SpiritMix - Soft and Cutesy (2.5D Anime)
Creator Website -
https://www.michaelpstanich.com/
Creator HuggingFace -
https://huggingface.co/michaelpstanich
Creator Donations -
https://subscribestar.adult/michaelpstanich
Original Release - 2023-14-7
< ----- ----- ----- >
Welcome to the world of SpiritMix and my first openly public blend! This model was an experiment to try and get a more "3D" looking anime model while retaining a bright and soft look. The goal was to keep those anime proportions and faces, along with those expressive faces, while still offering detailed environments and outfits. Honestly it turned out much better than expected, but struggles in a couple areas I'll cover later. Please share your results with me!
< ----- ----- ----- >
Model Name :
SpiritMix - Soft and Cutesy (2.5D Anime)
Experimental Name :
_SpiritMerge1_1-7-4-2_.safetensors
(My gens may have the experimental name in the data, but it's the same model, just a different name)
Base Model/Type
SD1.5 / AI Image Generation
Focus :
Soft 2.5D Anime Style with a soft 3D look
SFW / NSFW :
Does SFW and soft NSFW. Some SFW prompts can still show NSFW so I wouldn't consider this a 'safe for public use' model.
Bias :
- Heavy female human bias
- Tends to make faces look younger (subjectively looks young adult-teen looking in most cases)
- Likes to generate Lewd leaning images with certain prompts, even SFW prompts (silly SD, being horny >.>)
< ----- ----- ----- >
So what is SpiritMix - Soft and Cutesy (2.5D Anime)?
This Checkpoint Merge is a blend of various SD1.5 models which tries to create a soft and bright presentation without giving up too much detail and keeping that "3D" look of 2.5D anime. I consider this mix to be experimental but it's turned out quite usable so why not share? This specific merge was part of a merge tree of over 40-ish merges so the specific recipe is kinda lost, but contains various anime models, specifically starting with 'simple 2D' and 'flat and bright' models then merging with '3D / Realistic' leaning models. The mix pool contains both Public and Private models, all of which are mixes themselves.
Suggested Settings :
- Add "best quality" and "highly detailed" to positive prompt, avoid using "realistic" and "masterpiece" in your negative or positive prompt.
- Avoid ancestral samplers like Eular A, the random noise seems to cause issues and images don't resolve well.
- High-ish step count (I typically used 34 with DPM++ SDE Karras), helps to resolve small artifacts and issues.
- Image size of 512x512, 768x768, or 512x768, really struggles on wider images, deforming human arms/legs more frequently
- HiRes Fix, simple upscalers work well but requires high steps, latent tends to try adding non-existing detail and breaks figures (I mainly Used R-ESRGAN 4x+ at 0.36)
- Clip-Skip 2, tends to produce better results but isn't required
- VAE with soft color grading
What this mix does well :
- Soft and Cutesy Female Characters
- Young Adult / Teen-like Characters
- Cute Anime Faces
- Bright environments with simple structures
- Clothes, apparently
- Breasts X.x
What this mix struggles with :
- Hands can be a major struggle, higher resolutions and ControlNet helps
- Some body shapes don't form well and some body prompts are often ignored (For example, if you include Large Thighs and Flat Chest, it will prioritize Large and include Large Breasts instead, no matter the weights. The data sets most likely didn't include some body types so the tokens don't exist for them)
- Doesn't usually do well with Hardcore NSFW stuff, can do it but often messes up anatomy of certain body parts with missing details.
Other quirks / issues :
- Sometimes characters are cross-eyed when the head is lower res, kinda silly looking.
- As mentioned in Bias, some prompts will still generate lewd/erotic leaning results even with NSFW, nudity, ect in the negatives. Be VERY careful when including certain prompts with characters of certain sizes/contexts.
- In dark scenes characters are still often bright with a soft and almost flat look to them.
