Anyline -> MistoLine is what you're looking for. Restricts you to Canny-esque line-based inputs, but gives consistently good results.

thanks. I tried it but the results were not that great and often it was not following the pose. I did however stumble upon openposexl which worked great.

Fair enough! Glad you found something that works for you.

I am finding a similar struggle

+1, so i stick to v1.3, very clean base style, and compatible with all style loras

Ikena It's fine its just something we need to figure out how to work around with the prompts

Ikena standard Illustrious resolutions like 832х1216 and 1536х1024 (vertical pictures).

Ikena, no, I meant 1024x1536, sorry.

Ikena I understand that, but Illustrious 0.1 and 1.0 merges (even yours 1.3) draw with no problems a single person with solo tag.

The problem persists even on 832х1216 which is one of three default resolutions on Civitai.

thats not hard to answer the most ppl that trained models uses different models for finetune them and i know a model thats trained on danbooru tags only that does the same if the model is used for finetuning the whole model does it + vae fix for XL v2 is not 100% fixed it does alot of weird stuff in understandment of prompting so far if i build a model i use everytime the sdxl vae cuz its more stable then the vae fix for sdxl 2.0 . iam testing everytime around which models are here and what it does etc problem in this case too is as feedback pls check the prompt tag mature female for more adult faces (adult face,mature face details) it gens for some reason the bust/chest very big other versions dont do this so much but yeah its in some versions too 2.0 above . all other stuff out of everytime the shiny body is fine details on hands and face without adetailer is very nice thats a huge + so far from my side after testing it . + try to lower the cfg it reduces the shiny stuff a bit as info from me

Ikena depends on how much images u use and which poses etc are in if u use cowboy shot alot and finetune them with a broken model some prompts are missunderstood and its a focus on breasts etc then . + check the vae as i wrote in my other comment if u use the vae sdxl fix for 2.0 try to use sdxl vae for 0.1-1.0 and watch the results maybe this help u a bit to fix the problem with prompt understandment and maybe the focus will be gone

aah after some testing this point: the model ignores for some reason sometimes prompts in shots and angle and does what it want it does then a shot from above or behind what is not written. its a issue that is known from 2.0-2.2

Ikena yeah sry i translate it by my own with my bad english xD

so far u can try to bake in the SDXL VAE : after this try to test some gens and look at the results if it works better or not models that i made are only on pixai and not here on civitai and i use only the SDXL vae .

but u can try in training to use the sdxl vae and if u finetune ur base model with another version dont forgett to use the same sdxl vae that u build in the training. its a test worth , i helped one person that run in issues with the model too and asked her which vae she use : fact she used the sdxl vae fix, i said her use the sdxl vae 0.1-1.0 and the model worked better then before.

to explain SDXL vae fix for illustrustious 2.0 its not perfect at all it missunderstood prompts and does then weird things like focus on the called prompt as example u wrote matt chest lines or texture > focus on breasts/chest ( gens a extra feature like a transpartent shirt) can happend trough that

if the text is not rly better i use next time google translate i try my best to write it by my own cuz google translate is not allways perfect

Ikena for the bigger eyes its nice and a plus for full body gens, ok then i need to wait for ur update and check it again how i can manage to gen with mature female prompt without running in huge breasts

Ikena nah i use only mature female and if the face is still too young for my taste i add adult face / mature face + quality prompts and then boom huge breasts in oversize

yeah fp32 is a huge model and for fp16 its more variable , sure the fp32 model is more stable but the size is bigger then fp16 . in normal case if u have mature female include large breasts will be fine but without or with it gens huge ones also a size bigger then large .

sure try it out it should be fix the most things : and if u use a model that trained on alot of danbooru tags careful alot of ppl uses perfect hands , elegant fingers, or elegant feets, to control more the gen on good hands and feets in this case some of ur versions cuts the half image out and focus on that parts that happens on my older versions to cuz i tryed finetune my base style with some of urs but it destroyed the whole model . we need a new sdxl vae fix for 2.0 but i dont know if one has experience in it to create a stable vae for this .

so thats my experience so far and i test alot around and hopefully it helps u finaly to fix ur version for a good allrounder model thats are stable and understood the right parts of prompting without focus on that parts

Ikena Thank you, I'm looking forward to it!

