Flat Color - Style
Trained on images without visible lineart, flat colors, and little to no indication of depth.
ℹ️ LoRA work best when applied to the base models on which they are trained. Please read the About This Version on the appropriate base models and workflow/training information.
This is a small style LoRA I thought would be interesting to try with a v-pred model (noobai v-pred), for the reduced color bleeding and strong blacks in particular.
The effect is quite nice and easy to evaluate in training, so I've extended the dataset with videos in following versions for text-to-video models like Wan and Hunyuan, and it is what I am generally using to test LoRA training on new models now.
Recommended prompt structure:
Positive prompt:
flat color, no lineart, blending, negative space,
{{tags}}
masterpiece, best quality, very aesthetic, newestDescription
Trained on Z Image Base with diffusion-pipe
Same dataset/training settings as used for version 2.1 [z-image-turbo]
Training Config:
# dataset-zimage.toml
# Resolution settings.
resolutions = [1024]
# Aspect ratio bucketing settings
enable_ar_bucket = true
min_ar = 0.5
max_ar = 2.0
num_ar_buckets = 7
[[directory]] # IMAGES
path = '/training_data/images'
num_repeats = 5
resolutions = [1024]
# config-zimage-base.toml
output_dir = '/mnt/d/zimage/training_output'
dataset = 'dataset-zimage.toml'
# training settings
epochs = 50
micro_batch_size_per_gpu = 1
pipeline_stages = 1
gradient_accumulation_steps = 1
gradient_clipping = 1
# eval settings
eval_every_n_epochs = 1
#eval_every_n_steps = 100
eval_before_first_step = true
eval_micro_batch_size_per_gpu = 1
eval_gradient_accumulation_steps = 1
# misc settings
save_every_n_epochs = 5
checkpoint_every_n_minutes = 120
activation_checkpointing = true
partition_method = 'parameters'
save_dtype = 'bfloat16'
caching_batch_size = 8
steps_per_print = 1
[model]
type = 'z_image'
diffusion_model = '/diffusion_models/z_image_bf16.safetensors'
vae = '/models/vae/ae.safetensors'
text_encoders = [
{path = '/models/text_encoders/qwen_3_4b.safetensors', type = 'lumina2'}
]
dtype = 'bfloat16'
#diffusion_model_dtype = 'float8'
[adapter]
type = 'lora'
rank = 32
dtype = 'bfloat16'
[optimizer]
type = 'AdamW8bitKahan'
lr = 2e-5
betas = [0.9, 0.99]
weight_decay = 0.01FAQ
Comments (17)
What is your experience with Z-Image Base so far? Do you think we finally have a good successor to SDXL?
It's a very pretty model! Seems Tongyi put effort into improving the base aesthetic before releasing it, maybe at the expense of diverging from the derived Turbo model.
It takes a long time to converge when training though, so I guess it will take a bit of time to tell whether it would succeed SDXL as a community model.
you are the fastest lora maker in the world
Anyone able to do outlines?
I'm curious what training parameters you used for the z-image base model training.
Hello, I've added the training parameters to the model card "About this version" section
Curious why you dropped the learning rate to .00002, instead of the default ~.0001.
I've noticed personally the z-image base model converges really slowly and have been testing with bumping the learning rate up rather than down. Curious if you know something I dont, trying to figure out what works and what doesnt.
Maybe its because this style in particular is really simple, and so doesnt need to learn a lot?
Hello, I dropped from .00005 to .00002 when attempting to train z-image-turbo, then used the same z-image-turbo config to train z-image-base as a first blind attempt.
Convergence does seem slow and I've had difficulty training other datasets with this config, as you say for this simple style not a lot needs to be learned.
There's a lot of variance on each seed, and base generations seem to be very close to the dataset on one seed and then not at all on the next which makes it a little hard to measure too.
@motimalu Thanks for the reply!
I've been bumping up the learning rate slowly, going to try .0003 next, I read someone say somewhere that it started to break down around .0004 though, so we will see.
Even at these rates though, I cant help but notice it always feels like the model is 85%~ there, but never seems to converge completely. I'm going to try much more elaborate tagging, and much higher steps (Maybe 5k next, assuming it'll be good around 4k?).
My hypothesis for now is that my natural aversion to overcooking the model is holding me back and I just need to increase the step count dramatically. The problem with z-image-base is its really hard to tell when you start losing flexibility since it hides it so well.
Anyways, just trying to figure this stuff out, thanks for your time!
Hello, yes a trigger word could help.
Tags that the model understands to reliably generate something close to the dataset should be included at the start of the dataset tags I guess.
Negative prompts are also supported in the z-image-base model, so you could try that too.
really good
Cool one!
cool
Can the Flux K version edit anime images too?
It might work, I didn't train it with control/context images for edit though. Something I've been meaning to try
can you please create this flat artstyle - anbeauty0902 for noob or IL
this is really good