City Worlds 01 - Z Image Turbo
Details
Download Files
About this version
Model description
Finding all my models or profile elsewhere than Civitai is without my consent !
Please don't repost my files. Send people to my Civitai profile, thanks :)
-
This is my third Z Image Turbo lora ! ^^
I went through four failed trainings to bring you this one ;)
I didn't think this one would succeed... and so I didn't put the effects...
So... it's a test version... JUST FOR FUN AND TESTING.
--
Important details on settings below !
Also about the dataset !
--
DESCRIPTION :
Well so what is this ?
This is a background Lora with cyberpunk and city vibes. This the Z Image Turbo version.
--
City Worlds 01 :
About the dataset !
This dataset is quite old. It is a mix of 50 % realistic photos and 50 % generated stuff and anime.
Most of the captions are in tag format but as you'll see it doesn't bother Z Image.
So with it you can generate in Z Image :
real life city scenes
anime style scenes (or pixelated...)
Mix of the two above, having real life people in animation environment or animation people in real environment
Cyberpunk scenes
Futuristic scenes with spaceships
And those spaceships and buildings have a special design (Z image does a good job at reproducing the designs from the dataset)
--
Problems :
Sometimes generated images get too dark
Sometimes design spaceships get stuck in design buildings (dataset problem with some low res pics)
Some precision issues (those low res pics in the dataset...)
There are also Airships but those are ugly (no specific design in the dataset)
There are too many descriptions words (see below).... I was doing that in all my first loras.
--
HOW TO USE :
2 types of settings :
--
1 : Turbo
Strength : 1 tested
Steps : 9 - 20 (most of the time I used 20)
Checkpoints : Any
Adetailer : not used
Other : Instead of using the Lumina style, I used the Qwen one.
CFG : 1 or 2
Sampler : EULER
Sheduler : SIMPLE or BETA or NORMAL
--
2 : Not so Turbo... at all
Use this one if you like waiting more ;)
Strength : 1 tested
Steps : 40 - 50
Checkpoints : Any
Adetailer : not used
Other : Instead of using the Lumina style, I used the Qwen one.
CFG : 1 or 2 (2 is of course even longer...)
Sampler : DPM++ 3M SDE
Sheduler : KARRAS
About using DPM++ 3M SDE + KARRAS here :
Since Karras is unusable at a low number of steps and will produce distorted and ugly pics... After tests I found that raising the steps to 30-50 will make nice generations... and sometimes VERY nice generations.
But at the cost a slower generations...
But due to not enough time, I used Euler in most of my generations for this lora.
--
I made tests using multiple combinations between Base Fp8 version, Base Full BF16 model, and multiple text encoder versions. I used Forge NEO (easy install through Stability Matrix (choose Forge Classic package, then change version to choose NEO)), so settings might differ a little for ComfyUI.
--
Trigger word (optional) :
- citybkg
--
Sorry NO EFFECTS this time ;)
--
I'm puting here the description words that you can use and that are in the dataset :
Other trigger words and description words - words an () are optional but will work :
elevated highway
elevatted street(s)
river, lake, sea
(elevated) shore
tunnel(s)
high buildings
low buildings
design buildings
cyberpunk
airship(s) (in the sky) (closeup)
(huge) (small) planet(s) (in the sky)
day
night
evening
sun shades
ambiance
clouds
(very) detailed
illuminated
city lights
signs
kanji
japan
mega stuctures
neons
billboards (on buildings) (on tree(s)) (above buildings) (on highway)
traffic, car(s), buses
car rear, car front
car light trails
bus station
people, crowd
electric poles
bridges
(street) shops, shop under highway
shop signs
stores
space
futuristic
(floating) city in the clouds
wall(s)
street marking, highway markings
sunny
houses
vegetation
(giant) tree(s)
motion blur, speed
reflections
rain, wet
reflections
mist
smoke
metal road barriers
--
Some more in this ZIT version :)
Same descriptions words as above but :
each dataset pic has its own number : "phot0", "phot1", "phot20"... so that if you knew how each dataset pic looked, you could just mix the ones you want... but of course you don't :P
"airships" now become "spaceships" and "design spaceships"
haze
epic scale
massive towering (design) buildings
tiny spaceships
tiny buildings
spaceship focus
car focus
cyberpunk racecar focus
many road levels
many levels
cyberpunk colors
impressive
alive
vertical red light beam on building
red war paint
some lighted windows
construction site behind wall
huge tree
tree growing inside and over a building
moon in the sky
from low angle
mist in the background
haze in the (far) background
dense moving fog on sideways
light effects
partial view of a huge planet in the far background
--
Have fun ^^




















