UnNamed IXL V7!

After listening to the active users of UnNamedIXL, here is an update that should fix some issues with V6 while keeping the dynamism brought by it (V7 will replace V6 at next auction).

Special thanks to several heavy users/testers of UnNamedIXL V5!

• https://civarchive.com/user/faustoserone393

• https://civarchive.com/user/Blackskullart

• https://civarchive.com/user/wesfree

• https://civarchive.com/user/AMPLIFai

• https://civarchive.com/user/VicePalette

• https://civarchive.com/user/Brewce

• https://civarchive.com/user/Mulinshtein

• https://civarchive.com/user/Cookietinsewingkit

• https://civarchive.com/user/lonecatone23

• https://civarchive.com/user/viciviie

• https://civarchive.com/user/coolhotrod813

In order to make it look like what was expected by these power users, i leveraged some pictures from UnNamedIXL V5 public gallery to create a dataset. This dataset was used to create 2 LoRAs:

• UnNamed 5 to 7

• UnNamed 6 to 7

Both these LoRAs allowed V5 and V6 to get closer to an ideal V7 (internal models u7-5 and u7-6, with the LoRAs merged at 0.8) which were then merged together to get V7 (NuSlerp 7-5 + 7-6 at 0.4).

Skin and facial features should feel more natural with this version while being more dynamic / less static. V8 will probably a "Krea2" upgrade of V7 (dataset made with V7, img2img using Krea2) in order to get better hands, but is not scheduled for now, i'll see after some external tests. Thanks for reading!

Warning feedback

This is exactly the kind of honest feedback i love to receive. So, for anyone looking for this info, here it is:

Unamed v5 is superior to v6. v6 tends to distort into skin plasticity and weird mannequin features.

If anyone also experience this kind of stuff, feel free to DM or comment in order for me to get better results in the future <3

For now, i'll keep both available on-site (until an optional better V7).

UnNamedIXL V6!

It's an upgrade made while working on the ANIMA version of UnNamed. I am not quite sure it is doing exactly what i wanted but still like it! For less steps, you should be able to get better skin texture and poses. I am using Euler A + SGM Uniform, 30 steps and a special negative embedding in my pictures. Photorealistic can be added to the prompt optionally.

UnNamedIXL V5!

This will probably be the last release of UnNamedIXL. This version improve the details at the cost of slightly less following of prompts. It was made by making a merge of Illustrious Realism and CatPony-Real and using this merge as a refiner on top of V4 to generate a dataset. Generated images were used to train a "booster" LoRA which i merged with V4 (by pruning and reweighting some layers). Check the showcase for some sample/scheduler pairs that works nicely 🤩

Origin

Per request, i have uploaded the "Arn0-realistic" model found here as a "V0" of UnNamedIXL. This checkpoint merge was used during the initial build of V2.

UnNamedIXL V4!

Merry Xmas! I'll update the recipe and more info once i figure out a consistent negative prompt and parameters 😅

V4 is based on a careful merge of some WaiRealism and Perfection Realistic v5.0 to avoid loosing what makes UnNamedIXL different while fixing some funkiness that can happen in the background.

Involved block merge (in order):

1. WaiRealism (at 0.3): IN01, IN02, OUT03, OUT04, OUT07

2. Perfection Realistic (at 0.5): IN03, OUT03

3. UnNamedIXL 3 (after adding UnNamed Spice LoRA) (at 0.5): BASE, M00

For the record, here is a sd-mecha scripts for MBW + LoRA which also fix NaN in layers using IXL V2.0 (a plague in my previous merge):

import sd_mecha as sdm
from sd_mecha import merge_method, Parameter, Return
from torch import Tensor

sdm.set_log_level()

