>>31846 Nice, thanks for the bake

More Nobeta love

>>31850 She's cute

>>31850 Aw sweet. Ghost sex.

>>31842 bump

"""Easy"""Wan2.2

Let's see if 8chan strips metadata from mp4 files

>>31842 >>31853 Do you mean -5::artist collaboration:: or what are you talking about?

>>31857 something like that, it was a bunch of stuff + that

>>31853 >>31858 My usual positive prompt has this shit: masterpiece, best quality, no text, -0.7::canvas texture::, -5::artist collaboration::, and if I'm prompting sex I add "nsfw, uncensored" pretty much always not sure what else is there.

>>31859 I usually have masterpiece, very aesthetic, best quality, -5::artist collaboration:: and the nsfw, uncensored, but I can see how the canvas texture thing could help too. Maybe I'll give it a try.

>>31860 removes the annoying grain from some artist tags and I decided to add this to my preset, didn't seem to make things worse after a couple of tests I tried it with.

>>31856 Lol this is hilarious ComfyUI attaches a 600kB uncompressed workflow to each video or image generated by EasyWan2.2, which is stripped by 8chan after the file size is recorded, as a result all EasyWan2.2 videos uploaded to 8chan have their displayed filesize off by 600kB

anyone think visa or mastercard will let me sell a coloring book?

I hate that I can't keep the hair tone consistent

>>31865 Probably not. >>31866 Cute

I stole a bunch of prompts and shadowheart'd them

>>31865 NAI manage to accept card payments, so... maybe yes? very cute style

I'm really enjoying this drunk wives arc.

shuten anon consider the following

>>31869 >>31868 cute heart >>31873 based

>>31877 very cute! >>31860 >>31861 I got a nai sub again and started running into that grain on old prompts that used to look great. The new model has amazing prompt adherence compared to V2 when I last had a sub. But V2 seemed to default to producing something pretty if I added the quality and aesthetic tags. Now, it takes more prompting. stippled texture, dithering and halftone texture look like they could be worth some more experimentation, similar to canvas texture. They each have a slightly different effect on the style of the image.

>>31878 yeah something went very wrong when they baked v4/4.5 but at least those negs more or less solve the issue also feels like the quality tags give the model tunnel vision more than ever

Funny that recent discussion happens to be about NAI prompt wrangling because I've been bashing my head against it lately. 4.5 really likes to lean into its aesthetic tuning which seems to be a sort of scratchy, "sketch" style with thin black lines. There also seems to be issues with colors not saturating if you aren't explicitly prompting colors as part of clothing or the background. Overall it's very weird, like it's overbaked in some parts but underbaked in others. NSFW features often don't translate well between artists. -1.2::canvas texture:: is necessary for traditional media in my experience since those tags have to be highly upweighted and that causes artifacts. I don't see much good from it otherwise. I also don't see much of a difference between negging artist collaboration vs putting it in positive with negative weights. It also mostly helps with multi-subject pics, allegedly. Negging background dataset and clutter are probably the most obscure things that are actually helpful, thank you shutenanon for inadvertantly sharing those. My negs right now look like this: 3d, lowres, worst quality, low quality, very displeasing, bad anatomy, blurry, clutter, upscaled, jpeg artifacts, scan artifacts, negative space, blurry, film grain, chromatic aberration, too many watermarks, stippled texture, halftone, screentone, multiple views, unfinished, artist name 3d can be removed if it's hurting soft shading. >>31881 Cute style, I haven't used kantoku in a while.

>>31881 Cute kantokus. >>31882 I had similar observations about 4.5. For saturation issues I found it easier to use https://git.mmaker.moe/mmaker/sd-webui-color-enhance instead of trying to fully control it with prompts. Other than that, -1::sketch, flat colors, oekaki, jaggy lines, monochrome:: and variations is used in most of my prompts, without it the scratchy base style indeed seeps through pretty often.

>>31882 >artist name The "no text" tag usually cleans those and signatures up for me.

>Funny that recent discussion happens to be about NAI prompt wrangling because I've been bashing my head against it lately Most of my gripes were pretty much solved when vibe transfer was implemented, and once I figured out how to get consistent compositions I felt like I've been able to enjoy 4.5 to its fullest. Now it turns out we're getting a Character Reference feature which should help for OCs/lesser known characters and that excites me.

