/bbwai/ - bbwai

Stable Diffusion, Loras, Chatbots

BBW-Chan is supported by simple text or static image ads from a-ads.com.
Please consider whitelisting us in your adblocker if you enjoy your stay!


Mode: Reply
Name
Subject
Message

Max message length: 9999

Files

Max file size: 10.00 MB

Max files: 6

Captcha
E-mail
Password

(used to delete files and postings)

Misc

READ THE *RULES* BEFORE YOU POST!

[ / / ]

(336.11 KB 512x640 00107-3099090911.png)
(405.76 KB 512x640 00028-1742617437.png)
(426.82 KB 512x512 00159-1379098755.png)
(310.79 KB 512x512 00044-2671204358.png)
(330.53 KB 512x512 00104-1012007894.png)
(300.20 KB 512x512 00077-388568398.png)
Blueberry/Fruit/Non-Standard AI Body Shapes/Colors Thread Anonymous 04/11/2023 (Tue) 06:19:50 No. 3169
So I'm surprised this hasn't been talked about here yet, as I know there used to be at least one thread in the alt section, I think? Anyway, let's talk about pushing these models to the limit. There are lots of models that are now capable of generating larger body types; but it feels like there hasn't been much progress on generating more unique things. I know HG Swells sorta pioneered the photoshop method of generating an image, coloring it blue in photoshop, and then img2img-ing it, which is how the four sorta realistic ones were generated. However, both of the first two were generated as is without photoshop or anything other than stable diffusion. Of course, generating non-standard body types like blobs or the like is very difficult, as the further you get from the 'norm' the harder it is to tag it and teach it to a model. So post your creations, but also, let's talk about how we got them; if only because it should be fun to see what might be needed to get the models to create things that aren't 'normal.'
>>20621 and yes i did bring in all 3 of the current Blueberry LORAs, Throwing shit at the wall *Sometimes* makes a diamond
(2.26 MB 1533x1533 05612-874450937 final.png)
(1.98 MB 1360x1360 05644-1453406601 final.png)
(1.28 MB 1133x1133 05635-259189397 final.png)
(1.74 MB 1328x1328 05675-2656344566.png)
(2.31 MB 1559x1113 05700-2842861317 final.png)
(2.31 MB 1559x1113 05704-503412498 final.png)
The background is scuffed and I wasted a lot of time trying to inpaint in extras before resorting to an image composite, but I'm pretty proud of this one
(1.35 MB 960x960 05690-519287807.png)
>>21011 Here's just the berry
>>21010 >>21011 These are amazing! Whatever you're doing to make these is *muah* chef's kiss. Any thoughts on different art styles, or is that not an easy task?
>>21011 Abs unreal outputs here, nicely done Orbs are perfect
>>21013 Art styles don't have much to do with it. I've been following the advice of TheGuyWhoDidAThing from a few other threads in regards to generating realism, and his guide can be found on his DeviantArt page. The tl;dr is to generate the initial shape and proportions with a more drawn/cartoonish model, then switch to a photorealism checkpoint to iteratively upscale the image, removing superfluous tags and LoRAs as you go. The only reason I've been able to generate these shapes at all is with the new v3.0 version of the blueberry LoRA on CivitAi. Every other model generates mid-inflation shapes 99.9% of the time, but this one pumps out perfect fucking orbs as long as you're using the hyper tag. Hats off to Darkforces for training it so well. Godmode. I would also endorse Eddo's advice to incorporate several blueberry LoRAs for the first generation, but I would urge you to stop using the v1.2 of Darkforces model after that since it tends to force 2D art style and produce a ton of artifacts.
>>21024 Ayy, ty for the tip on animated/realistic models for upscaling, with give it a try later and see how we go, cheers
>>21010 >>21011 >>21012 Your results are fantastic! Amazing work. >>21024 Interesting. It must be the way I've been prompting but so far I ONLY get good results from version 1.2. The keywords I use are full and spherical blueberry inflation. It is definitely true though that the results get a lot of artifacts and a drawn look when you add any sort of background. Maybe I need to take another crack at version 3. I do like partial blueberries though and the 1.2 mixes well with the Hyperfusion LoRA. Examples attached here. The rounder berries are done with DreamShaper 8 and the more realistic images are RealCartoon-Realistic v13. >iteratively upscale the image So img2img? Anyway, thanks so much for sharing your work and your insights.
(884.44 KB 2904x1808 abandoned-factory.jpg)
(807.38 KB 2160x2304 tutorial-bb.jpg)