@merge_method
def fix_nan(
  a: Parameter(Tensor),
  b: Parameter(Tensor),
) -> Return(Tensor):
  if a.isfinite().all():
    return a
  else:
    return b

base = sdm.model("UnNamedIXL_V3.9.safetensors")
src = sdm.model("mine\\UnNamedIXL\\UnNamedIXL_V3.fp16.safetensors")
ill = sdm.model("base\\illustriousXLV20_v20Stable.safetensors")
lora = sdm.model("..\\loras\\UnNamedSpice.safetensors")

l = sdm.convert(lora, base)
a = sdm.add_difference(base, l, alpha=0.1)

b = fix_nan(src,ill)

blocks = {
    "BASE": 0.5,
    "IN00": 0.0,
    "IN01": 0.0,
    "IN02": 0.0,
    "IN03": 0.0,
    "IN04": 0.0,
    "IN05": 0.0,
    "IN06": 0.0,
    "IN07": 0.0,
    "IN08": 0.0,
    "M00": 0.5,
    "OUT00": 0.0,
    "OUT01": 0.0,
    "OUT02": 0.0,
    "OUT03": 0.0,
    "OUT04": 0.0,
    "OUT05": 0.0,
    "OUT06": 0.0,
    "OUT07": 0.0,
    "OUT08": 0.0,
    "VAE": 0.0,
}
blocks = sdm.convert(blocks,b)
blocks = blocks | 0.0

c = sdm.slerp(a,b, alpha=blocks)

sdm.merge(c,output="test.safetensors")

V3

I didn't want to stray too far from what made V2 a different realistic checkpoint, but i still wanted to give it a small lifting. I'll use this as a base for a future V4 ;-)

V3 + DMD2

Since i am going to play a bit with the DMD2 LoRA, i have integrated it in UnNamedIXL V3. This version is really experimental, prefer V3 if you want to do a classic realistic image as usual. Otherwise, this version is better used as a "refiner" or "Hires.Fix" model switch leveraging an other model to give a realistic boost on other composition. Here is an example of Hires.Fix using this version and both DMD2 version of AnBan and HoJ:

(click on show more for the rest of the story 😁)

Simply use Euler A, CFG 1, 8 steps in both standard and Hires.Fix run but switch to UnNamedIXL V3_DMD2 during Hires. You can also try other models for the base pictures as usual and do img2img only.

Here is a base picture using WAI14 (30 steps, CFG 5):

And the img2img, CFG1, 8steps, 0.4 denoise, x1.5 resize:

Previously

Brand new V2 is a full realistic model!

Build from scratch using:

• Perfection Realistic [ILXL / Illustrious XL]: https://civarchive.com/models/1257570/perfection-realistic-ilxl-illustrious-xl-nsfw-sfw-checkpoint

• PornMaster-Pro 色情大师- Illustrious & noob: https://civarchive.com/models/1045588?modelVersionId=1570999

• RealismHarem: https://civarchive.com/models/1478309/realismharem

• And a tiny bit of JedPointReal: https://civarchive.com/models/928145?modelVersionId=1787505

It then received the "ILX v2 upgrade" before some precise MBW with UnNamedIXL V1 to get its composition more to my liking :D

Usage:

• DPM++ 2M SDE Heun + SGM Uniform

• 40 steps

• CFG > 6 is prefered, but feel free to experiment

• Positive prompt: Can handle NLP 🎉

• Negative prompt: whatever suit your fancy, just be aware that Illustrious generate "young" characters

  • IllusN3g

  • Stable_Yogis_Illustrious_Negatives

  • A mix of ACTI

If you prefere "semi-realistic" bordering on realistic, V1 is still your best bet :D

V1: An "almost" realistic checkpoint, because pure realistic models feels too uncanny for me yet.

Long Story™

First, special thanks to @rambo943 who gave me the idea and the base for this model.

The release of RealismHarem and my tests (and the associated message received) gave me the hint that merging a bit of my other checkpoints into it could be a way to "cure" a real realistic model to better fit my current mood.

The merge started as a CosineB merge of RealismHarem into CuteIXL.

Then, i added MM Gothic NJ + Asian Beauty XL + Eyes High Definition + Snake Oil to get a specific style to the picture.

But since i threaded too much outside of the realistic feel i wanted, I redid a merge with the previous checkpoint before integrating a bit of HoJ.

Afterward, Web Realistic Illustrious got in the mix at small dose and step by step, on the course of 9 Release Candidates, i smoothed the result but adding and substracting RealismHarem and some of my checkpoints at block level until i got what i wanted.

Thanks for reading! 💖

Update for the more curious, some intermediary steps:

V0

RC1

RC4

RC4 vs RC5 vs RC6

CuteIXL vs HoJ vs RC6 vs V1