Discord - The Broken Chatbox - https://discord.gg/h3vB7S4FEw
HuggingFace - https://huggingface.co/michaelpstanich
-v- Donations! -v- =^-^=
SubscribeStar - https://subscribestar.*****/michaelpstanich
Ko-fi - https://ko-fi.com/michaelpstanich
< ----- ----- ----- >
SpiritMix returns with SpiritMix – Poly Cube, a 3D anime checkpoint ready to deliver the cuteness! This checkpoint was a challenge for me to create as it blends various styles of checkpoints together to create a unique looking 3D anime checkpoint with a wide range of concepts and variety. The goal was to create a 3D anime checkpoint with the variety seen in anime models like SpiritMix -Rival and its flexibility, but with the vibrancy and smooth 3D look of a western animation checkpoint style. Honestly, this was one of the first checkpoints I wanted to create but lacked the skill to do so, SpiritMix – Soft and Cutesy was my first experimental test of this concept so I guess you could consider this a spiritual successor? It was also a massive pain to create, and was only possible because of all the research and learning I’ve done since my first public model release! So far I’ve been blown away by this model, and I hope you are too! Happy Genning! (Don’t forget to post and share your gens! =^-^=)
< ----- ----- ----- >
Model Name :
SpiritMix – Poly Cube (3D Anime).safetensors [hash:91ee133c7d]
Experimental Name :
_SpiritMix6_1_2_0_2_3_1_1-1_7_1.safetensors [hash:e9ab14e28f]
(Some of my gens may have the experimental name/hash in the data, however it's the same model just without the baked VAE)
Base Model :
SD1.5 (Compatible with SD1.5 models and tools, may work with some SD2.x models)
Focus :
Create a cute and adorable 3D anime style that is both flexible and complex, focused on female characters.
R18+ Content? :
This model was primarily designed around safe and cute content and doesn’t lean R18+, however the model does include merges that add R18+ content so the model is not considered safe for public use.
Bias :
Heavy Bias toward cute/young female characters
Expressions are typically quite drastic/exaggerated
Prefers bright and vibrant colors
< ----- ----- ----- >
Quick R18+ Warning :
This model may contain R18+ data, as such, you should be careful with your prompts. Always include clothing keywords, specify for more safe outputs, and avoid lewd/erotic keywords when creating certain subjects.
< ----- ----- ----- >
So what is SpiritMix – Poly Cube (3D Anime)?
SpiritMix – Poly Cube (3D Anime) is a merge project with the goal of making a “3D Anime” checkpoint which has the flexibility and style you’d expect from an anime checkpoint but also has the look and vibrancy of something like a 3D western animation checkpoint with a rounded soft look to it. Poly Cube, which started as a project to create “rev animated but as fully bright anime” is the culmination of all my merge and block research since even before my first public checkpoint, including models of all types trying to capture as many concepts as possible. The method was to carefully balance anime, realistic, 3D animation, and toon style checkpoints by creating individual merge chains with traits/concepts I wanted to include in the final merge, then merging those into the final merge chain as the merge went on, then going back and re-adjusting weights to increase very small consistencies. One of the largest hurdles for this style of merge is not letting a single style dominate the entire checkpoint and getting concepts to converge into one another. To help with this, micro-merges were deployed and tested to see how blocks behaved when combined with troublesome checkpoints. As a personal note, faces, my goodness were faces such a massive pain! Faces constantly broke and each style wanted to conflict, so if you intend to try this style of merging I highly recommend learning how blocks and merges affect faces!
< ----- ----- ----- >
Prompting Assistance and Usage :
Recommended Settings :
Sampler : DPM++ 2M SDE Exponential or DPM++ 2M SDE Karras recommended
Size : 512x512, 512x768, 768x512, 768x768, 768x1024, 1024x768 (additional upscale recommended)
Steps : High if you want smooth edges, I used 40-50, gens well enough at 30 but corruption can happen.