>>31882 >>31883 This reminds me, is kantokuanon still around? the anon who posted picrel in the based64mix era (this image sticks in my mind, in a good way)

>>31885 I agree that VT goes a long way, but it can be a huge pain in the ass to wrangle compared to (comparatively) simpler artist tag mixes. Also it sometimes gives me weird artifacts but it's heavily dependent on references.

>>31887 you can combine VT and the artist mixes provided that you use lower VT strength/IE values i wish 4.5's IE could be less black box-y or that it at least worked like v4's but oh well at least you can save vibes for later would be cool if someone were to reverse-engineer the VT encoder but i doubt it'd be worth it just to save like 10 anals per pic at best

>>31882 >4.5 really likes to lean into its aesthetic tuning which seems to be a sort of scratchy, "sketch" style with thin black lines. yeah, it's really bad compared to even v4's base style and most traditional media-related "(medium)" tags don't seem to do shit (absurdly high weights aren't a good solution) once again just like on v4 >There also seems to be issues with colors not saturating if you aren't explicitly prompting colors as part of clothing or the background. works on my machine and i never have to specify colors unless i use "monochrome, spot color" or something like that, did you accidentally turn cfg rescale up? 0.5+ will start to kill your colors >Overall it's very weird, like it's overbaked in some parts but underbaked in others. i agree, this was the case for v4 as well but it's even worse here >NSFW features often don't translate well between artists. it's a compounded issue on one hand, some artists are woefully underbaked and some are very overbaked on the other hand, the model actually learned (or tried to) the styles properly and as a result everything is artist-dependent unlike v3's more rigid and prominent base anatomy >I also don't see much of a difference between negging artist collaboration vs putting it in positive with negative weights. the general consensus is that negative weights in the positive prompt is always better than using the negative prompt itself because something something the negative prompt is just a shitty hack that often backfires and isn't supposed to exist or something one of my biggest gripes is the lack of samplers, every single one of them aside from euler a (and maybe regular euler but i've seen it fail quite often too) is either completely fucked up or just worse enough to not be a usable alternative it's a shame because the original SDE is the best sampler when it comes to darkness

>>31887 my greatest tip for you is to use img2img, that + VT is easily the strongest tool in my opinion. Once you get your composition in a gen, you can set the img2img strength anywhere from 0.35-0.65 and by that point you can set VT reference/IE as high as you want and it wont change the comp. This is also how I learned you can reproduce an artist's style with greater accuracy than the model knows.

>>31886 Kantoku anon has not been around for a long, long time as far as I know.

easyWan22 gets frequent updates, and every time I reopen the workflow, new features appear, glitches pop up, or old ones get fixed. The structure is such that I can barely touch anything except the prompt and ON/OFF settings… Made it from a single still with manga symbols, so yeah… it looks kinda off as a video.

>>31891 RIP. really enjoyed his gens

>>31890 that's really smart. with a bit of testing getting really good results. very cute gens too >>31882 I experimented a bit with looking at what some of the neg tags did in positive prompt and then comparing with and without those tags in neg. chromatic aberration seemed not actually useful, I think there are too many good images in the dataset with that tag for it to be meaningful. it was almost the same as negging amazing quality, composition would just change a bit without being better. neg artistic error seemed to make an improvement. blank page also suprisingly good in neg. having both worst quality and bad quality in the negs also seemed to make it adhere to the prompt a little less without a quality bump. in no particular order, I found these neg tags to be good: lowres, artistic error, film grain, worst quality, very displeasing, dithering, halftone, screentone, canvas texture, stippling (texture), signature, watermark, simple background there's also a huge difference in tag behaviour between 4.5 full and curated. I accidentally did some testing on curated and it did not transfer at all. >>31883 this is interesting. I'm still a bit confused about -1::tag:: in the positive prompt. but -1::sketch:: definitely turns up the detailing by a lot! >>31889 I used to go for DPM++ 2M or 2Sa, but euler a seems like the only usable one now

I uploaded a video with the wrong resolution.

