Yifurr版本YifurrNewMoonE2 (ID: 519023)

A heavily tuned model merge I've been slowly adding bits of models that catch my eye, based originally on YiffyMix V3.6 merged with a few other models available on CivitAI and for which I take no credit for their creation. Uses a YAML file you will need to download and place in the models folder next to the model.

This model is, as I said, heavily tuned toward furry/scaly art, but differs in a number of ways from most of the models and merges I've used on the site here; images may be very washed out at low CFG scales, and it functions best at a CFG scale set higher than 20 (24 or higher is what I usually use for the best results), and it benefits immensely from using the Restart sampling method at about 50 steps. Hi-res fix and ADetailer highly recommended. Handles SFW and NSFW, though it occasionally needs to have appropriate tagging in either positive or negative prompt to ensure it does what you want. Not the most detailed at environments, but functions well with plain or colored backgrounds for character art.

This merge also benefits a great deal from tagging specific artsyles, and especially from mixing them together to make distinctive looks (this will be familiar to users of this family of models.) Without artistic or style direction, results will generally be sub-par in my experience, and your mileage may vary. Excels at making hybrid characters with proper use of weighting to generate unique original characters in portrait aspect ratios, while landscape images can yield multiple unique characters with the same method (see sample images for an example of this technique.)

All that said, I've been personally much happier with the results I get from these than any of the base models, and enough folks have been interested in using my version that I've decided to make it freely available.

Tips: Restart Sampling Method, 45-50 steps

Hi-res Fix (4x_fatal_anime_500000_G or R-ESRGAN 4x Anime6B work well) with 0.6 Denoising Strength

CLIP Skip 1

ADetailer enabled (face_yolo8n works well, tunes face details and eyes in particular, enable Merge mode in Mask Preprocessing for best results)

CFG Scale greater than 20 (24+ recommended)

Works with most LORAs and embeddings I've tested, just be judicious with the weights for best results

YMMV with other styles or subjects; heavily content-tuned merge

描述:

One of the better results of a long and annoying process intended to ameliorate the increasingly pronounced 'washing out' of images produced with low CFG values using this model family, as increasing additions to it have led to the minimum effective CFG scale being 30 or higher (which, as 30 is the maximum, began to produce images that were washed out no matter what scale you use.) This project is an attempt to recreate the general aesthetic of the previous merges while slowing the spread of the washout effect. Unfortunately, there's probably not much more that I can realistically do with this model merge series, so despite it being my favorite, I will probably need to switch to improving my other merges after this one.

All that said, this is still a Yifurr merge, so pretty much all the old tricks and tips will still apply here. This edition functions with lower CFG scales than previously, with somewhere around 18 being the happy medium this time around instead of 24, 28, or even 30. This should make it a little more accessible than previous, though it lacks the final layer of fine-tuning the higher-CFG scale versions have.

训练词语: