Type | |
Stats | 982 |
Reviews | (226) |
Published | Apr 19, 2023 |
Base Model | |
Hash | AutoV2 0FD16B71D2 |
DARK GEMINI Version 3.0
(I recommend using Waifu Diffusion 1.4 VAE kl-anime 2.)
(Not far down you'll find information regarding multidiffusion upscaling etc which is the final application to achieve the results my preview images have)
Dark Gemini was originally conceived as a dark fantasy gothic model but has since expanded into sci-fi, sci-fantasy, dieselpunk, cyberpunk, apocalypsepunk and a variety of genres both dark and neutral or in rare occasion also light. The focus is still on making sure the darker variety work before checking to see what else works and I'm happy to say that right now this model is incredibly versatile. Check example images to see the variety done using this model and kl-anime 2 VAE and sometimes the bad_hands embedding to help out those spaghetti fingers.
ADDED PROMPTS TO MOST OF THE PREVIEW IMAGES. THESE PROMPTS DO NOT INCLUDE THE FINAL STAGE WHICH IS IMG2IMG UPSCALE USING MULTI-DIFFUSION. UNFORTUNATELY CIVITAI HAS NO WAY OF SHOWING THIS STEP. I WILL FIND THE BEST METHOD OF POSTING THE NECESSARY INFORMATION TO YOU ALL. UNTIL THEN I CAN AT LEAST PROVIDE A LINK TO THE EXTENSION AND MY SETTINGS + A PROMPT EXAMPLE
https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111
STANDARD IMAGE SETTINGS
Sampler: DPM++ 2S a
Steps: 30
CFG scale: 7
Denoising strength: 0.4
TILED DIFFUSION
Tiled Diffusion upscaler: R-ESRGAN 4x+ Anime6B
Tiled Diffusion scale factor: 2.5
Tiled Diffusion:
Method': 'MultiDiffusion'
Latent tile width': 96
Latent tile height': 96
'Overlap': 32,
'Tile batch size': 1
TILED VAE
Encoder Tile Size: 1024
Decoder Tile Size: 128
Fast Encoder: Checked
Fast Decoder: Checked
Encoder Color Fix: Checked
The prompt can help elevate or hide elements of your image.
Example prompt: "Apocalyptic Art Style, Pseudorealism, Gigapixel Image, splash art style, Beauty Filter, cinematic lighting, digital illustration, digital art, 2.5d, masterpiece, best quality, highres, HDR, unreal engine 5"
Digital Illustration, Digital Art, Masterpiece, Best Quality and Highres are standard just to make sure it knows to keep them in the digital style.
Apocalyptic Art Style (without weights) to continue the theme.
Splash Art Style: Same as above
Pseudorealism + 2.5D They act synergistic to nudge the image from pure line art into a bit more of a hybrid, making the art less painted and more digital.
Beauty Filter, Cinematic Lighting, HDR: With beauty filter being the main addition to the prompt to add "beauty" to the subject with improved lighting using cinematic lighting and HDR.
Gigapixel Image, Unreal Engine 5: Gigapixel Image (while I'm unsure if it works) is a plugin and software used in high end upscaling, hypothetically by prompting it you add a bit of the style or flavor that the finished product of a Gigapixel Image receives. Giving it a 'high end' feel to it. While Unreal Engine 5 will help bring it all together into a common aesthetic, which won't dominate but influence the upscaled image.
These are my thoughts and motivations to why I included those in one upscale prompt.
----------------------------------------------------------------------------------
A Small Intro
GEMINI V3 and previous iterations are models generating art styles from anime to more western style splash art or If you want to render 2.5d (3d rendered in 2d, one way to accomplish this is to write: “line art, sketch art” in negative prompt.) After discovering the CLIP was off, I reset it and released Gemini 2.1 while keeping 2.0 as well for its different style. For 3.0 the Clip has been solid the whole way through and both of these guides apply to Dark Gemini 3.0 and will give you an understanding on how this model works and its quirks.
Aside from the more complex or convoluted prompts shown here, the much simpler ones with danbooru tags also work just as well but if you want to create very intricate, detailed and involved images then as long as you develop an understanding with this model. You won't be punished for trying.