You guys were right about mixing the LoRAs. I added notes on what I did with this one, all the important parts are there. Basically I put the LoRA keywords at the beginning and include the LoRAs themselves at the end. The middle of the prompt is all specific to the particular image. To me it seems like the Hyperfusion LoRA does add a touch of realism but I don't use it above 0.2 or 0.3. I find if you use it with (round breasts) or round_breasts you can get better-shaped boobs. So thank you to >>21024 for recommending v3 and mixing LoRAs. I'd just about given up on v3 but I'd been using it by itself.
(835.21 KB 832x832 00029-1509644321.png)
>>21142 I'm trying these out and getting nice stuff back, Cheers for the tips! will keep experimenting in inpaint
>>21192 Beautiful! So glad it was helpful. It took a long time to figure out. Something I forgot to mention was that I read somewhere the BB LoRAs were trained at 768x768 pixels, so you get much better results if you use something like that.
>>21142 I'll be honest, I've not gotten ANYTHING that looks remotely good with the spherical shapes. I've tried doing the mixing of loras, but in general if you want any kind of specific thing your tags are going to be over 150, and there be monsters once you get that that point. Forget about using real style models, I've never gotten much of anything to work with drawn ones and not for lack of trying. Perhaps this is a failing when it comes to my mentality; the way I see it, if the lora can't output what it needs on its own, it's not that great. And having talked to the guy who made it, he's not particularly interested in the concept, which is sad. Had I any skill in tagging or training, I would train it on my own system; I've got the hardware for it. Though I think that in the future it probably would be best to have a lora specifically for spherical shapes; how you could go about training that, I have no idea, since you get pretty far outside the realm of what models understand at that point. The way I see it, and this might just be me, blueberries are kind of like the last frontier on trying to really get models to generate something because it's so far outside the norm on so many levels. You have an intersecting of skin color, body shape, and clothing, and there's no real world examples to train off of. So you're dealing with a very small sample size of very small amount of artists. Which means that actually getting it to work is very hard. One last bit. I'm not going to say that TheGuyWhoDidAThing's method is wrong, but it's needlessly long and convuluted for like, 90% of generations. Rather than trying to upscale over and over, just generate at that size and inpaint, and use photoshop to correct as needed. I've generated lots of things at 'reasonable' shapes that would fit into hyper that aren't bad. if you CAN generate at say, 1k square, then just do that and inpaint at that. You don't really need to upscale so many times outside of automatic's repo. I'm also not 100% sold on his idea that prompt doesn't matter. Imo, prompt is extremely important, especially negative prompt. In fact, my methods usually revolve around generating, seeing what the model seems to struggle with, and then using the negative prompt to nudge it in the right direction. Ergo, if you can't get it to move in the right direction with 'lying on back' then using things like 'on front' in the negative solves the problem usually. But I digress. I really think we need to just find someone who can train a proper blueberry lora so we don't need to try and jerry rig one together like this.
>>21531 Unrelated on the LORA point, but I've been slowly collecting screenshots from TMC videos in order to train a realistic sphere LORA, but god it is just coming out as jackson pollocks, will post results when they arent incomprehensible
>>21548 Okay so quick advice on your end. I would not try to use TMC stuff as your base for spherical. I have tried myself to train a spherical lora, and my base was using mostly things from artists like Sidkid, sickstar, ect. There are a lot of artists that work with the spherical shape, and you could in fact use the real life stuff, but I don't know that it'll give you a good base?
I recommend trying out the realistic models like Majicmix v7 and Epic Realism. Tougher to get a good result, butt the good gens are really worth it.
>>21745 I am eternally impressed on the shapes you're able to get. Spherical is definitely very hard to get where it doesn't look terrible.
>>21752 My humble thanks. Something I should mention is that I find the frame size matters a lot when trying to get spheres. 768x768 seems to make slightly tall berry, so using something like 800w by 768h tends to be the best bet for a sphere. Also, like I always say, it is a numbers game. The sphere you see above was 1 out of maybe 80. There were a few others in there that looked decent but this was the best one.
