Recommendations:
steps:35 cfg: 3-5 sampler: euler
Description: This model is the result of extensive analysis and refinement based on multiple popular anime-style models, which often struggle with issues like detail loss, poor framing, and repetitive elements. To address these challenges, we’ve directly integrated a custom LoRA into the model, ensuring that generated characters are more coherent, detailed, and visually consistent.
While this model is still a work in progress and not yet perfect, it already delivers more stable and reliable results than most alternatives currently available.
Key Features:
Optimized Framing: The built-in LoRA automatically adjusts character composition, preventing unwanted cropping and ensuring all critical details are preserved.
Reduced Repetition: The model minimizes repetitive patterns, delivering more diverse and natural results.
Enhanced Anime Quality: Perfect for users seeking a highly detailed, expressive, and professional anime style, with rich textures and balanced proportions.
Included Rule 34 (R34)
Description
Background Clarity Specialist
Fine-tuned model focused on fixing blurry and poorly defined backgrounds. Improves background sharpness, structure and overall coherence while preserving the main subject. Delivers cleaner, more solid environments in both anime and semi-realistic styles.
FAQ
Comments (16)
Damn, was literally wondering to myself the other day if this checkpoint would get another update.
Out of the 230+ models i've collected and tried this is easily one of the best and not just for NSFW (Though that is the main draw :P)
Thanks for your hard work!
Thank you very much, I'm glad you like the model, share your images here from time to time please.
@Lacike I will. Been putting it through it's paces and testing some things on it, will have some images and feedback soon. Have you considered adding support for E621 tags btw? My current checkpoint supports them and it really opens up a lot of options tag-wise (not just furry related, it has actions, poses, expressions and concepts that you don't get with plain danbooru tags.)
@NeilSylver I'm doing this little by little,At the moment I'm studying in adding expressions, but not just the face, the whole body expression of each feeling, it's a little complicated, it requires posture study, but it will work. I'm being careful not to confuse the model by adding too many things. I want to be precise in updating. ( I don't just add the tags, I do all the description)
@Lacike Damn, that's some serious effort, and tbh it really does show. I wasn't kidding in my other comment when i said the previous version was one of my favorites, No. 2 in fact, Cinema Metropolis wasn't necessarily better qualitywise, i just found it slightly more flexible with the extra tags, but it was flawed in a few ways, struggled with shortstacks and alien/monster girls, making them taller/humanizing them respectively along with a few other concepts it struggled with. You've blown it out of the water with this version though. I've just about finished testing and am going to post a full pack of images of various types (Some NSFW for the gooners obvs) along with a rambling write-up of my thoughts, so look forward to that :)
@NeilSylver I will love to see your ideas.
@Lacike I tried to stress test it on a few things i know models struggle with. If i'd known you were refining expressions i'd have included more varied facial tags. I'll have to do another pack soon with face focus shots and different expressions/emotions too.
@NeilSylver This helps a lot, it is more important to have good images than quantity when training the model. And having the other person using the model besides me helps to better see where to start.
@Lacike No worries, i actually enjoy doing this kind of thing. I tried a fair bit, have just finished paring the images down to 20 and am going to write about what i was doing for each one just before i post them so there will be a lengthy ramble i'm afraid :) (Most are relatively tame but i did focus on a few NSFW themes. I'm just hoping Civit doesn't shuffle the order and frontload the most NSFW one lol, they'll be in order on my gallery though for some reason)
Been using this since it was released, thought i'd finally share my images and thoughts. Second attempt writing this out btw (accidentally backed out when i was halfway through) (╯°□°）╯︵ ┻━┻
Warning: HUGE wall of text about each image incoming, TLDR at the end.
Befor i ramble about my images i want to talk about the features mentioned above, first, the Optimized Framing is incredible, I don't think i encountered a single generation with excessive negative space, off-centred focus or something missing from the frame that i wanted in it. Looking over my image prompts i also realized that i never had to tag 'looking at viewer' or anything similar, it just knew when i wanted that to happen and did it.