>>31892 >>31895 these are impressive. clean loops even. maybe don't gen the original image with motion lines though, it looks distracting when animated.

>>31894 Thanks for sharing your findings as well. My negs now look pretty much like yours except -simple background and +clutter, negative space, blank page, blurry Anything past that feels like placebo for blanket negatives. In case anyone is curious or unaware, negative weights in your main prompt are *semantic* negatives, the model will try to steer in the *semantic opposite* of that concept. Think -1::flat color:: giving more color depth, or -1::simple background:: adding more details. The negative prompt box are *visual* negatives, during the inference process a second set of latents are generated using them and subtracted from the main prompt.

>>31900 >semantic vs visual This makes a lot of sense. The docs said something about the vector being scaled by the number in +-N::tag::. But what is still confusing is how the model would learn a semantic opposite that's meaningful? Like prompting for 1girl or -1::1girl:: actually kinda works as expected. It's also interesting that prompting 1girl or -1::1girl:: with add quality tags enabled and heavy undesired content preset both give a 1girl image. Something in the default positive and negative prompt strongly steers the image. If you're prompting for violence, I was getting a lot of bloody and horror gens while testing when I had no or few negative keywords. Simplifying the negs is something to keep in mind if a prompt is stubborn.

>>31885 Sakura's face is so perfect here. I need style like this on local...

so if v4's negative vibe trick doesn't work on 4.5 how DO you wrangle it?

>>31903 Didn't an anon just write how? Gen without vibe, then i2i with vibe at 0.6-0.7 denoise.

>>31904 tried it, it works but it's annoying and just adds more steps i also tried the negative strength trick with IE at 0.1-0.4 again just to confirm that it doesn't work and yeah it doesn't (i also got a super bright image just like another anon said) was trying to use this image and no matter the strength/IE values i use (i tried every IE in 0.1 increments) the damn towel is near-permanently burned into the image (even after negging beach towel, towel at -1.5)

>>31906 >tried it, it works but it's annoying and just adds more steps not to mention that the img2img strength slider also acts as an additional and global VT strength slider

>>31906 Yeah that other anon was me, though I'm not subbed anymore. You'll just have to get used to the annoying extra-steps I'm afraid. I don't think there's much here to come up with.

>>31883 >https://git.mmaker.moe/mmaker/sd-webui-color-enhance doesnt work for me throws errors like this ValueError: the input array must have size 3 along channel_axis, got (792, 424, 4)

>>31909 Ah yeah it didn't work for me as well I had Claude fix it for me, I think I only had to change mmaker_color_enhance_core.py Seems to work well like that https://pastebin.com/1X25vWvR

We have seia_(swimsuit) at home

>>31912 Wow you can get surprisingly close with some tag wrangling. Definitely one of the top chars I'm looking forward to testing references on, tho.

One, two, or three?

>>31914 Style wise?

>>31915 Yes.

extremely high quality thread so far, great job everyone

>>31914 All of 3 of them looks like xerbatt at some degree so idk, maybe the second one

>>31899 catbox please?

>>31919 https://files.catbox.moe/je939i.png

>>31920 thanks bwo

>>31893 >really enjoyed his gens Same. >>31892 Even if the Ai is very inconsistent at it, they give quite another vibe without their bodypaint.

At that time, I was desperate to see Yuuhi and Joel mating with single-minded devotion. >>31922 Nostalgic image! I remember seeing the potential of SDXL when I saw that even when a girl's legs reappeared from behind the boy figure, it rendered flawlessly without glitches.

>>31922 first image is nice but man the anatomy is broken on the last one it's almost painful

>>31922 #1 very cute. for certain scenery and landscapes local models still have a very nice vibe compared to latest NAI

>>31891 I saved a lot of his stuff, he was one of the good ones. >>31898 Very nice

>>31896 >>31895 very nice vids! any chance you share your workflow ? I'm just getting started on wan and kinda lost lol any tips and tricks are also appreciated, like how to do perfect lops is it related to prompt or do i need an node for it ?

spoilering some furry

>>31928 Honestly, I’m just using the easyWan22 workflow as-is without tweaking anything. I just turned on a generic NSFW LoRA and fed it a Kinomoto Sakura pic I found on Sankaku Complex!

