Underview - Z-Image Penis Model - v2.0

NSFW

Underview v1 - A Penis Model for Z-Image

This model is version 4 of attempting to train male anatomy with z-image turbo. I used the z-image de-turbo model provided by Ostris's Ai-Toolkit. After many attempts, I found that training at a high step count (this model is at 12k steps with 50 images at 512px with the de-turbo preset) and essentially over-baking allowed a low strength to be used so that quality of the image isn't badly degraded. This model is best used with a strength of 0.50-0.67, the quality should be similar to base (or better) and remains very flexible in outputs. It works best with a CFG more than 1 but less 3, I usually use 1.8 at 12-20 steps. You may need to do a low de-noise after to get some finer details. The model was trained with captions provided with the AutoDescribe-Images tool provided by hydropix on github. The trigger word is Underview, I specifically avoided training images of the top part of the glans as loras tend to get confused on which way is top or below, so keep vantage points at a lower angle and expect to get gens with the penis pointed up or to the side.

Underview v2

v2 is a modified version of v1 with an improved and much more consistent output. thanks to the amazing RealTimeLora v2 node shootthesound made!

After trial and error, the layers 24, 27, 28, and 29 could be turned off and eliminate most of the unwanted noise/artifacts that are present when the lora gets to full strength.

The lora also seems to understand higher angle views now, keeping it named "Underview" may be underselling it.

The RealTimeLora node pack also comes with a lora scheduling node that greatly improves the output.

The best schedule for the lora is (0:0, 0.2:0.65, 0.85:1, 1:0.65).

The v2 workflow is here: Underview v2 Workflow

The v1 version of the lora is still recommended to be used as a base pass (0.65 strength) before passing it to the v2 lora (with the above strength schedule).

v1 is better at following the prompted body types if you are wanting a less defined physique.

Description

v2 is a modified version of v1 with an improved and much more consistent output, thanks to the amazing RealTimeLora v2 node shootthesound made!

After trial and error, the layers 24, 27, 28, and 29 could be turned off and eliminate most of the unwanted noise/artifacts that are present when the lora gets to full strength.

The lora also seems to understand higher angle views now, keeping it named "Underview" may be underselling it.

The RealTimeLora node pack also comes with a lora scheduling node that greatly improves the output.

The best schedule for the lora is (0:0, 0.2:0.65, 0.85:1, 1:0.65).

The v2 workflow is here: Underview v2 Workflow

The v1 version of the lora is still recommended to be used as a base pass (0.65 strength) before passing it to the v2 lora (with the above strength schedule).

v1 is better at following the prompted body types if you are wanting a less defined physique.

FAQ

Comments (10)

HairyAIDec 24, 2025· 4 reactions

CivitAI

Any tips creating the lora? I've been fighting with AI-Toolkit for days! What settings did you use? Thank you!

hhhhotAIJan 31, 2026

do you find any solution?

HairyAIJan 31, 2026

@hhhhotAI i gave up lol

nte8

Author

Feb 3, 2026· 1 reaction

The best advice I can give is to make sure your training images are as high quality as possible, as any noise/artifacts in the base image WILL be present in the lora. I mostly use RAW images that have a resolution over 4000x4000px that I then crop/resize to a lower resolution.

These were my ai-toolkit settings for v1, I manually modified the layers afterwards to create v2
---------------------------------------------------------------------------------------------------------
job: extension

config:

name: zimage_uncutgems_v1

process:

- type: diffusion_trainer

training_folder: /app/ai-toolkit/output

sqlite_db_path: /app/ai-toolkit/aitk_db.db

device: cuda

trigger_word: null

performance_log_every: 10

network:

type: lora

linear: 64

linear_alpha: 64

conv: 16

conv_alpha: 16

lokr_full_rank: true

lokr_factor: -1

network_kwargs:

ignore_if_contains: []

save:

dtype: bf16

save_every: 250

max_step_saves_to_keep: 30

save_format: diffusers

push_to_hub: false

datasets:

- folder_path: /app/ai-toolkit/datasets/zimage_uncutgems

mask_path: null

mask_min_value: 0.1

default_caption: ''

caption_ext: txt

caption_dropout_rate: 0.05

cache_latents_to_disk: false

is_reg: false

network_weight: 1

resolution:

- 512

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: 6500

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: false

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: ostris/Z-Image-De-Turbo

quantize: true

qtype: qfloat8

quantize_te: true

qtype_te: qfloat8

arch: zimage:deturbo

low_vram: true

model_kwargs: {}

layer_offloading: false

layer_offloading_text_encoder_percent: 1

layer_offloading_transformer_percent: 1

extras_name_or_path: Tongyi-MAI/Z-Image-Turbo

sample:

sampler: flowmatch

sample_every: 250

width: 1024

height: 1024

samples: []

neg: ''

seed: 42

walk_seed: true

guidance_scale: 3

sample_steps: 25

num_frames: 1

fps: 1

meta:

name: zimage_uncutgems_v1

version: '1.0'

SunInShadowJan 28, 2026

CivitAI

It works perfectly fine. It works with all my Loras. It even works with my anime and cartoon-style Loras. Anthromorphic Loras are a little problematic, but they still work. It also works when I apply my cartoon Loras realistically.

wdwd132Feb 26, 2026

CivitAI

it's still the best lora so far -
but I was hoping that you could put more asian man in dataset when training so we don't get bias result

zramirez075Mar 6, 2026

CivitAI

Hello, Can I get please base workflow to run this model? Not sure to set the right options in the ksamplr.

royrychou346Mar 6, 2026

CivitAI

This is the best lora ever, Plz we need Z image base ver.

Curtis_1995Apr 6, 2026

CivitAI

best erect penis lora then kill the game. really need a zit or base bulge lora