Welcome to the world of SpiritMix! This time around we have a 2D anime mix which attempts to create a diverse model with dark, sketch-like outlines (preserving that 'drawn' look) while retaining the diversity and detail of the more complex anime models. The difficult part here was preventing the model from leaning too far in either direction, either eliminating the lines, becoming too detailed, or leaning too heavily into the lines and creating a flat image. Overall it works quite well, though the style may shift around depending on the setting and seed. Also, unlike my previous models this one DOES include NSFW data, so it does phalic objects! (Though NSFW will still be inconsistent X.x) Make sure to share your results!
Note : Please read the "Prompting Help" listed under suggested settings, this can help you get the results you want!
< ----- ----- ----- >
< ----- ----- ----- >
Model Name :
SpiritMix - Deep Line (2D Anime) [hash:0366ebd716]
Experimental Name :
_SpiritMerge3_1_2_4_6_0_1_2.safetensors [hash:3280d64c9a]
(Some of my gens may have the expiremantal 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 2D anime style with dark deep lines, high contrast character detail, and a sketchy overall look as if it was drawn by hand.
SFW / NSFW :
The model excels in SFW content, however NSFW content is viable as the mix includes NSFW data. Not considered safe to use in public!
Bias :
Leans toward simpler backgrounds (makes sense given the art style)
Creates 'thick' detail lines, such as a bias toward a thickly drawn eyelash style, has very thick/heavily inked eyes and other details
NSFW 'anus' tends to lean more 'gaping' style
Many other 'typical anime' bias in terms of styling and body form/proportions/anatomy
Tends to favor lighter skin colors, even when prompted for other colors (anime >.>)
< ----- ----- ----- >
Quick NSFW Warning :
This model ended up being very diverse, however this also means the NSFW capabilities ALSO become diverse. As such, be VERY careful with NSFW prompts (as is true with most anime models) to prevent creating subject matter you don't wish to see, and always include body types in your prompts such as (my personal favorate) 'female, woman, large breasts'. The model defaults to SFW output so you shouldn't stumble into danger, but it's worth keeping in mind.
< ----- ----- ----- >
So what is SpiritMix - Deep Line (2D Anime)?
This SD1.5 based checkpoint merge uses various anime style models to try and create a sketch-light style with deep and dark outlines and high contrast character detail. The focus was getting this sketch style without sacrificing the detail found in a lot of more complex anime models and overall it works quite well. To achieve this mix I used a mix of both private and public models, starting with some more detailed models then stacking on-top various models with stylistic traits I wanted to see in the final output, rotating between detailed and sketchy as neccesary to keep detail consistency and prevent the heavy "simplified" bias sketch styles tend to have. One interesting note for those other checkpoint mergers out there, one of the key aspects to getting this style was mixing in corrupted checkpoints that were considered failures and artifact! These artifacting models were mixed in through out with low weights to give the outlines their sketchy appearance (most likely from the artifacts). While this wasn't neccesary for this specific look, it helped retain a lot of detail and add a unique look to a lot of gens!
< ----- ----- ----- >
Prompting Assistance and Usage Information :
Of course you can use this model however you want, but it was mixed specifically for that dark-lined anime style, here's some tips based on my tests!
For complex scenes and detailed output you can use prompts like "best quality" or "high quality" with "highly detailed, detailed background" to avoid simplified output
If you're trying for simple output then avoid prompts like "detailed, masterpiece, complex" and use simplified prompts like "vector, simple, sketch"
You can also try mixing around these 2 configs in the positives and negatives to make the art style somewhat more consistent
If you see artifacts in the image like moire patterns you can try adding "hatching, cross-hatching, strippling, canvas" to the negatives, though this may simplify the image.
For characters I mostly just tested female subjects (sorry male lovers >.<) though this mix can do males.
For female bodies, if you want more adult woman I suggest "woman" in the tags, female alone will often vary the body based on other body prompts such as "large breasts", for younger looking characters use "girl" however avoid NSFW prompts when including this term.
For body sizing, obvious breast and chest tags apply, however so does "thick thighs, short height, tall height, chubby, skinny, ect."
I recommend always trying to include the body tags you're looking for to avoid body variety getting too out of hand.
I assume most of the same rules apply for male characters but with male style tags instead.
To force limited color palettes place "Monochrome" (and optionally "black and white") at the beginning of the prompt, then add color signifiers for different tokens like "green eyes, blue hair, orange jacket, yellow shirt", regional prompter can help with color bleed.
Suggested Settings :
Add "best quality" or "high quality" to positive prompt, adding various other tokens such as "highly detailed, detailed background" can help make the scene and character more detailed. "vector" and "simple" can help have the opposite effect, as well as style prompts like "sketch".
Avoid ancestral samplers like Eular a, the random noise can cause artifacts to appear more often on the edges of characters, some other samplers may show these artifacts more often as well but I didn't test every sampler.
High-ish step count for a clean resolve (I used 40 with DPM++ 2M SDE Exponential, though you could probably get away with much less like 30~)
Image sizes of 768x768, 512x512, 512x768, 768x512 seem to work well (512x is more consistent but may have more errors around areas like hands)
Hires Fix works well with lower denoising, latent and similar upscalers may add more detail than desired or artifact so I recommend a soft upscaler
Clip-Skip 2, tends to produce more creative results but it isn't required
VAE with saturated colors, anime style VAEs tend to work well but be careful with VAEs that oversaturate the image
What this mix does well :
Replicates drawn anime really well with lots of high contrast detail
Great with bright colors and colorful scenes, but also competent with dark scenes and moody atmosphere
Responds to color palettes well, such as monochrome or black and white
Has a large body and face variety (Though prompting may be required for some body types to show up)
Does contrast really well
What this mix struggles with (aside from typical AI issues) :
Without specifying body types in the prompt, the body may vary wildly
Occationally mixes up body proportions, creating things like tiny/shortened legs on sitting characters or weird head sizes
NSFW content occasionally tries to add manga-style artifacts (like hearts, smears, spiked lines, ect)
Struggles a lot with certain poses when purely prompted, ControlNet helps a lot with this though some NSFW/Nudity may not always resolve well
Backgrounds are often fairly simple, forcing complex backgrounds tends to reduce the artstyle / line weight
Struggles with darker skin colors, has a heavy bias toward lighter skin tones
Other quirks or issues :
The art style may vary from seed to seed and scene to scene, don't expect 100% consistent results
Because of the art style and unique mix, may not always respond well to Loras
When using limited color palettes through prompt it may simplify backgrounds more than expected
< ----- ----- ----- >
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/michaelpstanichIf 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
