2DN

2DN-ANIME2DN-ANIME v2.0Major improvements:Prompt obediencePrompt adhesion between seedsMore balanced contrast and details [increase CFG scale to increase contrast, if samples look too subtle to you]Better anatomyMore conceptsOneShotFacial technology integrated, which means you don't need to use ADetailer/etcMore detailed backgroundsBetter expressionsMore accurate artist stylesBetter compositionLess latent garblingBetter character-to-character interaction supportFiner texturesMore balanced landscapesBetter LoRA support (will not burn as easily with many LoRA's applied)I recommend an img2img pass at the end with the same settings, sometimes with 50% of the original CFG scale (e.g. to avoid toe multiplication)This is a hybrid v-pred checkpoint so you may need to hunt around for the best CFG scale to use with your image gen software.All but 3 sample images were chosen from the first generation that ran. The 3 needed between 3 and 5 retries.Samples generated with reForge -- you may need to adjust the settings in other software.2DN-ANIME v1.0Due to popular demand, I've decided to upload a "flat" anime version of 2DN, that retains the model knowledge and detail of the semi-real models. It is much easier to use and doesn't require exotic samplers or schedulers and will work with Euler A.This model is actually great for creating game CG assets! So the previews are all wide-screen. If I've stolen your prompt, my apologies, I am a thief who likes your work.Some software (such as the CivitAI generator) like settings like this:CFGSCALE: 5STEPS: 29SAMPLER: EULER AI use reForge with:CFGSCALE:1STEPS:29SAMPLER: Euler Ancestral CFG++SCHEDULER: SGM UniformFor comfyUI steal the settings or workflows from anyone kind enough to post their metadata. comfyUI confuses the heck out of me so unfortunately I can't help there.For reasons unknown to me ADetailer fixes feet really well with this model (as in it actually removes extra toes). I just use the hand model, and most of the time it thinks feet are hands, so it catches them.Please try using the model first without stacking LoRAs like crazy on it. It already has my NAI v-pred fix integrated - so chances are some detail LoRAs will fry the outputs. You can easily prompt for different styles using booru artist tags/combos.2DN-NAI v1.0This is my best general-purpose model! All of the sample images have the same seed! They were not cherry picked!I wanted to make a model that leveraged the sheer creativity and knowledge of NoobAI (NAI) v-pred v1.0 - so after much "hitting head on brick wall" here it is!The culmination of all of the 2DN series model vision - a highly detailed model that lets you do anime, realism, slop, or fancy art, with the ability to produce striking images thanks to v-pred.Sample metadata:masterpiece, very aesthetic, best quality, newest, absurdres, highres, very awa, BREAK 1girl, solo, colorful, glowing, eyes, realistic, bodysuit, safe

Negative prompt: lowres, worst quality, low quality, bad anatomy, bad hands, multiple views, abstract, signature, furry, anthro, 2koma, 4koma, comic, manga, text, watermark, username, artist signature, patreon logo, patreon username, twitter username,

Steps: 14, Sampler: DPM++ SDE CFG++, Schedule type: SGM Uniform, [using reForge]

CFG scale: 2, Seed: 3493447269, Size: 832x1216, Model hash: 565548f2eb,

Denoising strength: 0.2, Hires CFG Scale: 2, Hires upscale: 1.5, Hires steps: 20, Hires upscaler: R-ESRGAN 4x+ Anime6B,

Token merging ratio: 0.5, Token merging ratio hr: 0.5,

ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 17, ADetailer denoising strength: 0.2, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer use inpaint width height: True, ADetailer inpaint width: 1024, ADetailer inpaint height: 1024, ADetailer use separate steps: True, ADetailer steps: 60, ADetailer model 2nd: hand_yolov8n.pt, ADetailer confidence 2nd: 0.3, ADetailer dilate erode 2nd: 4, ADetailer mask blur 2nd: 4, ADetailer denoising strength 2nd: 0.4, ADetailer inpaint only masked 2nd: True, ADetailer inpaint padding 2nd: 32, ADetailer version: 25.3.0

