SDXL
CognitiveRiver
- Đã tải lên vào
- 15:43 14 thg 5, 2026
- Lượt sử dụng
- 9
- Đánh giá
- Chưa có đánh giá
- Quyền
- Allow image generation & sharing
- Cho phép người dùng tải xuống mô hình của bạn
- Sử dụng thương mại
Mô tả
New GoalI was tempted to start a new model name as my original goal was 4 steps and skin texture. This new model is best around 11-15 steps, but skin detail is still good (though if you are comparing to some FLUX models you will be disappointed). What started to get me about realistic models was the lack of support for non-real prompts; if I wanted to prompt a non-real skin color like blue or purple / and eye colors (like red or purple) I couldn’t. So, my focus this time was realistic, but capable of producing fantasy / sci-fi.I spent a ton of time on this but I'm not charging... so tips are appreciated.IllustriousThis is my foray into Illustrious models. I was so close to finishing a Pony v3 version and maybe I still will, but I don’t see as much interest in it these days (feedback is welcome). Anyways, Illustrious is a hard model to develop a realistic style for. I was originally drawn there because of the light contrast in v-pred... And I tried to create a version but after so many failed attempts I gave up for now. I still think there’s some good lighting control in this version but not as good as v-pred could have gotten. But lots of Illustrious models have either too smooth skin or this weird film grain that looks splotchy so wanted to give it a try.I’ve trained this model from the ground up, training on about 1500 images which I manually tagged. My aim is to give a lot of control words to users... I’m still trying to find the right balance of tokens vs natural prompts and padding, so tags might not always work. It is also a DMD2 model, so trade off on CFG guidance vs Steps. I definitely will be tinkering with this in the future on newer releases.THIS IS A HIGH RESOLUTION MODEL. It is trained on 1280 pixel images.SettingsVAE: VAE is now baked in.Sampling Methods: Euler A gives best results this time around, wanted to keep it simple. LCM can work too but it tends to smooth out details like skin and textures, but you might want this. One trick is to use LCM as main sampler and then Euler A for Hires to add a bit more detail back.Schedule Type: Karras works best, Align Your Steps and Turbo works pretty good too.Steps: As mentioned above, I balanced this around 11-15. If you want to get to 4-5 range, you can mix in lora:dmd2_sdxl_4step_lora:0.4. However, there are some sacrifices in details and your ability to use much range on CFG (which is already limited).CFG Scale: Between 1 - 2. The lower the number, the more neutral and realistic the colors, higher is a little bit more vibrant. But let me warn you, if you use a low number it hinders your ability to use a negative prompt if that is important to you. At 1, negative prompt is disabled, at 1.5 I had a hard time getting the negative prompts to be consistent. I found at 2 that it actually starts to listen. If you use the 4-5 range above, you’ll be limited to 1 - 1.25 (again limiting negative prompt).Resolutions: This is a high resolution model, so it’s best at scale 1280, however, you can absolutely still use it at 1024.1280x12801152x14081024x1536896x1664Hi-Res. fix: Because the resolution is already high, I would limit your Upscale By ratio to 1.25. You can do 1.5 if you want to but it will be slow. I use 4x_NMKD-Superscale-SP_178000_G as my upscaler of choice. Steps match sampling. Denoising at 0.25. Pretty standard.TAG GUIDEThis guide is a reference, it does not mean the model will follow it all the time and it is not comprehensive as Illustrious itself has tons of tags.QUALITYYou don’t need all the masterpiece, very awa, year 5000 filler tags. Use if you want, but it’s a waste of a token if you ask me. If you want you can use anything with the word quality / resolution, like best quality, high quality, amazing quality, high resolution, or bad / low quality in negative. They are not needed, the quality should be pretty good out of the box.CAMERAZoom:Closeup / Portrait, Medium Shot, Cowboy Shot, Full ShotThis one frustrates me and will try to keep improving, I had more tags but they are very inconsistent.Angle / Orientation:High Angle / From Above Low Angle / From Below Side View Back View / From Behind Upside Down (head is closes to camera than body) Titled Angle, Dutch Angle (Had this working and now it’s not)Focus: Say the part of the body you want to focus on for a closeup. Examples:Feet Focus, Face Focus, Back Focus, Breast Focus, Butt/Ass Focus... You get the idea.STYLERealistic Style, Anime StyleBase Style (don’t have to add anything): Just my own personal style I like and is defaultRealistic Style: More photo like, more skin blemishes, skin detail, less cleft chins, more variety in ethnicityAnime Style: Smoothed out skin, larger eyes, variety in hair colors and styles and facial expressionsCosplayThis is still experimental, I had it working so well in my not released Pony 3 model. Sigh. Anyways, if you add the tag cosplay, it will take the character and have someone look like they are dressing up as that person, rather than it look