These ones are all Animerge 2.7 Been thinking about publishing techniques and even prompts (or most of prompts, the important bits like before). The only snags are: (1) it would be disappointing if, after spending time writing and publishing, nobody was willing to do all the reading and put the hours in to create images (2) I've realized that at least 50% of the work is inpainting and Photoshop. I can give guidance about inpainting but Photoshop is way beyond the scope of what I can spend time teaching. It's not something I could do for weeks but I thought I would mention it in case there actually are people who would read it.
>>22222 Checked.
(791.07 KB 840x784 00012-616383122.png)
(708.07 KB 832x784 00088-653418860.png)
(825.20 KB 912x784 00009-4197498116.png)
(716.53 KB 832x784 00017-1446906960.png)
going to post some random ones from the last few days
(762.76 KB 840x784 00041-2511497771.png)
(866.56 KB 840x784 00049-1514921731.png)
(2.67 MB 1680x1568 00135-2000871843.png)
(826.16 KB 840x784 00044-1967985461.png)
(2.98 MB 1936x1472 00003-2546830168.png)
(916.66 KB 888x848 00047-1698083762.png)
(2.59 MB 1696x1584 00057-2994766390.png)
(720.83 KB 816x784 00054-1837930798.png)
(759.14 KB 816x784 00038-588168166.png)
(728.25 KB 816x784 00045-1974570961.png)
(789.12 KB 888x784 00059-816391532.png)
(2.98 MB 1664x1568 00138-737250404.png)
(730.35 KB 832x784 00016-1364601735.png)
(2.87 MB 1664x1568 00063-276770521.png)
(3.14 MB 1664x1536 00018-1448272087.png)
(2.86 MB 1648x1568 00079-3728884975.png)
(1.38 MB 1648x1568 00001-1681527512.png)
(739.76 KB 832x912 00106-910517935.png)
(878.05 KB 832x912 00144-4095521395.png)
(804.75 KB 832x912 00011-1891920879.png)
Bump
Good
>>22825 ????
>>22843 My friend are you okay, ramblings of a mad man
>>22844 Bbwchan welcomes pedophiles, rapist, killers, human trafficers. Are you ok????????
>>22844 "Dont expect professionalism here" - nerd look like mad scientist lol
Take your meds son
>>22847 Youre on bbwchan. Do yourself the favor old man.
They dont make medcine for what I got.... (Autism)
>>22849 Do autistic people have sex? How can I be autustic when I fucked more women then all of you lol
>>22850 Just my also autistic wife :) Honestly it helps, hyperfixation goes wild
>>22854 Oh that too like, whats your point big dawg?
>>22854 >>22857 Old man. Look me in the eyes I dont fear no one. Fuck u going do.
(4.80 MB 2048x2048 00050.png)
Had some success with training a LORA (pic related, usual SD problems with hands/feet/face inpainting required). Moving to SDXL seems to have made a huge difference. If anyone here has tips for pushing training into the degenerate unknown, we should share them. Under SD1.5 it really felt like I was fighting the base models with how they viewed human anatomy. Also for those generating locally, are there any tweaks for SDXL stability? Automatic1111 is crashing on me once every 4-6 generations.
>>23738 I recommend using A1111 Forge if you aren't already. It's a branch of A1111 with a bunch of optimisations.
(2.47 MB 1536x1536 00021-2094892416.png)
(2.58 MB 1568x1280 00107-30208007.png)
>>23768 Thanks for the tip, I switched over to forge and its much more stable. I really appreciate that force is smart enough to scale VRAM allocation dynamically based on image size instead of crashing on OOM errors.
(2.15 MB 1568x1280 00113-40417150.png)
(2.14 MB 1568x1280 00085-972812512.png)
Some more success I have had. Anyone experimented with regularization images for LORA training? I feel like the LORAs I have trained tend to pull in the backgrounds far more than I want.
>>24125 I'm very curious what your training settings were. You should also upload your lora to civitai so we can test it, because it seems like you're getting fairly good results. In general, the way you avoid getting backgrounds to be trained into the images is to make sure you tag for them and also make sure that you have a wide variety of them, so it understands that the backgrounds aren't what you're training.
(2.33 MB 2000x2000 08240-555519790 final.png)
(922.27 KB 2000x1367 office gossip final.png)
(1.42 MB 1024x1024 07489-3388350134.png)
(1.21 MB 1024x1024 07486-3047726762.png)
(1.50 MB 1280x1280 08139-3653228954.png)
It's been ages but here's a dump of some of the stuff I've been making lately. I've got more lying around in some folder that I still need to touch up, so I might post those later.
(2.91 MB 2000x2000 08224-2733944198 final.png)
(4.32 MB 2500x2500 07726-2571960231 final.png)
(1.50 MB 1280x1280 08211-3009183112.png)
(1.60 MB 1280x1280 08058-3646711441.png)
(3.74 MB 2000x2000 07689-3122057166 final.png)
(1.45 MB 1280x1280 08097-895082607.png)
(1.17 MB 1280x1280 08063-3801092021.png)
(1.44 MB 1280x1280 08085-3099642456.png)
(1.51 MB 1280x1280 08075-3927030143.png)
>>28244 These are all good, but the realistic looking ones are incredible!

Delete
Report