Secondly, the Reduced Repetition aspect is readily noticable as long as you don't lock your prompt down with excessive tags. Varied, detailed backgrounds of every type as far as i could find, same goes for poses and expressions, with more than enough difference between images to keep them fresh.
Finally, the Enhanced Anime Quality, it has clean, sharp outlines and great shading that isn't overbearing, but can be pushed to different styles readily, keeping the crisp and clean look (I personally had fun with 'toriyama akira \(style\),' then added retro artstyle, 1980s \(style\), to compare with his old art style.)
Now, let's see if these comments have a character limit as i ramble about each image i'll post. If they're shuffled in the gallery, you might have to go to mine to see the correct order :/
All of these were generated with txt2img with no quality tags other than those in lazypos and lazyneg, at 1344x1536 with a 1.5 upscale and face detailer (Though i could argue that they seemed unnecessary for the most part.) I also varied the backgrounds as much as i could, using single tags for the most part. There were many more, so count yourself lucky Civit has a 20 image limit.
First there should be four images of Princess Zelda, one without a style, then ones with 'photorealistic', 'anime screencap' and '3d' respectively. I intended this to be a fingers and proportions test with a closeup background, knowing that some janky merge models out there wreck furniture proportions without a refernce point. Proportions were consistent, I saw no issues with the fingers despite them being obscured in some manner. These also function to show the variations between backgrounds, poses and expressions with them all varying quite well between images despite using the same seed.
The next two are the goth punk girl and warrior, both serving to see what outfits and backgrounds are generated through minimum prompts, along with a lighting test. The punk girl surprised me with the fine details, a stable shirt print, tattoos and appropriate piercings, though there is a slight error on the forearms regarding the fishnets (not as much as i expected considering how tricky they can be when overlapping details, though 'fishnet sleeves' did reinforce it enough to stop any issues. Backlit by the moon let's say, without any unwanted lighting on her front.
The warrior is more of the same, with a well detailed outfit and background with a minimal prompt along with great lightning effects proving backlight, I also added 'sunken cheeks, sunken eyes, thick lips,' to see the effects on her features.
Then we have the anime and finger test, with the ganguro girl. I'm certain the same looking character is in every model, but she demonstrates the art style well, and i wanted to see what would happen with her hands if i prompted 'pink nails', expecting limp arms with her hands resting on the seat, yet every generation was indeed different and natural looking. I would also like to draw attention to the background, where motion blur was seemingly added, when most models would provide static shots, that surprised me.
Next is the horror monster/non-human test, seeing how well the model interprets H. R. Gigers xenomorph. It's certainly very recognizable and symmetric and the model doesn't overbellish to the point it isn't.
The next two are blackhole sun (♪♫♬) spacescapes, characterless, background shots to test the optimized framing to see what it would do. The first was very centred and symetrical so i added 'asymmetrical composition' to see if i could trip it up in some way but it handled it amazingly.
Next is a quick furry test with a reflection test that really surprised me, if not for the arm it would be perfect, it also avoided my intended trap of putting the prompted bell on the back of the collar like so many models do with front-worn accessories when the character faces away.
(Not far to go now :P)
Robot girl next, checking for joint and symmetry issues, and to see if 'laboratory' would generate a generic lab background but it didn't
Now we're getting to the spicy ones, i'll be brief with these and just point out the main things then move on quickly.
Goblin girl, checking toe count and body type, seeing if they're consistent enough. I actually didn't see a single issue with toes across multiple gens when i was certain i would, body type and size stayed consistent too, background varied greatly considering.
Frog girl, testing how it handles bizarre concepts like non-standard hands and long tongues. One missing finger but very consistent on the tongue with no strange nonsense going on like other models when i attempt this.
Vague concept test with the girl on a machine, checking to see if the model breaks down and makes something incomprehensible. It didn't, maybe this is the Rule34 at work but it's fairly distinct and plausible looking.