just like them. Still working on this, more successful if you use the realistic styleNo PeopleI’m realizing I didn’t do a great job of this so this is one area for improvement, but if you want to get rid of people showing up, here are a few tips:Use the subject you want and then photography. Examples:Food Photography, Landscape Photography, Macro Photography, Product Photography, Wildlife PhotographyNo person, No PeopleLightingThey don’t all work all the time but give them a try:dim lighting Dark lighting / room studio lighting natural lighting soft lighting candlelit lighting indoor lighting outside lighting bright lighting moody lighting dramatic lighting stage lighting [color] lighting / filter. Best paired with Dim Lighting and Dark Room / Lighting BODYSkin Color: fair skin, light skin, pale skin, olive skin, tan skin, medium skin, dark skin, black skin, white skin, grey skin, red skin, orange skin, yellow skin, green skin, blue skin, purple skinSizes / Shapes:Thin, Petite, Slender, Slim, Curvy, Chubby, Fat, Athletic, Muscular, Thick Thighs, Thigh GapBreast Slider: flat chest, small boobs, perky breasts, medium boobs, large boobs, huge boobs. Can use boobs or breastsButt Slider:small butt, large / round butt, big / huge buttFacial Control:High Cheekbones Sharp Face Round Face Long Face Cute Nose (gives more of a rounded nose) Parted Lips Open Mouth Full Lips (puffy lip look) Tongue Out Eyes ClosedMakeup / Markings:Eyeliner Blush / Blushing [color] lipstick Lipgloss (shiny) Tattoo / Tattoo on [body part] FrecklesEye Color:Blue Eyes Green Eyes Red Eyes Pink Eyes Purple Eyes Orange Eyes Yellow Eyes Black Eyes Brown Eyes Grey Eyes Amber Eyes Hazel Eyes Heterochromia Glowing EyesExpressions:Happy / Smiling Sad Neutral Expression Serious Expression Sexy Expression Playful Expression Angry Surprised Expression / Shocked ExpressionPoses:Standing, Sitting, Kneeling, Squatting, Doggy Pose, Laying on Stomach, Laying on Side, Arm Support (leaning on an arm), Leaning Back, Handstand, Leg Split, Leg(s) Up (Leg if you just want one, Legs if you want both), Yoga Pose, Bent Over, Legs Spread, m-pose (laying back and legs are spread with knees up), Bent KneeAnthro (Furry) / CreaturesIt is really hard to get a model to do either a real creature, or a furry. I think I have it 80% there. Pony was really good at Anthro, but terrible at creatures. If you want a boy version, it is better to use the tag male, but I didn’t train much on male furries, sorry.[Creature] = Just creature
[Creature] Girl / Woman = A woman dressed up as that creature.
[Creature] Girl / Woman, anthro / anthropomorphic = Changing into that creature (mileage may vary and could be a complete furry)
[Creature] Girl / Woman, furry = Well... A FurryI trained on lots of creatures but it’s not as robust as I’d like, robust meaning, be able to extend to things you haven’t trained (Pony was really good at this), so your mileage may vary.Also trained on non-real creatures:Slime Girl, Tieflings, Elves (including dark elves, and night elves), Demons / Succubus, Fairies, Giantess, Monsters, Goblins, Orcs, Dwarves, Lamia (half snake, but pair with half snake half girl for better results), Oni, Robots, probably missing some...PONYGoalI really enjoy the SDXL Lightening models, for their speed and image quality. However, I didn’t like the fact that SDXL seem to smooth out the skin to where it seemed too “perfect”, a little too plastic for my liking. One of the best workarounds is to generate in SDXL and then use ADetailer with a 1.5 model for the face. My goal was to see if I could get close to generating faces without that extra step or at least not switch models. This model was developed for photography only, not for anime or art, there are better models out there for those.PONY VERSION 2I feel like I relied too much on hires fix and did not give enough focus to regular image quality in the last release. I’ve fixed that this time around; you’ll see the first several image examples are normal quality. By fixing this, the high quality (hi-res fix) images turn out even better than before (some examples there too). Also with this release, I relaxed the settings so as to allow a greater degree of freedom. Lots of good Pony models out there, but they don’t seem to be TURBO models, this gives you the choice between fast (4 steps) and good quality, or regular speed (15+ steps) for great quality. You don’t even need a long prompt or all the score prompts, first image is as basic a prompt as you can get.SettingsVAE: Baked inSampling Methods: LCM works best. Euler A can work in a pinch now at certain settings but, it’s not as good as LCM, as the skin texture looks a little fake but feel free to try (hi-res fix seems required). Stay away from DPM.Schedule Type: A lot more options, at 4 steps, DDIM or Karras. If you do more than 10 steps, you can also use Beta and Normal, 15 or more can also add SGM Uniform.Steps: From 4 steps on. If you want more details without turning hires fix on, feel free to increase this.DDIM / 4+Karras / 4+SGM Uniform / 15+Beta / 10+Normal / 10+CFG Scale: 1 - 1.5. I actually like 1.25 the best, but feel free to find what works best for your style.Other: Everything els