>>31929 That third one is really hot actually but he must have a dick of steel. It's bearing like 200-300 pounds.

>>31929 >spotted fur I see you are a man of culture as well

EasyWan2.2 has a tendency to deviate from the brightness of the original image

Fixed negative prompt in machine translated Chinese? What purpose could this possibly be serving 色调艳丽 – Vivid color tone 过曝 – Overexposed 静态 – Static 细节模糊不清 – Blurry details 字幕 – Subtitles 风格 – Style 作品 – Work (or piece/artwork) 画作 – Painting (or drawing) 画面 – Frame/scene/image 静止 – Motionless 整体发灰 – Overall grayish 最差质量 – Worst quality 低质量 – Low quality JPEG压缩残留 – JPEG compression artifacts 丑陋的 – Ugly 残缺的 – Defective/incomplete 多余的手指 – Extra fingers 画得不好的手部 – Poorly drawn hands 画得不好的脸部 – Poorly drawn face 畸形的 – Deformed 毁容的 – Disfigured 形态畸形的肢体 – Deformed limbs 手指融合 – Fused fingers 静止不动的画面 – Motionless frame 杂乱的背景 – Cluttered background 三条腿 – Three legs 背景人很多 – Crowded background (many people in the background) 倒着走 – Walking backwards So this is what causing the color tone shift

>>31927 cute. two finger nipple cover is kino

>>31936 Thanks and yes it is but I'm now overdosing on lolidom.

>>31938 coincidentally I was overdosing on the same thing in the NAI text gen. text gen is insanely intense, I'd forgotten just how much so.

>>31939 Based, I tried NAI's textgen for a bit as well and it was alright. I'm using local LLMs now but I'm a bit limited by my hardware.

>>31935 The negative Chinese word for 'wan' is necessary… As for the color tone shift, it’s due to the high-speed LoRA, so you need to use Refiner’s ColorMatch.

>>31939 finding someone still using NAI's textgen is like running into an uncontacted amazon tribe deepseek would rip through your people like smallpox

>>31942 deepseek doing loli nsfw? hmmm also, cute confused loli

how you Anons manage your loras ? i have a shit ton of loras but still using folders for them there are many projects in github but idk which to use any recommendations ?

>>31944 Had all my loras in one folder with loraname_type_subtype_model_trainer.safetensors naming structure, worked for me. Example: guweiz_style_artist_vpred05_lazylora.safetensors

>>31944 I used to just have folders for artist/style/concept/wip and name them accordingly as [type]-[name].safetensors, [type]-[name]-[dim]-[version] for ones I made myself.

help

Man I wish anime models could take instructions in natural language It'd be so much nicer to give simple instructions to the model while inpainting or just say what you want instead of fiddling with tags for hours to achieve the desired result

>>31950 godspeed anon o7

>>31951 sometimes switching to flux for inpainting with natural language and then doing a final i2i pass with anime model works pretty well. NAI understands natural language surprisingly well too.

>>31943 yeah deepseek is hardly filtered compared to the rest of the large models, i've gotten more refusals from the chatgpt data they trained it on than the model itself

>>31941 Oh thanks, didn't know about that option

what's the newest and hottest shitmix? noobaiCyberfixV2_10vpredPerp is fine except it loves long torsos and hates lolis.

>>31956 I've been having fun with GLM-4.5 lately.

>>31958 I saw some people shilling https://civitai.com/models/1217645/sih and gave it a shot, it's pretty good. It can handle "normal" cfg ranges just fine (4-6 on non-++ samplers). A little less refined in base res but the colors are good and after upscaling it's similar to 102d with less noise. >>31959 GLM-4.5 is really good, especially for coom. It doesn't self-censor or beat around the bush like other recent "big" models so you can escalate into explicit NSFW very easily. No idea if there's a free option but it's dirt cheap, I'll use it as my daily driver and switch to Opus 4.1 for a few messages when I need the big brain EQ.

>>31956 >>31959 what service do you use to run the model? I tried synthetic for some sfw stuff >>31960 no concerns about data retention?