Ikena ok I can wait

Verdade

not realistic at all for me. try putting realistic, 3d, hyperrealistic, photorealistic in your negative.

penguiny2357  Even do so, this checkpoint still focused on realistic-like 2.5D in the first place. so I doesn't matter, really

Add "cel_shading" in your prompts to produce anime-style images.

Ikena Overall it's a solid model, but some data isn't up-to-date yet – like certain character skins and specific poses. I'd really like to see the model generate these with reasonable accuracy without relying on particular LoRAs. Hope you can update the dataset in your next iteration. Keep up the great work.

Ikena Certain poses may be new, but based on my testing, some older poses or perspectives aren't being generated with reliable accuracy. This could stem from limited Danbooru data coverage, or the fact that many high-download models haven't prioritized training in these specific areas.

For frequently used perspectives like wide shot or female POV, even applying weight parameters doesn't yield consistent results. As for niche poses such as outstretched legs (legs extended forward) or feet on table, model still relies heavily on LoRAs to generate these compositions correctly.

Regarding characters: beyond new additions, even some old characters or their alternate skins (without relying on LoRAs) fail to achieve expected results despite stacking similar feature tags. Sometimes the original character's traits overpower its sub-variants – a prime example is Azur Lane's Laffey. When generating Laffey II or her Playboy Bunny skin, details often skew toward the heavily trained base Laffey model, resulting in missing/chaotic clothing elements. (Of course, this also stems from scarce training data for these specific forms/skins.)

Ikena Got it. Personally, I'd still suggest slightly strengthening the training for older perspective poses and legacy characters/skins. Your model is genuinely impressive though

Ikena "with the most actual data" is not means they are all new datas, why many said that is because Rouwei actually received multiple non-booru datasets include somethings you maybe won't found in other base models.

Ikena it's Okey, I exactly just need zani and fleurdelys 😈, thanks you so much for your amazing work and for giving this for us free. God of hentai bless you. Also you plan to update the amenatsu?

Ikena hold on, let him/her cook. Zero pressions. You already did the world in difference adding an updated database. And again. Thanks for sharing your amazing work. 🙏😁

Ikena I would like to suggest if you want ofc. If she isn't added: wilhelmina from brown dust 2. She is a goddess. And would be perfect to have her too. Just in case you want too ofc it's just a suggestion. 😅

Ikena by the way,i try the prompts Zani,but not get the character picture,whats wrong with me,thank you,must use lora??thank you

Ikena thank you,its wonderful,thank you very much

In fact, the 3.0 version is overly stylized, and the issues with the human anatomy remain unresolved. It may be suitable for generating relatively good-looking images with a few prompts, but it is not ideal for users who try to precisely control the image through prompts.

deadmanno3 I see, that's unfortunate. I like the style, but having to do so many different generations to get the result I want is a bit too tedious for me. Is there a previous version that doesn't have these issues?

BingusChungus Perhaps you can try version 1.3A, which has less stylization and higher adaptability for Lora. The downside is that it doesn't recognize new characters well enough. Alternatively, version 2.2 is also a good choice, as it appears to be more mature than version 3.0 for now.

Exactly the same problem. Hassaku has a very poor prompts response, plus frequent problems with anatomy.

Ikena On the V3.0, the visual output still exhibits excessive brightness, with darkness-related prompts (e.g., "dark", "dim") often proving ineffective for generating low-light imagery or atmospheric scenes. I'd recommend addressing dark tone generation in your next iteration.

Ikena This link has expired,i can't join the server. Though to be honest, I don't use discord much myself anyway.

Use artist tags or style loras.

Yes you can

If you grab the reforge version of A1111 or use comfyui you'll be all good. It won't be fast, but it won't crash constantly.

bro, i'm using comfyui on my gtx 1050 3gb and it works fine, but you need to be patient. I suggest you to use LoRA, such as byte dance hyperstep LoRA for faster image generation

@BullDream 🙏

@BullDream how?

@kozulff2327 i'm using some of LorA's that speed up generation time without loosing too much detail and using optimized workflow, you can check my images info that i had generated

same altough i understand why style B (->Amanatsu) does not receive updates since Amanatsu is not used even near as much. I end up ussing hassaku with a style lora most of the times anyway.

If it doesn't return does that mean we need to bid for it?

I agree you bro

Yeah, 3 was a mistake, I returned to wai nsfw, it feels like it has less biases and more juicy