wtf thats weird. ok, see you on disc then!
I've been learning stable diffusion for the past 5 months and what I've learned is that A LOT of things can be done through sheer prompting wizardry and a tiny bit of luck/inpaint/ControlNET. Let's be honest, those extensions are downright USELESS.
Inpaint can do pretty much anything. You can move from a rough cn img + inpainting to pretty much anything.
But that will take you way more time than using the extensions.
Now, every day SDXL gets better. When I made this video what Im doing here was harder than it is atm. Calling them useless is a bit much, since I bet you wont acomplish half of this with SD 1.5 raw without a lot of inpainting and luck.
As you just CANT use 2 loras like this without them interfering.
@@Not4Talent_AI Things aren't as simple though. I was also initially very hyped for SDXL until I tried it. Let's be honest, there is NOTHING in SDXL you can't do in SD 1.5 as of right now. SDXL merely means highres basedata the model was trained on, nothing more and nothing less. Sure, they could have added some more tags/data and trained it more thoroughly (hence why some models even in SD 1.5 do respond better to certain prompts and produce better results than others since they have that specific source data trained and anime models happen to be in 90% of cases just plain better than regular/realistic models since, well, more tags/specific data) but not much. You can get more detailed images in SDXL...if you go ABOVE FHD resolutions and that means...yeah, slow adoption, slow developement (since it's very hard to pick decent training data at FHD for source images) and practically no difference when it comes to real usage.
P.S Lora containing several characters is a thing, btw, and works flawlessly even without latent couple. I've had MUCH better luck just straight up writing 'in the bottom-right corner, on the very left of the image, character's cap is x' and it just WORKS. It didn't a year ago but now it does so why not use it instead? And merging similar style loras is also an option.
Ngl Id love to see images done like that in text to image. Cuz I have tried this all before and it didnt work. Starting by the fact that SD 1.5 doesnt know left from right.
I dont know everything tho, so if you found a way to do it, that would be very interesting to see.
If you can send me a prompt, model, and loras to use. I could see if I can replicate what you are saying. And if it is replicable, then that changes a lot.
Not only for me but for the whole community.
In case you can, ty for sharing!
@@Not4Talent_AI model: CounterfeitV3 (just can't seem to find a more universal and well-trained model), I use mainly basic common-sense phrases like (on the left/on the right), (on top of image/in bottom of image). It's far from 100% accuracy, but the results are pretty much comparable to Latent Couple/Regional Prompting without masking. Skipping the articles (a/the) does save some tokens so I believe it's better to skip them whenever possible. SD will still in 90% of cases understand the tag that is being referenced.
That aside, I think it mostly comes down to using as little tokens (prompts) as possible so that it prioritizes every word equally. I think many people noticed, the longer your overall prompt gets, the more you start relying on (token:value) to outweigh the first 75 tokens. So, the most important aspects of your composition should always be in the first 75 tokens. Unless it's a token (not lora /specific prompts/ that affect its further activation), those don't really affect token priority.
So, my main prompts are always under 75 (ideally 50-60) tokens, I do not use loras that require the addition of many specific lora-related token (they just mess with your overall image and often times limit your initial prompts by chaining them to that lora specific training dataset, I use 'flexible loras' only where you don't even need to add any specific additional tags other than basic to make it work), I use some prompt 'wizardry'.
Every model has 'face close up' dataset. For some reason, it has insane priority over most other tags. Type it among the very first prompts, do not remove it, and try adding anything else related to human, perspective, clothing, background. The results are very interesting indeed. And while you may change perspective (view from below/above/behind/far away), clothes, background, add more characters, 'face close up' will keep affecting the face of a character regardless of how far away they are and keep drawing them in 'manhua' style as if it's still the only thing in focus screen, although adjusting its depth/perspective appropriately. It works up until...you guessed it, 75 tokens.
Beyond 75 you start fighting for your proper prompt recognition with Sampling Method, CFG, VAEs and (prompt:value).
Maybe there are some other 'tags' that have priority over others, that's what I'm trying to find out right now.
Ill try when I get home, even though Id love a tested promt or generation data.
Ill try creating any of the images above without the extensions, just using what you said.
So, trying to get a character punch another one. I also try to keep promps under 75 tokes, even tho you can controll this without it affecting your image as much using BREAK.
