It is actually more of a style lora, It was trained on 1200x800 images. I am aware of the color cast it comes and go's and is the result of a smaller imbedded dataset to add some style to the lora. I have been using it for a while at higher or lower powers and haven't found it to destroy the quality of the base model. But I have found if you use the heretic clip loader it will ruin ruin outputs and make it look overtrained. so I recommend not using the qwen heretic clip loader. I have really tried this at multiple prompts completely on or off and haven't found it too destroy the quality. I'm surprised its giving you that issue! But I am aware that I am using the Q8 gguf and there are a lot of other versions that is way I included a screenshot of my models as loaded in comfyui so people could see how I use it. I do appreciate the feedback though! I'd be curious to know what model version of zimage you run and which clip.

I really prefer the look mine! hahah to each there own. But i'm so happy you made that comparison!

oh holy shit 11gb lora! I must try this

I've love to pick your brain about when I should be using a higher rank?

---

job: "extension"

config:

name: "Quality_1_Training"

process:

- type: "diffusion_trainer"

training_folder: "output"

sqlite_db_path: "./aitk_db.db"

device: "cuda"

trigger_word: ""

performance_log_every: 10

network:

type: "lora"

linear: 128

linear_alpha: 128

conv: 128

conv_alpha: 128

lokr_full_rank: true

lokr_factor: -1

network_kwargs:

ignore_if_contains: []

save:

dtype: "bf16"

save_every: 1000

max_step_saves_to_keep: 16

save_format: "diffusers"

push_to_hub: false

datasets:

- folder_path: "/path/to/images/folder"

mask_path: null

mask_min_value: 0.1

default_caption: ""

caption_ext: "txt"

caption_dropout_rate: 0

cache_latents_to_disk: true

is_reg: false

network_weight: 1

resolution:

- 1024

controls: []

shrink_video_to_frames: true

num_frames: 1

do_i2v: true

flip_x: false

flip_y: false

train:

batch_size: 1

bypass_guidance_embedding: false

steps: 10000

gradient_accumulation: 1

train_unet: true

train_text_encoder: false

gradient_checkpointing: true

noise_scheduler: "flowmatch"

optimizer: "adamw8bit"

timestep_type: "weighted"

content_or_style: "balanced"

optimizer_params:

weight_decay: 0.0001

unload_text_encoder: false

cache_text_embeddings: true

lr: 0.0001

ema_config:

use_ema: false

ema_decay: 0.99

skip_first_sample: true

force_first_sample: false

disable_sampling: true

dtype: "bf16"

diff_output_preservation: false

diff_output_preservation_multiplier: 1

diff_output_preservation_class: "person"

switch_boundary_every: 1

loss_type: "mse"

model:

name_or_path: "Tongyi-MAI/Z-Image-Turbo"

quantize: false

qtype: "qfloat8"

quantize_te: false

qtype_te: "qfloat8"

arch: "zimage:turbo"

low_vram: true

model_kwargs: {}

layer_offloading: false

layer_offloading_text_encoder_percent: 1

layer_offloading_transformer_percent: 1

assistant_lora_path: "ostris/zimage_turbo_training_adapter/zimage_turbo_training_adapter_v2.safetensors"

sample:

sampler: "flowmatch"

sample_every: 250000

width: 1024

height: 1024

samples:

- prompt: "man with red hair, playing chess at the park, bomb going off in the background"

- prompt: "a man holding a coffee cup, in a beanie, sitting at a cafe"

- prompt: "a horse is a DJ at a night club, fish eye lens, smoke machine, lazer lights, holding a martini"

- prompt: "a man showing off his cool new t shirt at the beach, a shark is jumping out of the water in the background"

- prompt: "a bear building a log cabin in the snow covered mountains"

- prompt: "man playing the guitar, on stage, singing a song, laser lights, punk rocker"

- prompt: "hipster man with a beard, building a chair, in a wood shop"

- prompt: "photo of a man, white background, medium shot, modeling clothing, studio lighting, white backdrop"

- prompt: "a man holding a sign that says, 'this is a sign'"

- prompt: "a bulldog, in a post apocalyptic world, with a shotgun, in a leather jacket, in a desert, with a motorcycle"

neg: ""

seed: 42

walk_seed: true

guidance_scale: 1

sample_steps: 8

num_frames: 1

fps: 1

meta:

name: "[name]"

version: "1.0"

It seems it simply comes down to a difference in our aesthetic preferences. I've tested nearly every z-image checkpoint released on Civitai and other forums, but I have yet to find a model that can replace the official one. Most of them are still garbage.

I currently use z-image turbo bf16 with the standard VAE and text encoder. Occasionally, I use the abliterated text encoder shared by huihui when my prompts get blocked: https://huggingface.co/huihui-ai/Qwen3-4B-abliterated

I recommend using Rank 128. I’m sharing my ostris training settings with this message. I hope this helps. but it's for 24gb gpu setting.

@RABBAI  I agree the alternative checkpoints are all garbage. Its a waste of time to make them until the true base model is out. thank you for the information. I tried to figure out what was best for getting the details I wanted. I saw a jump in quality when I went from rank 32 to 64 but was worried about overdoing it and wasting hrs of runpod time so that's why I jump up to 64 rather than 128. There is a lot of people on youtube that do tutorials on lora training that clearly know no more than me which is frustrating. I will likely take another stab at this lora in the future, I'll run rank 128 and perhaps I'll prune the style elements out of the dataset. I'll likely also replace all the photos with best quality versions but that will be a task ugh.

@iamgeekusa Yes, I look forward to your new model. However, don't spend too much time on it. I used to train a ton of Flux models, but I found that they all became useless over time. Who knows, maybe some crazy company will eventually release a fully NSFW-capable model.

By the way, the Ostris preset I shared is set to 1024px, but you need to change that to 1536px. I must have sent the one I was using for some recent tests with a 1024 dataset. Cheers.

@RABBAI he did after all call it the bean effect

@RABBAI I think if I make another I might lean into it being even more of a style lora and a NSFW lora. Mystic xxx is clearly better for just basic stuff. Its kinda why I did mine anyway. Although if you mix mine with mystic xxx you can get some fun stuff. mystic .60 mine 40

No offense, but in your comments here you just come across as a bit too pretentious to me.

Oh well, whatever floats your fancy I suppose.

@Latent_Dreamscape really? I was just happy to get any feedback not sure how I'm being pretentious. oh well.

you should try my new one with the workflow in my newest article, I think you'll be impressed

yea that was a stretch ahah but tastes vary, I'm not a fan of the current AI thick eyebrows lady people keep training that seem to be pony based

yeah I wondered how man people would get the crumb reference

we can see your images bro it's not a secret ":3"