>>31960 nta, is glm-4.5 decent for regular roleplay/narration or am i still better off with gemini/deepseek/-god forbid- gpt? (i don't have opus, i don't have a sonnet proxy and even if i wanted to give OR money i don't want to give them *my* cc with *my* name on it)

>>31961 >no concerns about data retention? I use OpenRouter which accepts crypto if you're especially paranoid. OR pseudo-anonymizes data purely by how it works and the providers I use (Novita and Parasail) claim not to retain data so that's about the best assurance you can get without going full local. >>31962 Yeah, I like it a lot for general use, genuinely third place for me behind Sonnet 3.7 and Opus 4.1. It's around the same smarts as DeepSeek despite being half the size and pretty creative. If you do try it it works best with a very simple system prompt (and optionally instruction to length limit).

>>31963 >Yeah, I like it a lot for general use, genuinely third place for me behind Sonnet 3.7 and Opus 4.1 never got to try 3.7, rip also i guess i know what i'm doing today (hunting down the latest merkava link and hoping it has glm)

>>31964 3.7 is still around with no planned obsolescence date, it's 3.6 (aka 3.5v2 fuck you anthropic) that's getting discontinued in October.

>>31965 still not getting a chance to try it unless a new proxy pops up and it doesn't require me to pay in crypto, i cba to get into fake computer money anyway, glad there's a new good enough generalist model, i was getting tired of having to juggle gemini/gpt + deepseek for loli

>>31963 thanks for the info, I'll check these out still very surprised the models don't argue about >The girl squirms helplessly, moaning loudly. You fuck her with a steady rhythm, pounding her sweet little cunny.

where my shutens at?

>>31961 I use OR too and pay with crypto.

>>31961 i just use deepseek's api directly, i'm least worried about something causing me trouble down the road via a chinese company than anything more western

>>31968 based have a heart

>>31972 Do you apply a vibe before img2img and after upscaling? Copying the prompts doesn't come out quite as soft.

>>31973 I gen with a vibe no img2img unless im going for something specific, & I gen at high res without upscaling; cost way more anals though

It makes videos where the weirdness isn’t so noticeable, so I kind of ended up making a tentacle_sex video without thinking too much—but I actually like how it turned out.

very glad i wasted money on a 3090

perspective is probably fucked up in the last one and I'm not sure how to fix it, but I kinda like it still

>>31976 Could you please share the workflow and prompt of your first video? I can tell it's not the latest EasyWan2.2 by the size of it >>31978 >water level decreases as liquid is injected into her ass Is this prompted? Video gen is just so damn more coherent than image gen

>>31981 >Video gen is just so damn more coherent than image gen bro what

>>31978 imagine what you could do with a 5090

>>31982 I know it sounds ridiculous but that's my impression so far Local video models are better image models than local image models, they generate nonsensical motions most of the time yet their capability to preserve styles and character details are above that of any SDXL models

>>31971 hehe cute, thanks anon

>>31976 1,2,4 are hot! >>31979 silhouettes look great, tried prompting that once before and failed

>>31984 video reminds me more of live2d rather than actual animation tbh

drunk and sweaty wife

>>31981 this should be everything. nothing special, just boomerprompting. no, it did that on its own every time >>31987 yeah with real people it's a lot more natural but i bet it was trained on a bunch of live2d

What the fuck happened to the "main" thread? I've been out of this for a while, the 4chan hdg looks like an aftermath of a nuclear war and subsequent zombie outbreak.

>>31990 excellent wives by the way

>>31991 i need to get me a better gpu some day

>>31993 thanks, have a few more

>>31992 It's been shit for quite a while yes, same reasons as always, insane shitpost, bots and miserable moderation support

>>31992 It went to shit around the time NAIv3 released because of retarded console warring. Became usable shortly when Illustrious dropped, then immediately went back to shit. Pretty much a dumping ground for shitposts that the jannies refuse to do anything about, which I can't really blame them for since it contains the schizos. /trash/slop/ is ironically the best thread for imagegen on 4ch but you can't post loli and it's started to attract schizos recently too.

>>31995 fuckin wew that second one >>31960 that's a pretty solid shitmix, appreciated

>>31996 >miserable moderation this. open troll posts and spam run rampant while minor stuff gets the deleted and you get a ban for even reporting.

