Type | |
Stats | 137 |
Reviews | (13) |
Published | Nov 6, 2024 |
Base Model | |
Hash | AutoV2 10BFCC9F33 |
WARNING: HERE BE DRAGONS!
This model is made though merging with ComfyUI and, therefore, incompatible with Model Toolkit. Do not use Model Toolkit to check/prune this model, for this is impossible without destroying the model itself. See this article for more information.
What is this model?
ConcoctionMix (now Pandora's Box) is my experimental branch for OpenSolera, my Stable Diffusion 1.5 anime-style model. The model is my foray into ComfyUI merging and continuation of CLIP experiments (thanks @daskmaster for the CLIPs). CCM-a1 is probably one of, if not the hardest model to work with that I've made, with future models growing in complexity and pain.
The model should work the same as other Stable Diffusion 1.5 checkpoints, just much harder to merge with other models without ComfyUI.
Checkpoints:
a1 [Vodka]
a2 [Vermouth]
a3 [Mojito]
LoRAs:
As of a3 [Mojito], the following models are added:
Hyper-SD CFG 12-step LoRA (despite being a Hyper-SD merge, it doesn't really do that much better on steps)
CLIP: zer0int's ViT-L-14 smoothGmP
VAE (baked in):
a1 [Vodka]: PPPAniMix - v3
Main VAE: nubby/kl-f8-anime2-blessed
How to use this model?
Make sure to check the "About this version" section of each model as well for some tidbits and info.
Prompting:
This model (in my testing) uses booru tags very well, but can use natural language when necessary.
For best results, a style must be prompted (in artist and/or medium) to give this model a style to work with. Versatility is this model's strong point. Base model (without style prompt) jumps between styles very often. and can sometimes drift into 3D/2.5D art styles and sometimes realism.
Like most anime checkpoints, this one also has a female bias, but it can be countered.
This model is mainly tested against ComfyUI's prompt parser and reForge's default parser.
Character recognition should work well with really popular characters, but not much else.
Parameters:
(Anything in bold has been tested and working decently fine)
Sampler + Scheduler: Almost anything works, with recommendations being: Euler a, DPM++ 2M Karras/AYS, DPM Adaptive, UniPC-simple, DDIM (with ddim_uniform),
Steps: 20+
CFG: 6-12 (recommended range: 7-12)
CLIP Skip: 1-2
Resolution: 3:2 aspect ratio tested, 640x640, 640x960, 960x640, 512x512, 512x768, 768x512 all works to some degree
Hi.res fix: Recommended, but not necessary for close-ups. For multiple people, this is a necessity. (This is wrong as of CCM-a2. Hi.res fix for CCM-a2 is more required to get strong results)
There's also a unique workflow built for most of my SD1.5 models for almost any txt2img purposes. Use that for best results.
See OpenSolera's page for some recommendation and guides
On a technicality, this would be OpenSolera-a6 [Plex] (or at least a version of it), but due to it being hard to work with for everyone and the difference in style, this model is separated from the series.
Documentation
This section will be used to list my findings and recommendations on this model. Maybe I'll make it into a separate article if it gets big enough.
Model metadata
The first 2 models, a1 [Vodka] and a2 [Vermouth] have fully available metadata within the model themselves. However, since a3 is made from 2 parts, I'll upload both images in this section instead.
a3 [Mojito]
Both image includes metadata as well. Load into ComfyUI to check.
Prompting-related findings
Realistic prompts: (realistic, photorealistic: 1.2). With this, this model could feasibly be a DreamShaper knock-off since it can do both realism and anime easily. Also remember to lower the strength of any other negative tags (worst quality, low quality,...) when using theses tags.
Please note that for a2 [Vermouth], these tags should be thrown into negative prompts to get the anime style. If you want, you could somehow make a 2.5D style out of this
Without clothes prompt, this model will generate nude in many cases, even when you don't want it, so please, give them clothes.
Styles and mediums:
For all mediums, add \(medium\) at the end of each set. Tag strength at 1.2 is also recommended to enforce the style/medium properly (on base ComfyUI parser, no normalization)
Western comics: working, use western comics
Alphes (style): unsure
Traditional art mediums: some works, currently: watercolor (works as expected), ballpoint pen (gives pencil-like drawings), airbrush (not sure what style it gives), charcoal, colored pencil (kinda), color ink and ink (compatible with color tags), crayons (kinda), gouache (not sure), graphite (mostly), marker, nib pen (unsure), oil painting, pastel, photo (gives realism), watercolor pencil,... Additional tags that synergize is paper, canvas,...
Digital art software: photoshop (realism), painttool sai (drawings, as expected), clip studio paint (UI on every image, drawings), krita (drawings), medibang paint (unique style), ibispaint (soft style, might feature UI),...
Dakimakura (body pillow): somewhat working, recommend simple background tag (sometimes). The fabric can be dyed by setting the background color (sometimes). Double dakimarura also works as a medium if that's more your style
Cards: for trading cards, use card, for tarot cards, use tarot card or tarot. If you with to have regular playing cards (can't make actual cards atm), put trading card in negatives
Traditional art styles: abstract, art deco/art nouveau (different movements, but it blended into one another), fine art parody (maybe), impressionism, ukiyo-e, sketch, surreal
Posters: works pretty well for propaganda and movie posters. Other forms are untested
Calendars: works, but why would you want to generate a calendar outright with this thing? (genuine question)
Clear files: technically works, but again, why would you want that when you could generate a simple (and better) portrait/landscape pic and put that up instead? (genuine question)
Blueprints: technically works, not sure what to do with it though
Close-ups: works surprisingly well with specific body parts (limited to what humans have)