Asian Beauty — MiniMax H3-Image
Novarchi · Creative Models
An always-on beauty LoRA for MiniMax-H3-Image. It lifts East Asian portraits toward a polished Korean / Japanese / Chinese idol look: refined features, fair skin with real texture, slimmer figures, cleaner poses and more confident expressions.
It keeps your prompt. Scene, wardrobe, light and composition stay what you wrote; only the subject and the styling get the lift. Trained on glamour, lingerie and nude photography, so sensual prompts work out of the box.
Quick start
Trigger: none, the LoRA is always on while loaded
Weight: 1.0 · 0.6–0.8 for a softer touch
Sampler: 6 steps, CFG 1, LCM or euler, simple scheduler
Size: up to 16 MP, e.g. 3:4 at 3552 × 4736
Prompting tips
Name the ethnicity (“a 24-year-old Korean woman…”). Captions were written with explicit ethnicity, so the LoRA stays controllable.
Non-Asian prompts stay non-Asian. A Russian woman with chestnut hair stays European at full weight.
Want real skin? Avoid “cover photograph” / “magazine cover”. They push H3 toward airbrushed gloss. Describe a camera and real light instead: candid photograph, full-frame camera, prime lens, slight film grain, visible pores, no retouching.
ComfyUI: with the default dynamic VRAM, H3-Image LoRAs load in bf16 and lose about 15 % of their strength. Start ComfyUI with
--disable-dynamic-vramor use about 1.15.
Showcase
10 scenes at weight 1.0, LCM 6 steps, 3:4 16 MP: Seoul rooftop in red satin · Kyoto ryokan yukata · 1930s Shanghai qipao · Tokyo hotel lingerie under neon · Hong Kong after the rain · morning after in the kitchen · office after hours · Bali villa · studio silk nude · snowy onsen at night. Every image carries its full prompt, seed and workflow in the PNG metadata.
License
Derived from MiniMax-H3-Image and subject to the MiniMax-H3-Image Community License and its Acceptable Use Policy. All depicted people are fictional adults (20+). Label posted images as AI-generated. No minors and no impersonation of real people.
Description
## v1 — step 6000
### Training details
**Base model** | MiniMax-H3-Image (pruned, fp8_scaled) |
**Trainer** | ostris AI Toolkit · fp8 DiT kept quantized · nvfp4 Qwen3-VL text encoder |
**Network** | LoRA · rank 32 / alpha 32 · attention + MLP of all 50 blocks (AdaLN excluded) |
**Optimizer** | AdamW 8-bit · LR 4e-5 constant · batch 1 · EMA 0.998 |
**Dataset** | 6,218 curated photographs
**Content** | East Asian majority, ~26 % European control · fashion, swimwear, lingerie, glamour, nude · cosplay ≈ 6 % |
**Resolution** | aspect buckets 1024 / 1280 px |
**Captions** | detailed natural language (Grok 4.7): ethnicity, age, body and breast size, wardrobe, pose, light · 5 % dropout |
**Curation** | near-duplicates, censored frames and anything youth-coded removed |
**Steps / epochs** | trained 8,000 steps (≈ 1.13 epochs, 7,108 samples per epoch) · **released: step 6,000 ≈ 0.84 epoch** |
**Hardware** | 1 × RTX 5090 32 GB · ~2.2 s/step |
### Why step 6000
4000, 6000 and 8000 were compared with base H3 at 16 MP on six fresh scenes with fixed seeds. The look sets in around 4000 and plateaus after it; 6000 ≈ 8000, and 6000 is the best balance. For the lighter 4000-style touch, lower the weight.
### Notes
- Faces lean toward a K-/J-pop idol ideal. Use 0.6–0.8 for plainer, everyday faces.
- Bodies trend slim and toned. At full weight, swimwear and fitness poses can show defined abs.
- Skin is smoother and dewier than base H3 but keeps pores. Prompt wording matters more than weight.
- Legs-and-skirt fashion styling is cleaner than base, but this is not a dedicated fashion LoRA.
- No real identity is baked in. Faces change with the seed.



