I'll explain a few things on how this model responds to prompts:
If your prompt contains a lot of different details that your character is doing.
Try to first condense it, make it more simple but precise. If your prompt still ends up being very long.
Then make sure the prompt loops at least once, this will make the AI take another pass to read through it.
I don't know if repeating the input for the prompt you want the most has any impact,
but if it does then I don't see there being any other way than going beyond 75 tokens and letting it do a second pass.
Example style prompts:
Nu-Gothic Art Style
Victorian-Gothic Art Style
Splash Art Style
Anime Art Style
Dark Anime Art Style
Dark Fantasy Art Style
Norse Art Style
V3
Disaster Art
Dieselpunk
Cyberpunk
Gearpunk
Clockpunk
Apocalypse Art
Art Style can be switched out with Artwork, which can alter the results a tiny bit giving you the same aesthetic but different renders. Art can also be removed from the prompt to create “X Style Infusion” or “X Style Aesthetic”.
----------------------------------------------------------------------------------
THIS SECTION APPLIES TO 3.0 and 2.1
In 2.1 after the clip has been reset, many prompts dislike when you place style or design tags at the start of the prompt. It should instead be moved to the final stretch when all art, style, color, design, composition, light, camera etc is written. Since we'll be using weights, the effect of the location of the prompt has less impact than in 2.0.
If you want to render 2.5d (3d rendered in 2d, one way to accomplish this is to write: “line art, sketch art” in negative prompt.)
For example:
To get two characters to fight and actually make contact, actually stab or slash. It helps to add tags such as: full contact, close quarter, stabbed, loses limb, blood gushing, wounded.
These can be placed sparsely around the entire prompt and experiment with their weights until you get something you want. Of course it also matters how else you write the prompt so I will give you an example and you can trial and error from there!
((extremely detailed:1.22) two lethal swordsmen dueling to the death in the phantom ruins:1.22), (loses limb:1.08), (blood gushing:1.1), (Facing each other:1.06), (Full Contact:1.04), (Stabbed:1.04), (One has short blonde hair and one has long dark red hair:1.20), (each of the two is wearing unique enchanted (ultra detailed:1.22) legendary armor:1.22), (wounded:1.14), high intensity, (swords daggers and magic:1.17), (magic circles:1.10), (2021 anime art style:1.25), (digital illustration:1.20), (digital art:1.20), (splash art style aesthetic:1.15), (dramatic angle:1.15), (masterpiece:1.25), (best quality:1.25), (highres:1.25), 8k, (intricately detailed:1.22), (exquisite line art:1.16)
This is accompanied by a negative prompt modified from the general one I use (which still applies to 2.1.
Down below is some more information if you want to replicate the image I rendered.
Steps: 30, Sampler: DPM++ 2S a, CFG scale: 7, Seed: 2930654821, Size: 1024x640, Model hash: 49fc1bfcac, Model: Dark_Gemini_v2.1, Clip skip: 2
----------------------------------------------------------------------------------
THIS SECTION APPLIES TO 3.0 and 2.0
Placing one or two style reference tags at the start of the prompt will focus more on achieving that look than it will on your active prompt of your choice. This however happens a lot less when the tags are much simpler.
For example
Nu-Gothic Art Style Vs Portrait Of
Nu-Gothic Art Style or Nu-Gothic Artwork and similar prompts will have a much greater impact on the overall prompt when placed first.
Rather than in the final composition phase of the prompt. When using Portrait Of, it still won't know what style.
So don't forget to let it know during the final phase of the prompt, or you might get some strange and very poor results.
Here's three examples of the prompt layout I used when creating the example images. Replace XCONTENTX with your typical prompt content be it a character, an item, a landscape etc. Then XCOMPOSITIONX are style, design and command prompts that affect how the prompt will look like, beyond what's already written (these three weren't the only ones I used, but should serve as examples to get you started!)
