🌸 Virtual Japanese Woman YUMEKA – Qwen Edition (v1.0)

Virtual Japanese Woman YUMEKA – Qwen Edition (v1.0)

Trained on 1024x1536 images to reproduce a photorealistic Japanese woman style named YUMEKA.
Not a glamorous idol, but rather the vibe of a “beautiful lady you might see in everyday Japan.”

🔹 Trigger word: "yumeka" (most likely works even without trigger)
🔹 Recommended weight: 0.7 – 1.0

⚠️ This character is fictional and does not exist in reality.
⚠️ Facial appearance may vary depending on other LoRAs or workflows used.
⚠️ Highly prompt-dependent: please refine features (bangs, blush, lip gloss, gaze, angle, expressions) via prompt.

🟢 Tested on Qwen with good results, though still in the testing phase — generation may fail at times.

🔧 Workflow Node Example (Sampler settings)

📌 The sample images were generated using two different samplers: some with res_2s_ode, others with exponential/res_2s (ClownsharkKSampler).
Results may vary slightly depending on which sampler is used.

ClownsharkKSampler example settings:

• sampler_name: exponential/res_2s

• scheduler: bong_tangent

• steps: 8 (if unstable, try 10–12)

• steps_to_run: -1

• denoise: 1.00

• cfg: 1.00

• eta: 0.50

• seed: randomize (example: 1046916965579049)

• sampler_mode: standard

• bongmath: true

🟢 In my case: Sampler = res_2s_ode, Scheduler = bong_tangent, Steps = 12 → higher chance of stable results. Please adjust within 8–12 steps.
⚠️ Euler, Simple, or strict 8-step settings may cause collapse (at least in my ComfyUI “nunchaku” build, reason unknown).

📂 Workflow settings are embedded in the sample images for reference.

❌ Redistribution or selling of this LoRA file itself is prohibited.

📌 Additional note (for reference only)

I sometimes test by writing prompts in Japanese or English and then using a translation node to convert them into Chinese (commonly used by Alibaba’s models). This is just an experiment.
For everyone else, writing prompts in English as usual is perfectly fine.