>release month and no exact date Got this ready for the delay announcement.

Testing JANKU v5 checkpoint. I'd been using v4 for a long time. JANKU v5 [Checkpoint]: https://civitai.com/models/1277670?modelVersionId=1896532 Retro Anime Illustration [IL LoRA]: https://civitai.com/models/1227514 JANKU page says you can gen large images up to 1536x1536 (or 1024 x 1536) without problems so I tried that, then hi-res upscaled it anyways of course. The issue with doing that is the resulting image is too large to inpaint & takes a century to process, so I had to separate it into smaller sections, but I made things even worse by using a LoRA that introduces film grain which means if one part of the image is inpainted then every part must be inpainted to make the film grain match. I tried some other AI tools to remove grain (they suck & ruin the image), but the best way to even grain out was to simply SD Upscale it which comes with its own issues & can reduce or break details rather than improving them. A third layer of fuckery was using Regional Prompter. When inpainting a Regional Prompted image I have to switch between the two prompts depending on what I'm inpainting. SD Upscale can't properly separate anything with Regional Prompter either since I believe it stupidly applies the mask to every tile while ignoring the whole. In the end, I got it evened out enough to mask any remaining grain variations in Photoshop. Next time, I'm gonna do something simpler. I should just delete the LoRAs that introduce film grain since they make inpainting a hassle. /textwall >>1713128 I haven't even played though all the shit they added to the early access build over the years.

>>1712471 It's difficult, but WAN is supposedly better at prompt adherence. Made some attempts with Hunyuan, but got a lot of random outputs. My last resort, yet still unreliable, method was to make an edited version of the image with that specific clothing removed, then use that second image as a keyframe in Framepack. Had her hands been gripping the garment we intend to remove in the image it may have worked better, but best I could do was sliding them off via telekinesis as she wouldn't remove them herself. Also included one gen that had a wind blast suddenly blow her top up since it made me lol (it wasn't told to do that). Simple/known action are much more manageable, but when the AI doesn't know or decides not to do something it's a lot of random fishing for anything that works. LoRAs are often a necessity. When I can't get it to do something I try to bruteforce it by mentioning the same action repeatedly in the prompt hoping the AI will take the hint. Ironically, removing her top is significantly easier & happened multiple times when "undressing" was in the prompt. Hunyuan Video LoRAs 0.3 weight https://civitai.com/models/1255010/anime-style-for-hunyuan 1.1 weight https://civitai.com/models/1307852/innies 0.8 weight https://civitai.com/models/1319348/hunyuan-porn-salt Gave extra weight to the innie pussy LoRA, but it failed to reveal anything. Using a keyframe image was the only way.

https://civitai.com/models/1916227/il-gunfighter-ballads-and-trail-songs-cover-big-iron-stance?modelVersionId=2168823 Sometime this month is the 66th anniversary of Big Iron, so I made a LoRA out of parodies. That makes 8 months of daily LoRAs. Intended 6 then had lots of spillover due to on-going Jimmy Neutron project and made it to 7, then I had enough spillover from Pony conversions of that to get most of the way through 8 so I decided to push to finish it (that's why I've been updating some of my old stuff: Easy way to get something "new"). Just one V2 (which I don't think came out too well) tomorrow and that's totally it. Already got a few things WIP, so I won't be totally gone, but I absolutely won't be remotely as frequent.

>>1708823 You could make a matrix chat and maybe people will join it

I can't tell if my gens are getting worse of it my standards are just getting higher. Either way don't know how to make the art look less generic

>>1709903 >>1711080 >autistic shit like Regional Prompting, multiple characters, and such Here's a demonstration of Regional Prompter used alongside ControlNet. These tools do not necessarily need to be used together, however using them together tard-wrangles the AI more strictly. In my experience, this is the most reliable method to force two characters into specific poses with specific camera angle. As you can see, they're posed together without separation. This requires specific masking of the two characters & since ControlNet always places them in nearly exact positions, they can be separated by Regional Prompter reliably. Mistoline: https://civitai.com/models/441432 Mistoline goes here: stable-diffusion-webui-reforge\models\ControlNet 1st image is my end result after all edits, inpainting & upscaling. Hoodie stripes weren't coming out right so I added them manually in Photoshop, then inpainted my edits. Same for the eight-ribbon. The style LoRA I used: AkaBleak Style [IL LoRA] (DELETED): https://civitai.com/models/1374482 Archive: https://civitaiarchive.com/models/1374482?modelVersionId=1552991 >>1717540 If different art styles aren't enough, you could try Mistoline with pose images to get more interesting poses & camera angles.

