TL:DR
Check all versions of this model, they do not replace one another, they're different mixes for different things.
What it is
Trinity Mix is a 2.5D model that tries to go for decent realism with great contrast, saturation and dynamic range; and that can do both SFW and NSFW. It's designed to be extremely compatible with detailers, character loras, and with NSFW loras in general. It will take anime and furry character loras decently well and render them in its style. Furry versions are better served with the models in the Shasta branch, generally, as they tend to produce better fur textures. This mix can compete in image quality with XL models, and can easily generate images with up to 1024 base dimensions without big issues. My usual base size is 768×1024 for portraits
The Branches
The different Branches are named after different nuclear test codenames, because I believe this model is the bomb. Unless otherwise specified all the versions are equally valid, and the choice depends on what style you are going for. Do not necessarily default to the latest version
FatMan: The base model. All other branches are based on it
Vanya: These mixes are intended to portray extremely voluptuous female bodies without compromising quality too much. They should react more strongly than normal to appropriate size tags without needing loras like hyperfusion and hyperbreasts that can have severe style leakage at higher weights. V1 keeps things moderately reasonable, while V2 goes full on huge. Breast size tag are as per Hyper Fusion: Large breasts/Huge Breasts/Gigantic Breasts/Hyper Breasts
Shasta: These mixes are intended to give FatMan knowledge of furry/anthro/kemonomimi/monster people anatomy and they can all produce great fur, feather and scale textures. Different versions use different mixes of furry models in them, so test them to see which ones suit your style best
VAEs
All models come with a baked in VAE, but you're free to disregard it and use any of the usual suspects
Quality
Recommended quality tags
((best quality)),((masterpiece)),(detailed),(realistic),(8k), (4k),
Recommended negative
(worst quality, low quality:1.4), (greyscale, monochrome:1.1), cropped, lowres, username, blurry, trademark, watermark, title, multiple view, Reference sheet, strabismus, clothing cutout, side slit,worst hand, (ugly face:1.2), extra arm, bad foot, text, name
Detailers, quality, and style loras
I use these frequently, and many of the images shown will be using them
Detailers
Furtastic Detailer. This works specially well with the models in the Shasta branch
Other probably work just fine, they're just not part of my usual workflow
In general, do not mix multiple detailers in one generation, unless using them at very low strengths
Quality
Furry generations should use the negative embeddings provided by sopenit494 with his fantastic Furtastic model
Envy Better Hands works very nicely with this model, and IME better than other tries to fix hands, such as the bad hands embeddings. The model will still have issues with hands though. as it's still a 1.5 model
This model reacts very well to both positive and negative strengths of Anime Realism Control Lora. This lora can be combined with detailers to get quasi-photorealistic results This was done with the same prompt and seed, and with the FatMan version of Trinity Mix
Prompting style
I have found that this model works best if you prompt in the following way:
<Quality Tags>
<your usual tags for 1.5 generations>
<a narrative prompt describing the same thing as the tags listed above>
One trick that is fairly easy to implement here at TensorART is to make your prompt as normal for a 1.5 model, then copy it, hit the "Enhance" button under the prompt checkbox, and then paste your original prompt before the generated text. You want to keep your quality tags and your regular tags at the start of the prompt though. You might want to surround your tags with parenthesis and add some extra weight to the initial tags. This greatly increases the quality of the generation in my experience
Getting furry characters
Shasta versions are mixes with furry models, not full furry models themselves, so they might require some extra cajoling to get them to generate furry characters without a character lora. As such, two techniques tend to work fairly well, depending on the generation. They can be combined if necessary.
Use (anthro, anthropomorphic:1.x). Use up to 1.5 weight here, more than that tends to break the model, 1.3-1.5 is usually enough
Use the AND construct e.g. (((wolf:1.5) AND (girl:0.5)):1.3)
Clip Skip
This model works with both Clip Skip 1 and 2. Clip skip 1 tends to have a higher quality celling and lower quality floor.