Hierarchical CLIP Within Boost: 1, ["off", doesn't affect the images, I'm working on a fork of reForge with hierarchal prompting which is not used here]2DN-LatentLens v2.0Improvements to model fidelity and knowledge. Difficult to use.Checkpoint note: if you're getting bad anatomy and warped results, turn down the denoising in your hiresfix pipeline and/or reduce the number of steps. Weird faces? Turn down the denoising in ADetailer. The model is designed to give you details, and can be more aggressive than other models. Closeups messing up? Turn off ADetailer altogether!2DN-LatentLens v1.0It uses the right sampler and scheduler settings not available on CivitAI or any other online image generator service. It is also very fast.Introducing 2DN LatentLens.LatentLens is a v-pred model, based on NoobAI XL v-pred v1.0It uses a novel technique called LatentLens, which increases model responsiveness to prompts, coherence and quality. For example, most of the preview images use neither hires fix, face detailers nor upscaling. LatenLens makes models more flexible, allowing you to explore various creative prompting strategies while seeing consistently coherent and vibrant generation results.This model works with NoobAI artist tags (not pictured) and is good at explicit content (not pictured).My intention when working on this model was to "let people's dreams come true easily".For best results:Include in positive prompt:masterpiece, best quality, amazing quality, very aesthetic, absurdres, very awa,Include in negative prompt:lowres, worst quality, low quality, bad anatomy, bad hands, multiple views, abstract, signature, furry, anthro, 2koma, 4koma, comic, manga, text, logo, artist signature, patreon logo, patreon username, twitter usernameRecommended settings:Guidance: 1.1Steps: 20Sampler: DPM++ SDE CFG++Scheduler: SGM UniformSoftware: reForgeThe model is also available on TensorArt2DN-Illust v1.0Introducing 2DN Illustrious.This model uses IllustriousXL/NoobAI as a base. It was fine-tuned using 65,000 high quality photos for 12 epochs, two of which were preconditioning. Note that I specifically avoided training on cosplay photos and selected a very diverse dataset (LAION etc). It was further refined with custom LORAs (such as my cfg scale LoRA).For those of you coming from Pony there are some important things to note:Your Pony LoRAs won't work. This is not as big of an issue as you might imagine. Illustrious/Noob has more than 100x the concepts and tags available for use than Pony. Including obscure tags you can only find on Danbooru. It also has tens of thousands of artist and characters tags if you want to make anime images or mix them into the realism style.You don't need to use the pony score tags. They will only interfere with the images.REALISTIC MODEI recommend the following base prompt:highly detailed, realistic, fine fabric detail, absurdres, highly-detailed, best quality, masterpiece, very aesthetic, <your prompt>For the negative prompt you should almost always use:lowres, worst quality, low quality, bad anatomy, bad hands, multiple views, abstract, signature, furry, anthro, bkub, 2koma, 4koma, comic, manga, sketch, ixy,Inserting artist tags of creators who make highly non-realistic artworks such as ixy and bkub improves the quality of the outputs.Using artist names in the positive prompt will make the model produce anime style images. For a 2.5D look reduce the weight of the artist tags to 0.5 for example.Facial expression tags will result in exaggerated results. Lower the weight of these tags also to control their effect when generating realistic images.Lower CFG scale settings will produce less saturated images with lower contrast.You can control the temperature/colour cast of the images by specifying for example „warm light” in the positive prompt and „blue, blue light” in the negative prompt.If the model produces „babyface” during the ADetailer stage, set ADetailer positive prompt to an anime artist name such as „quasarcake”.ANIME MODESimply prompt as you would any illustrious model, including artist tags.—Highres-fix or upscaling is highly recommended (0.3 denoising strength works well). ADetailer is essential for non close-ups of subjects.Recommended CFG scale is 6.5. Adjust to taste.Training was performed on a RunPod A100 GPU, costing $200USD. The model will be buzz gated for a short while. Your money will go towards offsetting the training costs.2DN-Pony v2.0I'm happy to release v2.0 of the original