In any case Ill tell you if I agree that the extensions are useless or not after the fact.
If you already have an image prepared with gen data it would be great tho
i use masked lora and it fixes all of theses issues.,,, specially since composable lora is deprecatd
Ive never even heard about masked lora, and I checked the extensions not so long ago. Is it good?
@@Not4Talent_AI it does what latent copue and segment anything feel like should do. when you install it and expand it, there is a video that edxplains everything it does. it is the perfect solution to mix multiple loras
amma check it out, ty!
@@CarlosFerrabone
cant find it on the available tab. Do you have a link?
@@Not4Talent_AIits not allowinfg me to past a linj here, ill send you a direct message
if you can draw a decent drawing you can use just imgtoimg, controlnet and inpainting to make 2 character , but once again drawing decent image also need talent
Yep 😂😂 knowing how to draw will always be the best option. But also the hardest to be able to do
latent couple doesn't work anymore
what? why? It does for me
@@Not4Talent_AI I can't seem to download it. I tried the default method and installing from using a link
weird. If that keeps happening, you can use Regional prompter instead. In our discord there is an explanation on how to use it like Latent couple. by Lefourbe @@obsidian3136
Dude, you could have fucking said I needed composable lora sooner. That requires a restart which wipes everything I just spent time on writing.
Its literally in the first 15 seconds of the video man.
And also, if you hit generate with what you wrote. After restart you can just either click the blue tick icon under generate to load all your last parameters. Or just import the generated image to load them
Also SDXL 1.0 is also now part of SD 1.5 have test several SDLX lora and checkpoints and they work.
I dont think thats the case, could be wrong though. Do you have any examples I could test?
2 different software and super sped up tuitional, whut
I just used 1 software. Well, unless you count photoshop that I used for like 30 secods
@@Not4Talent_AI i mean learning Composable LORA & Regional Prompter at the same time is abit confusing
ohhh I see. You dont need to learn both at all. Just stick to either Latent couple or regional prompter. (I recomend Latent couple, but both are perfectly fine to use on their own) @@hairy7653
you can always just draw it
If you know how to thats the best way 100%
Why learn to draw? It will become a useless talent soon, whether you like it or not. So accept it and stop crying
Disagree, learning how to draw is not only good for creating art. And want it or not a person that know how to draw will always get better results than someone who doesnt. They also have legal protection, which AI doesnt. And businesses tend to contract professionals and not just people with a skill that most people have.
An artist will be able to tell what is wrong in an image and how to fix it, wile someone who has only used AI most likely wont.
Ofc as the tech advances the line will get blurrier and blurrier, but I still think that, in this lifetime, learning to draw will always be a valuable skill to have.@@fernandosouuza3881
>:c
i can draw ok, but not as anatomically correct or detailed as i'd like, when i started comin up with characters and story i was doodling and paying artist to redraw. i would rather use whatever resources i can to bring my vision to life and not spend a large amount of money to do it if possible. if i could i would though
Great video!
I'm not getting any of the commands to control the weight of composable LoRa during different steps. I always get a flat horizontal line at LoRA weight of 1.00. Chaging between # and \u0023 doesn't help.
Although I get "composable LoRa load successful" I only get a horizontal line. I tried all possible syntax and functions. please help!
this is the simplest command I'm trying [::10]
ty!
Thats very weird btw, not sure what it could be.
Are you using sdxl or sd1.5? It should work for both I think but... idk maybe it has something to do with that, or the version you are using?
sadly im just getting errors when i try using latent couple .
RuntimeError: tensor.device().type() == at::DeviceType::PrivateUse1 INTERNAL ASSERT FAILED at "D:\\a\\_work\\1\\s\\pytorch-directml-plugin\\torch_directml\\csrc\\dml\\DMLTensor.cpp":31, please report a bug to PyTorch. unbox expects Dml at::Tensor as inputs
no idea what that means
hmmm that's a pretty weird error yeah., Is your LC a new instal?? If not, you could try erasing it from the extensions folder and re-installing. Else, try just using regional prompter instead. You can achieve pretty much the same results. Even if the prompting is a little different.
Someone also recommended "masked lora" extension, but I havent tried it yet.