This model is a reimagining of the fictional character Manuela Vicious, by Tower13Studios.
Flux.2 Klein 9B
Checkpoint: Flux.2 Klein 9B
LoRA strength: 0.8–1.6
Sampler: euler, euler_ancestral, res_multistep, or dpmpp_2s_ancestral
Scheduler: simple or beta
Steps: 4–12; 30–50 for base model
CFG Scale: 1, or try 1.2–3; 2–5 for base model
Trigger word (for superior results): ATRX_MANUELA_FLUX2KLEIN9B
Z-Image —Turbo
Checkpoint: Z-Image Turbo, or try Z-Image [Base]
LoRA strength: 1–1.6
Sampler: dpmpp_2s_ancestral
Scheduler: simple or beta
Steps: 4–12
CFG Scale: 1, or try 1.2–3; CFG Norm: 1–1.12
Shift: 7 (first sampler) and 0.6 (second sampler)
Trigger word (for superior results): ATRX_MANUELA_ZIMAGE
Negative prompt (as needed): "Asian woman, black hair"
Qwen-Image 2512
Checkpoint: Qwen-Image 2512
LoRA strength: 0.8–1
Sampler: res_multistep
Scheduler: simple
Steps: 20+, or 4–8 with lightning LoRA
CFG Scale: 2–5, or 1 with lightning LoRA; CFG Norm: 0.94–0.98
Shift: 3.1
Trigger word (strictly required!): ATRX_MANUELA_QWEN2512
Qwen-Image
Checkpoint: Qwen-Image
LoRA strength: 0.75–1
Sampler: euler
Scheduler: simple
Steps: 15+, but 20+ is better, or 4–8 with lightning LoRA
CFG Scale: 2–5, or 1 with lightning LoRA; CFG Norm: 0.8–0.9
Shift: 3.1
IMG-2-IMG denoise: 0.2–0.6
Flux.1
Checkpoint: Flux.Dev 1.0
LoRA strength: 0.8–1.4
Sampler: deis or euler
Scheduler: beta or simple (or ddim_uniform for different look, but add steps)
Steps: 15+, but 20+ is better, or 8–12 with turbo LoRA
CFG Scale: 2–5
IMG-2-IMG denoise: 0.65–0.85
Pony
Checkpoint: Real Dream SDXL Pony 12
LoRA strength: 0.6–1
Sampler: DPM++ SDE or euler
Scheduler: turbo or simple
Steps: 20–30+, or 8–12 with turbo LoRA
CFG Scale: 2–5
ADetailer recommended
Description
Prompts must include trigger word for superior results: ATRX_MANUELA_FLUX2KLEIN9B. Can produce artistic nudes pretty well on its own, but can be further assisted when paired with a NSFW LoRA. With structured prompting, other characters described in the scene can fairly retain their own distinctive features without too much bleed over.
The model was trained at full rank and is quite large (1.1 GB). I've attempted a number of methods to shrink it without retraining, but have so far been unsuccessful.
The dataset for training consisted of 50 high-quality photorealistic images AI-generated using Imagen 4 in Google Flow/Whisk, with references drawn from previously rendered images of the same character. Most images were touched-up using Photopea. SeedVR2 in ComfyUI was used to upscale images from 1024x1024 to 1280x1248. Most captioning was initially carried out in Qwen Chat, but then checked and tweaked through human curation.
No known reference to any real person was included in any part of the training.