A tame sunset shot as a palate cleanser, testing multi-color gradients (managed three prompted sky colors easily) and a clothing type most models actually struggle a bit with. Any kind of string effect that isn't flat on an object tends to break down on most models (try prompting for bow-users drawing a bowstring) Despite this, it actually rendered consistent backstraps, functions as another anime test too i suppose.
Multi-armed, multi-eyed monster girl performing multiple 'actions'. I was certain this would trip the model up in some way but other than a slight fusing of two middle fingers (far better than seeing a thumb on each side and five fingers). This also functions as another furniture proportion test, with some janky merges making very small 'beds', this one is well sized and stayed that way across multiple generations.
Party gang, testing to see if multiple actions would break anything or generate awkward poses or extra fingers. It didn't and this was the most fun to test pose variety on with every image being different :P
All girl party gang, testing for duplicated appearances on a simple prompt, as you can see, all different hair colors, random accessories and an outfit to boot
and last but not least, your mom /s An extreme perspective, body type and action test. It handled this far too easily imo.
TLDR Download it, it's seriously impressive. Far more stable than any random merge checkpoints, incredibly flexible if you know how to prompt, and apart from problems caused by my own wonky prompts there was little for me to complain about (Oh, and if you struggle to make a dark scene like the other person in the comments 'dark background, dim lighting, hard shadows,' then creep up the weight on dim lighting.)
Long story short, i loved playing around with Lacike's Mega Cock, it might be hard to handle but with a light touch, you'll be surprised at what you can coax out of it.
Thanks for reading (or not). Hopefully this persuades some of you to seriously give this a go (or that i'm absolutely mental). If you somehow enjoyed this, seek help, but first check my profile where i do this nonsense for every lora i try. Peace!
I thank you from the bottom of my heart for that. I spent so much time running details and trying to do everything as best as possible that when I finished I was so overwhelmed that I just posted the model. Your tests will be very welcome, now I saw a point that I hadn't noticed, and honestly I'm very happy with the result, thank you for being part of my small work.
@Lacike Not at all, i genuinely enjoyed myself putting it through it's paces, and i'm using it as my main checkpoint right now. There were a few concepts that i couldn't include because of the image limit, such as my face tags test where i altered the face anatomy with working non-tags like 'broad nose' and such (I did use sunken eyes/cheeks for the warrior though.) If you ever need someone to test something like facial expressions, let me know. I also like seeing what characters i can prompt for without any lora's, so if you have a popular series you want me to try to showcase what can be done with raw tags, i'd be more than happy to give it a go!
@NeilSylver This about details that change the expression using face anatomy without tags interests me a lot, I have already started to work on your suggestion about expressions , and once again thank you very much for your help, I will continue to improve the model even more, I am very happy with what you did.
@Lacike It's only something i recently discovered when testing Interfusor's 2026 style lora.
https://civitai.red/posts/30564922 There's quite a few non-tag concepts that work surprisingly well, the broad nose one was from my tests with an orc where i wanted her to have a less anime-esque face. Your model actually provided a variety of different types too! I believe that if a concept exists in an actual tag like long/short/broad/narrow that it can be transferred to other things too with varying success (Don't take that as gospel though, it's just my personal experience.)
@NeilSylver This is interesting because when I train a model I try to attach as much as possible details to the tag, The richer in detail and well defined the image the more consistent the result, I will do a search in this direction. So in addition to the pure expressions, the model will have a variety of face shapes to work with without getting stuck with certain characters faces . (In my tests, when he puts a certain type of detail on the face, the model is pushing the entire face of a character who has the detail)
@Lacike It's definitely worth experimenting with, a lot of shape/size modifiers actually work for noses and eyes i've noticed. I'm still trying to find a solid one to make chins less pointed and anime-like for character like orcs generated without a lora. It's hard to pin down if i should apply them to chin/jaw or jawline though. I'll experiment a bit when i have the time and post my results directly here for you (without an essay this time lol)