"(X art style1.30), (X style infusion:1.20), (digital illustration:1.25), XCONTENTX, (digital art:1.20), (X artwork:1.20), XCOMPOSITIONX, (masterpiece:1.35), (best quality:1.35), (highres:1.35), 8k"
"(X artwork:1.30), (X Aesthetic:1.20), (Concept Art:1.20), XCONTENTX, (digital art:1.20), XCOMPOSITIONX, (masterpiece:1.35), (best quality:1.35), (highres:1.35), (8k ornate octane render live 2.5d:1.31)"
"((Portrait of:1.38) (Character Description:1.39):1.45), XCONTENTX, (X Style Infusion:1.35), (X Artwork:1.40), (Extremely Fine Line Art:1.20), (Intricately Detailed:1.30), XCOMPOSITIONX, (masterpiece:1.35), (best quality:1.35), (highres:1.35), 8k"
You can also look on some of the images that come with prompt information and use it for testing or as a guide. Eta Noise Delta is always 0, 1 or 31337. I always use kl-f8-anime2 for my VAE.
----------------------------------------------------------------------------------
THIS APPLIES TO EITHER VERSION:
There are other purely experimental tags that can affect your prompt:
Clockpunk
Gearpunk
Diesel Punk
Cyberpunk
Darkcore
Futurecore
Ethercore
Keep in mind that the model is mostly tested on Gothic and Dark Fantasy themes in a range from anime to western style art or the hybridization within. It's also been tried and evaluated by a handful of beta testers that are themselves model makers and given the thumbs up to exit beta stage. They tested it for waifus, husbandos, chibis, animals, famous characters (it failed a lot on that, so might need an embedding) as well as it's intended use. Myself and the testers also tried cyberpunk, futurepunk, dieselpunk, steampunk etc with some promising results but as that isn't the primary focus of this model I can make no guarantees to your success. However I'd love to see your successes!
It’s been tested using the kl-8-anime2 VAE and Skip Clip 2 with delta noise of 0 or 31337, ETA Noise for Ancestral and DDIM is 0.67 on 2.0 and 1 on 2.1
The negative prompt that has served as a foundation for most of my prompts in 2.0 and 2.1 is the following:
low quality, no quality, bad quality, low resolution, lowres, normal resolution, no detail, low detail, deformation, text, title, Description, logo, watermark, jpg artifacts, jpeg artifacts,
disfigured, deformed, bad anatomy, extra leg, extra arm, extra finger, no head, no body, half body, melted face, cropped, cropped head, cropped body, missing finger, 6 fingers, 7 fingers, 4 fingers, 3 fingers, extra thumbs, no thumb, noodle fingers, contorted fingers, spaghetti fingers, wavy fingers, extra long fingers,
spaghetti body, contorted, contortion, siamese twins, conjoined twins, conjoined body parts, unnatural pose,
big breast, massive breast, major breast, big tits, massive tits, major tits, hentai, eroge, nukige, pornographic, porn, XXX, erotica, loli, lolicon, skimpy, slutty, NSFW, nude, nudity, naked
----------------------------------------------------------------------------------
LICENSING, MONETIZATION ETC
The model uses the CreativeML Openrail with some modifications.
I permit personal (where you do not receive any monetary return from your use of the model, whether they be voluntary donations or mandatory fees) and legal use of the model but I'm not responsible or liable for anyone's use or actions of and with the model.
I permit free merges of my model. I don't want to impede any creative works, so long as they're not commercial and allow all users free access (and donation free) to the models made.
Any personal commercial use of Dark Gemini (all models and derivatives or merges) is prohibited without the express and explicit permission from myself.
Personal commercial use is defined but not limited as commissions, donation based ventures (anything where the model is used and people can voluntarily pay, or through services such as patreon.)
Any business venture commercial use of Dark Gemini (all models and derivatives or merges) is prohibited without the express and explicit permission from myself.
business ventures refer to websites serving to generate images with a paid business model, through for example but not limited to credits, monthly subscriptions etc
Any non-commercial generative services using Dark Gemini (all models and derivatives or merges) to provide any form of generation without the express and explicit permission from myself.
Sale of Dark Gemini (all models and derivatives or merges) is prohibited without the express and explicit permission from myself.
You can contact on [email protected] if you have any questions about commercial or business inquiries.
//Cryonicus