Qwen-Image 2.1 — Krea2 35mm Film Texture & Realism LoRA: v1.0: krea2_film_texture_qwen21_r64_000002500.safetensors

ver1.0

## 🎞️ Krea2 35mm Film Texture & Cinematic Realism LoRA for Qwen-Image 2.1

This LoRA brings **authentic 35mm motion-picture film texture, natural skin pores, tactile fabric weaves, and organic highlight/shadow roll-off** to **Qwen-Image 2.1**, while keeping **100% of Qwen-Image 2.1's native capabilities** (Text-to-Image, single & multi-image Edit / I2I, character consistency, and text rendering).

Vanilla Qwen-Image 2.1 has top-tier prompt adherence and editing power, but often produces overly smooth plastic skin, synthetic surface sheen, or a digital yellow-green cast. This LoRA removes that artificial digital look and replaces it with grounded, photorealistic cinema film texture — **with zero trigger words required (100% Plug & Play).**

---

### ✨ Key Features

* **Plug & Play (No Trigger Word Needed)**: Trained directly across all 384 diffusion transformer blocks. Simply load the LoRA and write your normal prompts.

* **Zero Loss of Native Capabilities**: Works seamlessly with both **T2I** and **Multi-Reference Image Editing TextEncodeQwenImage21)** without breaking character identity or prompt adherence.

* **Natural Skin & Surface Micro-Texture**: Restores realistic skin pores, fine hair strands, wool/tweed/velvet fabric weaves, leather grain, and stone/wood surfaces.

* **Organic 35mm Film Color & Lighting Roll-off**: Eliminates the flat digital/plastic look, adding rich shadow depth, natural halation, and fine 35mm film grain.

* **Trained at Native 1920×1088 (16:9 FHD)**: Trained on **2,951 VLM-audited cinematic stills** at native 1920×1088 resolution (zero resizing or cropping) with captions 100% aligned to the official Qwen-Image 2.1 prompting structure.

---

### ⚙️ Recommended Settings

| Parameter | Recommended Value | Notes |

| :--- | :--- | :--- |

| **Trigger Word** | **None Required** | Just load the LoRA and use your normal prompt |

| **LoRA Strength (I2I / Edit)** | *0.50** | Best balance for Character Sheets & Multi-Image I2I (locks 100% facial identity) |

| **LoRA Strength (T2I)** | *0.50 – 0.65** | 0.50 for subtle natural realism, 0.60–0.65 for stronger 35mm film mood |

| **Base Model** | Qwen-Image 2.1 | Fully compatible with both bf16 and int8_convrot weights |

| **ComfyUI Placement** | LoraLoaderModelOnly | Connect right after *Load Diffusion Model** (before QwenImage21Cache / ModelSamplingFlux) |

| **Sampler / Scheduler** | euler / simple | Official Qwen-Image 2.1 flow-matching schedule |

| **Steps / CFG** | 40 steps / CFG 2.0 | Use ModelSamplingFlux max_shift: 0.693548, base_shift: 0.5) |

| **Resolution** | 1920×1088 / 2752×1536 (2K) | Works across all aspect ratios 16:9, 8:9, 3:4, 9:16, 1:1) |

---

### 🧪 Training Specs

* **Dataset**: 2,951 curated cinematic & editorial frames at native 1920×1088 (filtered from 3,000 generations via automated VLM quality audit)

* **Architecture**: Rank 64 linear: 64, linear_alpha: 32), bf16, timestep_bias: low_noise, targeting all 384 transformer LoRA modules (including img_mlp)

---

### 🇰🇷 한국어 안내 (Korean Guide)

**Qwen-Image 2.1**의 모든 본래 기능(T2I 생성, 단일·다중 참조 이미지 편집/합성, 캐릭터 일관성, 텍스트 묘사력)을 그대로 유지하면서, 순정 모델 특유의 매끈한 플라스틱 피부결과 디지털 색감을 **실제 35mm 영화 필름 카메라의 피부 모공·원단 질감·깊은 암부 계조와 필름 입자감**으로 바꿔주는 LoRA입니다.

* **트리거 워드**: **불필요** (별도 트리거 워드 없이 LoRA 연결만으로 즉시 작동합니다)

* **추천 강도 strength_model)**:

* *0.50 (기본 권장 / I2I 캐릭터 시트·다중 이미지 편집용)**: 원본 얼굴 인상을 100% 그대로 유지하면서 자연스러운 피부결과 필름 톤 적용

* *0.55 ~ 0.65 (T2I 단독 화보·영화 스틸컷용)**: 한층 진한 35mm 필름 질감과 시네마틱 무드 연출

* **ComfyUI 연결 위치**: *Load Diffusion Model (디퓨전 모델 로드)** → *LoraLoaderModelOnly (0.50)** → QwenImage21Cache → ModelSamplingFlux