>>1717581 Annoyingly, my prompt in the guide image has unnecessary junk & mistakes, but the general idea of how it all works is still in there.

>>1717581 >>1717792 >>1530177 Follow up per request.

>>1535107 https://kusa.pics/play/smash It seems to be the same, though it is been "fixed". I tried with both a lamborghini and a nissan GTR and it gave both of them a 1/10.

>haven't genned anything in a month I've grown bored of 1girl (or multiple girls) I think. >>1717581 More interested in where you got the pose and if I can make my own. I typically look for stock photos or porn images with the pose I want and quickly edit them in paint.NET to get what I want then do a depth pass.

>>1718275 gibe them baginas

>>1717540 I recommend familiarizing yourself with composition tags: https://danbooru.donmai.us/posts?tags=tag_group%3Aimage_composition >>1718300 >3 Cute.

>>1718300 >I typically look for stock photos or porn images with the pose I want Same. >where you got the pose Traced from a porn image found on google image search. Lowest effort method I do is simply paint out the background of a photo to remove background influences (Example 1), but I found a cleaner input helps so tracing a photo into a cleanly defined reference (Example 2) can help with complex things. Rendering source posers "bald" removes influence on hair, but sometimes I edit a blob of hair into the guide to force a certain length or style. Example 2, traced some image with the brush tool in Photoshop. Example 3, shape tool + stroke outlines in Photoshop. First Vivian x Miku pose guide was made using that method. Another method I do is taking the outputs of Stable Diffusion, editing & reusing them as the pose guide. Can also edit together aspects of multiple gens. This method keeps pose, but re-rolls other things. Like I'm telling the AI "Keep that pose you did here, but change all this other shit," >haven't genned anything in a month Cool styles though. I was hung up on video gen for a while, but got back into image gen more. I thought I'd get back into LLMs, but I wasn't that impressed with the Beyonder model. It retains shortcomings I saw in lesser models so it's okay, not mind blowing.

>>1718438 Ye horndog. Good gens as always. >>1718446 We use a similar workflow then. I've seen people suggest to use screenshots from games with a photo mode like Cyberpunk, or games like Koikatsu and HoneySelect that allow one to pose characters but it's too much work compared to simply opening PS and cleaning up a stock pose. >I was hung up on video gen for a while, Still don't have any interest in current video gen. I've always preferred paintings, illustrations, comics etc. >but got back into image gen more. Things have been crawling to a stop for local image gen it feels, yet these NoobAI and Illustrious models are still lacking in terms of understanding in a lot of domains like backgrounds, complex scenes etc. Often it's like pulling teeth trying to get them to gen the things you have in mind. They're basically only geared towards 1girl. Styles are not a problem, composition and understanding are. Also might just be me being out of the loop but it feels as Civitai has also died down quite a bit from months ago where you'd get constant new updates and interesting models. All the people I was following barely post anything (me included). >Cool styles though. 2nd: https://civitai.com/models/1614294?modelVersionId=1826939 3rd: https://civitai.com/models/113055 Model: https://civitaiarchive.com/models/1000708?modelVersionId=1272871&is_nsfw=true&is_deleted=true The good news is that I've busy trying to make a vidya game. Pics unrelated.

The OP is kind of out of date and I haven't been genning for about 3 years now. How is it currently, what's in use now? Will we ever get simple dedicated program instead of hacked together webapps running as localhost? If one were to start genning from scratch, what would they need to do?

>>1718572 Yeah, I had also thought about using some 3D posing software for guise images since I'd have unlimited freedom for poses & camera angles with that, but that does make it even more of time sink. I think I have the 3rd style LoRA already. I don't recall if I have the 2nd one though. I hadn't tried that model. The evolution of it to v4 looked like it was trying to get more consistency & contrast, then up to v6 it's hard to spot differences. I suppose checkpoints can reach a level of refinement where there's not much more you can do to them. >>1719583 Yeah, should at least include a link to SD reForge, but I looked at that & there's Forge/reForge2 branch being talked about on their git so even a link to reForge may be outdated.