>>31981 >>water level decreases as liquid is injected into her ass >Is this prompted? Video gen is just so damn more coherent than image gen that's so stupid i had to do a double take

>>31991 Could you catbox it? In general video gen workflows cannot be embedded as stealth metadata because of their size

>>32002 oh, sorry. here: https://files.catbox.moe/qj4j5p.png

>>32003 Thanks, so the hair jiggle is from the base model or the NSFW lora I guess

>>32004 the nsfw lora does like to make things jiggle but wan 2.2 is pretty good at doing that on its own as well

>>31992 one (1) nai v3 user baited the localkeks one (1) time and they retaliated by making sure the general would never ever be usable ever again going forward

>>32006 it's the same two or three people on both sides of every argument, you should know this by now

>>31992 They're just inventing new schizos and larping as them on the daily now. It's crazy.

>>32007 It probably wasn't like that at the beginning but yeah it doesn't matter now, that general is fully ran by schizos and bots.

is it controversial to say that I like noob shitmixes more than 4.5 for lazy 1girls

>>32010 I don't think stating your opinion could really be considered controversial I think 102d has nice default anime style and it's kinda just easier to work with compared to 4.5 for lazy prompting.

>>32010 after using NAI again after a long stretch of only local, I think this is accurate. but once you get your particular favorite tags on NAI it is also pretty easy to get lazy 1girls

>>32007 after the nai posters left, sure

>>32010 wrangling 4.5 is needlessly frustrating at first, i don't blame you if you'd rather not deal with it for just 1girls

>>32014 shut up bitch

>>32016 lol make me, schizo

qrd on running WAN i2i? do sex position loras off civitai just work now?

>>32018 i2v oops

>>31992 I've given up on reading the threads for a long time Now I just scrape those threads for pics with metadata, and in rare cases a lora is posted maybe extract the reply chain for context, fully automated process

>>31992 the tranny shitposter known as cathag happened i think we should genuinely kill him

>>31992 what happens to people when cunny makes them seethe

>>32018 Start from here https://github.com/Zuntan03/EasyWan22 In my limited experience Wan2.2 loras with the exception of a few well trained ones malfunction half the time

>>32018 i just use the generic nsfw lora and everything i've tried works fine

Just dumpin some cute wife

>>32013 can you do a creampie of the second pic?

>>32027 cute!

>>32031 thanks!

anyone knows if there's extension that can add hsv noise to gens in post? would be more convenient than opening external editor

digging through the internet archive trying to scrape together enough of one artist's work to train a decent lora and man link rot really blows

forget it, his style isn't stable enough and i could probably get more or less what i want with norasuko plus some ~2005 capcom watercolors. it was mootan, he's got a few on sadpanda. a thread on /v/ that had some of his DAoC fanart got me curious.

I really should’ve fixed the finger count before turning it into a video. It didn’t catch my eye in the small preview, but after enlarging it with the refiner, it became obvious. Still, I’m happy with how the motion looks, and I feel like with a little more tweaking, it could get even better.

>>31987 It's heavily influenced by the style of the original image WanVideo2.2 performs better if there's realistic shading in the start frame

Sorry for posting so much. I’d be happy if there’s anything among them that you like.

>>32040 i like the animations, they're neat

>>32040 don't think you're posting all that much, it's fine, especially considering some anons like it. honestly though since you asked i almost completely ignore animations because to me they still look bad

>>32040 I've saved quite a lot of them. If I had more than 8GB VRAM I would try genning some as well.

>>32043 earning extra credit

>>32047 my little hitler can't be this cute...

Streaming wife

She must be raking in those donos.

>>32051 As long as she allows Monero.

>>32043 what are the artists for that first one? very crisp.

>>32049 😭 uuooooohs in the chat

>>32054 literally me

>>32056 cute hinata

>>32053 https://files.catbox.moe/kqbxgx.png

Trying to do anything with lipstick, without it turning into full bimbo lips, is a pain in the ass

>>32043 could you post box of the other two, p-please?

>>32060 The second and last one are really hot.

>>32062 hungry shotacon trainers not that much older than their prey uoooh