The world of SpiritMix has opened once again! For this entry we have a bit of a weird one, SpiritMix - Rival! Rival is a spin-off of SpiritMix - Deep Line which starts with the same base but alters the weights and adds some along the way to create a checkpoint with similar concepts but has the opposite goals and visual style. Unlike Deep Line, this merge focuses on supporting complex prompts, complex upscaling, and giving a more 'general' detailed anime vibe the AI generation scene has grown use to. Rival unfortunately suffers from some common issues I'm not a fan of so I do plan to revisit this style of Checkpoint in the future (most likely updating this one), but even still, this checkpoint offers some uniqueness into my currently released Mix list. Make sure to share your results!
Note 1 : I'm releasing this checkpoint as is to see what people think of it, however I'm personally still not happy with how it turned out so I will be returning to release either an update or newer version when I get the mix done right.
Note 2 : Since this model relies heavily on upscaling to produce good results it may be more difficult to use than other checkpoints.
Model Name :
SpiritMix - Rival (2D Anime) [hash:b77916be18]
Experimental Name :
_SpiritMerge4_0_5_0_4_0_3.safetensors [hash:c0a58902f3]
(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 an opposite to SpiritMix - Deep Line focused on complex upscaling potential and lots of macro detail.
SFW / NSFW :
This model does well with SFW content but can lean lewd within certain prompts and contexts, does nudity well and NSFW works with the right prompts but may be inconsistent. Not considered safe to use in public!
Bias :
Tends to make characters a bit tall, often favors larger body sizes
NSFW tends to overexaggerate quite a bit
Many other 'typical anime' bias in terms of styling and body form/proportions/anatomy
Defaults to lighter skin tones, doesn't always obey skin color related prompts
< ----- ----- ----- >
Quick NSFW Warning :
This model will often slide into more lewd territory with certain keywords. Ensure you're careful when designing your prompts, include nudity and nsfw in the negatives if you're strictly looking for sfw content and include the appropriate body prompts when making nsfw content.
< ----- ----- ----- >
So what is SpiritMix - Rival (2D Anime)?
Unlike my other mixes which start with a concept and have their own unique recipe, the current release of Rival is an abandoned branch of SpiritMix - Deep Line. The idea was to take the same recipe, modify the weights, and change a few checkpoints in the mix to create the 'opposite' model. While mixing I decided to try and make the model designed to be used with complex upscaling technics with a lot of potential for macro detail. This was accomplished by adding painterly and detailed models while removing and lowering the weight on more stylized checkpoints. I plan to return to this mix in the future with a completely unique recipe in the future to accomplish these goals without the compromizes seen in this initial release.
< ----- ----- ----- >
!!! IMPORTANT !!!
Prompting Assistance and Usage Information :
Recommended Settings :
Sampler : DPM++ 2M SDE Exponential or DPM++ 2M SDE Karras recommended
Size : 512x512, 512x768, 768x512, 768x768 (additional upscale strongly recommended)
Steps : Fairly high, 40+40 for basic gens, 50+50 for complex upscaling at least
Hires fix / Latent upscale : (Safe) R-ESRGAN 4x+ with 0.34-0.36 Denoise (For complex upscales read upscaling section)
CFG Scale : 7-12, typically stay around 8.5-9 for most gens
Clip-Skip : 2 recommended but not required
Prompts :
You can use this model however you choose, however I specifically tried to mix with latent upscaling in mind. This does mean the results are somewhat inconsistent, but with the right prompt and some good post work the results can be quite pleasing.
Prompting should include quality tags, the set I used was "masterpiece, high quality, highly detailed, detailed background" though you can remove masterpiece to reduce the painterly look of the model. high quality and best quality tended to have the same effect, but I didn't test it enough to see if Best Quality is actually better or not.
When prompting body sizes you may need to emphasize certain keywords to have the intended effect, especially if you're targeting smaller characters (the model prefers tall, thick woman in most cases)
When creating smaller characters emphasize "girl" and don't include "female" as female weights a larger size.
Upscaling :
This is where things get tricky and require some fiddling, but I'll try and set some guidelines.
For "safe" upscaling to reproduce the simpler style of the base model, use R-ESRGAN 4x+ with 0.34-0.36 denoise. This will keep your gen looking very similar to the original while cleaning up artifacts and adding details to unresolved areas.
However, if you want a complex and detailed image you'll want to play with either Latent upscalers or other complex custom upscaler models. In Automatic1111 I used hires fix with the latent at 0.55-0.62 denoise or latent (nearest) at 0.62 denoise for the best results. This upscale is very heavy on RNG, especially for hands, and you may need to switch between them and alter your denoise and adding/removing a single step from the upscale to try and improve the upscale. Trial and error is the name of the game here but you can also employ post processing methods like inpainting. (ControlNet helps to a certain degree, you can make the base gen then use that gen into controlNet to then run the gen again to improve your results, however depending on the artifacts this may not always help)
Colors / VAE :
When using certain style tags the baked in vae may oversaturate or overexpose the image, to fix this you can force another VAE with softer colors or use a lora to reduce brightness of the gen itself.
What this mix does well :
Creates a lot of detail and allows latent upscalers to run rampent with details
Maintains an anime aesthetic through any concept thrown at it
Allows a wide range of 'moods' and 'tones' through color grading and variance
Supports stylized keywords like "retro" for retro anime styles and coloring
Most common loras work really well
Responds really well to control-style tools like ControlNet or ADetailer
What this mix struggles with (aside from typical AI issues) :
Default gens without latent upscale may look a bit melty or unrefined at times
Tends to generate tall and large female characters, in some cases proportions will disconnect from the surrounding environment
Struggles with distant faces, use tools or latent upscaling to help
Hands, they're normally an issue but given the focus on upscaling hands struggle an extra deal as latent upscaling adds more chances for the hands to get screwed up (May require post-gen work like in-painting)
Requires additional prompting in some cases to get specific looks, use negatives to help guide clothing/environments/concepts
Other quirks or issues :
Doesn't like limited color pallets, will try and inject color whenever possible
NSFW prompts may be difficult to get right and require experimentation with different keywords+weights
< ----- ----- ----- >
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 over on Ko-Fi!
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!
https://discord.gg/h3vB7S4FEw