Hires fix / Latent upscaling : (Safe) R-ESRGAN 4x+ with 0.36-0.45 denoise / (unsafe) Latent-Nearest or Latent-Bicubic with 0.62 / 0.66 respectively
CFG : 7-12, typical stay around 8-9 for most concepts
Clip-Skip : 2 recommended but not required
Recommended Quality Prompt :
Positive Prompt : best quality, highly detailed [optional : detailed background, depth of field, blurry background]
Negative Prompt : (low quality, worst quality:1.2) [optional to avoid watermarks/signitures : watermark, logo]
Prompting :
Prompting is actually pretty simple, quality tags will ensure the model stays with the 3D Anime look (include ‘best quality, highly detailed’ in positive prompt ‘low quality, worse quality’ in negatives). SpiritMix - Poly Cube follows prompts relatively well for an SD1.5 model and should work well with both simple and complex prompts, however complex prompts can start to conflict with objects and cause minor artifacts. Avoid using style keywords to try and alter the 3D style, ironically, including ‘3d’ and other similar keywords lowers the general output. Using keywords like ‘vector art’ can simplify the texturing on characters/objects in interesting ways, ‘drawing’ keywords will give characters outlines and start to change the overall style.
Upscaling :
SpiritMix – Poly Cube stays relatively sharp even with latent upscaling, so I recommend switching between my safe and unsafe recommendations. Personally been rather impressed with the latent results but you can use whatever for upscaling and the checkpoint should keep the style consistent.
NSFW / R18+ :
While SpiritMix – Poly Cube wasn’t designed with R18+ content in mind, and was not extensively tested in this area, the checkpoint does contain R18+ data and can do many R18+ concepts with relative ease. However, you need to be careful with any R18+ prompts you try to create! This checkpoint leans cute, and often interpreted as young, characters, so when making R18+ prompts include keywords that define the body and shape well to avoid surprises! While I never ran into any ‘cute’ r18+ when testing, if you start playing with fire you may get burned, I recommend prompts like ‘large breasts, thick thighs’ and other keywords that would alter the character’s appearance to be older while avoiding R18+ prompts when generating cute characters (including ‘nsfw, explicit, nude’ in negatives can help here if want to avoid any posibility). This worked well in my testing, but that testing was very limited, so careful!
Colors / VAE :
Different VAEs seem to have a rather noticeable effect on Poly Cube, so I recommend starting with the baked in VAE, but don’t be afraid to try any SD1.5 VAE you have and see if you like it. People seemed to prefer the more vibrant/over-saturated checkpoints in testing but you can try a more muted VAE for a softer look similar to my SpiritMix - Soft and Cutesy checkpoint.
< ----- ----- ----- >
What this mix does well :
Generates a soft 3D Anime look that sits between animation and anime really well
Consistently clean output, regardless of upscaling
Does exaggerated emotions really well (probably from both anime and 3D checkpoints working together well)
Creates very cute looking characters and concepts “sugar cube” was a fun one to play with!
Responds really well to basic prompts, producing a wide range of images from just a few words
Very flexible and responds to body keywords well
What this mix struggles with (aside from typical AI issues) :
Some concepts may lean more 3D animation rather than 3D anime, others the opposite (from lack of data in those areas)
Style keywords can alter the checkpoint’s output in undesired ways
Distant faces without upscaling can look a bit mushy and inconsistent (seems to pick up the anime faults rather easily)
Has some anime-specific quirks I couldn’t completely remove
Low-Res gens can fail to give the full 3D look depending on sampler (use the recommended samplers to avoid this issue)
Other quirks or issues :
Style tags often have strange affect on generations
Leans toward generating cute characters and concepts (use targeted prompts to guide the gen otherwise)
< ----- ----- ----- >
Hey there, so like, this stuff takes a lot of work and I have many other projects, if you'd like to support my work then why not consider a donation? You can send me donations either on SubscribeStar or Ko-Fi! More donations means I can do more models and update existing ones! (You can also follow for my various other projects if you’d like!)
https://subscribestar.*****/michaelpstanich
https://ko-fi.com/michaelpstanich
If you have any suggestions, feedback, or ideas you want to throw my way, head over to our community Discord, The Broken Chatbox! We also have a channel to share your generations!
https://discord.gg/h3vB7S4FEw