>>1718572 >Things have been crawling to a stop for local image gen it feels, yet these NoobAI and Illustrious models are still lacking in terms of understanding in a lot of domains like backgrounds, complex scenes etc. Often it's like pulling teeth trying to get them to gen the things you have in mind. They're basically only geared towards 1girl. Styles are not a problem, composition and understanding are. There's a part of me if NoobAI/Illustrious have pushed SDXL's capabilities to its limits, and if SD3 being cucked trash and Flux being too complex to modify have kept any developments from being made. Then again, I get the feeling that AI's capabilities as a whole have seriously plateaued, especially locally-run models. What was the last gamechanger we've seen, Deepseek?

>>1719990 WAN

Does anyone know which site this is from?

>>1720363 From simply searching the name of the model, I found… https://crushon.ai/

>>1713128 Date just dropped, September 25th. Now prompt her winning GOTY awards.

>>1719990 This is something the industry has noticed and worried about. It appears AI progress was on a Sigmoid curve and we're approaching the maxima. Some researchers believe if the models are fed enough data they'll suddenly leap forwards, but there's no evidence for that and a bit against it. Now asking again - how would one start genning locally today? The methods I knew are wildly out of date.

>>1721452 AI need a better structure to do their things currently its like trying to force a hentai dick on a real pussy

>>1720061 I don't care for image-to-video, but I admit that it is a step forward. Still, if image-to-video is the only real frontier, then that's kind of boring. >>1721452 >Some researchers believe if the models are fed enough data they'll suddenly leap forwards Well, how much more data is there for AI to access? With AI churning out mountains of low-effort garbage and making legitimate sources of data such as Stack Overflow and other support forums obsolete, it's just going to start feedback looping. Also, I've seen other research that says as AI is fed more data, its capabilities converge with other models fed a similar amount of data since they're all drawing from the same well. It might be a case of diminishing returns, like you said

>>1721452 probably not as out of date as you'd wish github.com/Panchovix/stable-diffusion-webui-reForge https://civitai.com/models/1217645/sih https://danbooru.donmai.us/wiki_pages/tag_groups

>>1715969 This is relevant to a character I was messing with earlier. >>1702733 The model can't draw a revolver, holster or a hand holding a revolver very well. It also always draws her boots differently, but I'd heard character LoRAs fail at that a lot.

>>1722871 Neat. As for the revolver: Illustrious is pretty bad at revolvers in general and the Big Iron LoRA made no effort to fix that (most of the pics aren't even using a revolver). The recommended resources does however link to a Single Action Army LoRA I made (and that has a Model 10 I also made linked, but that's generally too new for Big Iron).

Did rough lines again, but probably went too rough. 0.9 weight - Kaolla Su [IL LoRA]: https://civitai.com/models/1362218 0.5 weight - Ken Akamatsu [IL LoRA]: https://civitai.com/models/1361986 0.8 weight - Sketchy Valentine [IL LoRA]: https://civitai.com/models/1209865 0.8 weight - Pixel Art [IL LoRA]: https://civitaiarchive.com/models/466454 (microsoft paint \(medium\), pixel art, sketch, graphite \(medium\):1.4), (watercolor \(medium\), hatching \(texture\), traditional media:1.2), partially colored, soft color, >>1723097 I did try that Single Action Army Revolver LoRA while making that image, but JANKU was still doing a messy job drawing it. I think the character LoRA influences things too much since artist tags do very little without lowering the weight on the character LoRA. In the end, I decided not to complicate the tug-of-war going on & omitted the revolver LoRA altogether, falling back on Photoshop & inpainting to make the things it was struggling with look somewhat presentable.

>>1721452 I just want AI to be able to do pixel perfect sprites and I'll be set for life. There are many good styles yet but many still have these blurry pixels and all. Being able to create 2D games of your dreams will be the best thing ever, and finally we won't be limited to the same few characters who happen to have official sprites.

>>1723498 Even if possible (there was an extension for this IIRC) manual intervention would still be necessary for consistency. For example, if you're trying to generate a character from various angles & poses that may work if their clothing is simple. If they're wearing anything with stripes, forget it. The position, thickness, spacing, etc. of the stripes will generate differently every time. It's the same outcome with certain accessories.

>>1723498 you can do downscale upscale shenanigans, the elin thread on cuck/jp/ has done some individual work on the subject but nothing ultimately standardized

>>1677667 Can you use pre-made voice profiles for it?

Is this powerup even worth my pearls?