Type | |
Stats | 240 |
Reviews | (30) |
Published | Mar 8, 2024 |
Base Model | |
Usage Tips | Clip Skip: 2 |
Hash | AutoV2 C3492DF445 |
Check out my other sibling model used as the base mix for this model: ImmortalEternalMix
Check out my other sibling model for semi-realistic anime illustration style art: MaiMixSR
Check out my other sibling model for anime illustration style art: MaiMix
NarakuYuzuMix is an semi-realistic anime style model which aims to produce high quality stylized art with little prompting at low steps but even greater results with more complex prompts. The NarakuYuzuMix series are the first models released by me that are mixed with other base models rather than just LoRA merges. This is the final proof-of-concept that reveals how after multiple LoRA additions (greater than 300) any turbulence and instability of the base models can be merged/mixed out whilst still maintaining all the relevant detailing and model specifics.
All sample images are first generation/no edit images that have been generated without the use of textual inversion/negative embeddings at 512x768 resolution (no upscaling) with no extra features (e.g., Hires.fix, ControlNet, Adetailer, Inpainting, LoRAs) to display the strength of the model. For real world use it is recommended that EasyNegative (or similar negative embeddings) is used for even higher quality results. Hires.fix and/or Adetailer is recommended for images where the character is more distant for better results.
NarakuYuzu R3V3L is a version using NarakuYuzu LuciVerS as a base model which has been further improved via further merging of models and LoRAs to improve details, and the consistency of good hands and eyes, compared to NarakuYuzu LuciVerS. This model also has improved prompt adherence and improvements regarding non-human subjects. This model is designed to show the absolute possibility of SD 1.5, and was released during the EVE (3V3) of SD 3.0 which would likely change the whole SD industry.
NarakuYuzu LuciVerS is a version using NarakuYuzuHyperD as a base model which has been further improved via further merging of models and LoRAs to improve details, and the consistency of good hands and eyes, compared to NarakuYuzuHyperD.
NarakuYuzuHyperD is a version using EternalMix Version X (based on GhostNotes) as a base model which has been further improved through merging with various other models to produce better backgrounds, subjects, and better details than the previous uploaded models.
NSFW is supported but NSFW LoRAs may be needed for better results.
Recommended Settings:
Sampler: Euler or DPM++ 2M Karras (25-50 steps)
CFG Scale: 4 to 11 (7.5 preferred for balanced outputs)
Clip Skip: 2 for more variation but model works well with 1 as well.
Negatives (if not using EasyNegative): (lowres, worst quality, low quality, monochrome noir, sepia, cropped, normal quality, jpeg artifacts)), AND bad anatomy, AND hideous extra bad hands, AND auxiliary missing malformed fingers, AND unrefined duplicate body parts, AND fused overlap unseparated conjoined, AND ((strabismus, (red eyes), bad eyes), AND fused merged additional arms, deformed, malformed, twisted, mutilated, watermark, signature, username, paintings, sketches, error
Prompt Example (optimal prompt length for best outputs):
cinematic, cinematography, available outdoor light, sharp absurdres aesthetic, waist up half portrait, phantasmagoria, 1girl
Extreme Wildcard Prompt Example (Over-Prompting):
professional|commercial|finished colorful|pristine|dark|saintly reflective|iridescent|crystalline|textured immersive|surreal|phantasmagoria intricate|delicate|exquisite detail floral|snowflake|sunset|outdoors|indoors tattooed fashionista|executive|goth|student|swordswoman|punk|seraph|witch angled|dynamic|focused portrait masterpiece CGI AND attractive|seductive|trendy|beautiful|mature|enticing and womanly|alluring|proportional|motherly|subservient|affectionate|casual|erotic with jewel-like color contacts and ((((salon dyed complex hair wearing a sleeveless|simple low-cut shirt uniform belt jacket)))) AND contemporary|Desi|Arabian|Oriental|European|multinational|cybernetic traditional|elegant|sophisticated fractal|endless|entrancing|infinite|perpetual|abstract mural natural|ambient|moody polychromatic|cinematic|transcendent ancient|opulent|fancy cityscape|dusk fantasy beauty showpiece outline BREAK aesthetic|pleasing|artistic|tasteful official 8K HDR anime|photorealistic oceanic|scenic|picteresque depth|layered|dimensional|spatial landscape|background|theme|scenery|silhouette original film grain polaroid style AND enhanced|featured|complete|final|enriched rainbow|pastel|neon|psychedelic continuous|flowing|curvaceous|diffuse|eternal encased|juxtaposed|overlap underwater bubble jellyfish butterfly bloom concept AND absurdres ultra high quality sharp|extravagant|defined|precise|crisp resolution|clarity magical|cryptic|arcane mystic|majestic|dramatic|action cosmic energy modern|excessivism|analytical perfect|fantastic award|prize winning digital movie|blockbuster game release coverart wallpaper poster illustration image by wlop|WENJR|ufotable with hand-drawn thick|distinct brush|ink|watercolor stroke|impasto edge line shift|shimmering volumetric lighting using Unity|Unreal|Octane Engine|Render trending on Pixiv|Instagram|Artstation
Model Design Aim:
NarakuYuzuMix is designed as a proof-of-concept semi-realistic anime weight merge model using a larger number of LoRAs (>300) and base models (>50) than standard mixes aiming to further train and fine-tune models to produce new model characteristics and improvements over base models. This newer model is also likely to produce better results when checkpoint merged with other models but would likely loose it's own characteristics with more and more merges due to dilution.
Special Thanks To:
The huge LoRA and model creator community that is part of Civitai as more than 300 different LoRAs and 50 different base models (too many to list) were used as part of this model and in various parts of testing.
Disclaimer:
You cannot use NarakuYuzuMix models to deliberately produce nor share illegal or harmful outputs or content
The author claims no rights on the outputs you generate, you are free to use them and are accountable for their use which must not go against the provisions set in the CreativeML Open RAIL-M license