Phoenix

trained on juggernaut XL, noise offset is 0.018

I recommend using the word Phoenix.

You should be able to change the color of the bird. I prefer a multicolored bird except for ethereal. In my mind this is a normal awesome creature that is capable of engulfing itself in flames. I tried to train some flaming birds with ashy talons and some with regular talons. Great effort was put on generating quality talons with 3, 4, or 5 claws.

V3: Introducing Immolation support! Now, the phoenix boasts a mesmerizing look with parts seemingly crafted from liquid fire. I've fine-tuned the model to eliminate quirks, ensuring the fire no longer looks feathery, and the bird doesn't appear to hold feathers. It's performing exceptionally well with hybrid creatures even with simple prompts. I welcome your feedback on this latest version! Your thoughts and constructive criticism are invaluable to further refine and enhance the model. Thanks for the ongoing support and collaboration!

V2 notes:

USE ethereal for less bird like more spiritual phoenix

using flaming to increase fire.

use fireless to reduce the amount of fire.

you can also put 1 or 2 of those 3 words in the negative prompt.

I can not promise you fireless will always generate a flame free bird of prey nor can I promise flaming will always have fire but it is better then juggernaut XL default

File names is fixed, training reached my intended duration. I've observed an interesting quirk: sometimes the phoenix appears perched on feathers, possibly due to my stupid reluctance to exclude a few long-tailed phoenix that I really liked from the training data. Resolving this, as well as an issue where the fire itself occasionally takes on a feathery texture, proves to be a challenging task. I believe this happens because the model is trained on the feathers and the area around the bird as being on fire. Given the magical nature of the bird, the feathers also glow, leading the model to interpret them as a change to fire. I'm actively working on V3, aiming for a release before Christmas, depending on my progress. Currently, I'm employing LoRA at low strength to generate high-quality images of objects on fire, using them to fine-tune and regularize the model.

V1 notes: two training issues. Planned 24 epochs, but a file overwrite with save state enabled caused space shortage, ending training prematurely. Additionally, file renaming occurred due to a side project on NYC trains, affecting only the file name in the code, with no impact on your end unless you deep dive into the file you won't see the train names and stuff none of that is in this lora just the previous file name is hidden inside the code. I DO NOT RECOMMEND DOWNLOADING V1. IT IS NOT BETTER THEN v2 or v3 in anyway.