👑 MOSS-Image SFT + Krea 2 Refiner + InstantID Hybrid
The 3-Stage Ultimate Pipeline: Composition → Image Refinement → Identity Fidelity
This workflow solves the classic AI portrait dilemma: how to combine complex text understanding, crisp cinematic textures, and precise facial identity preservation in a single, automated ComfyUI workflow.
🔄 The 3-Stage Architecture
[ Stage 1: MOSS-Image SFT ] --> Scene composition, complex prompt comprehension & natural body anatomy ↓ [ Stage 2: Krea 2 Refiner ] --> 2nd-pass image enhancement, micro-contrast, cinematic lighting & texture sharpening ↓ [ Stage 3: InstantID Hybrid] --> High-fidelity face identity transfer, gender/position targeting & pore detail injection
### 1️⃣ Stage 1: Scene Generation (MOSS-Image SFT)- Powered by the Mosskva diffusion model + Qwen 3.5 9B text encoder.
- Excels at complex compositions, intricate scenes, and natural anatomical proportions.
### 2️⃣ Stage 2: Full-Canvas Image Refinement (Krea 2)
- Takes the latent output from Stage 1 into a dedicated **Krea 2 (kreros1.0)** KSampler refinement pass.
- Eliminates AI blurriness, refines clothing, hair, and background micro-textures with Krea 2’s cinematic color grading and sharp details while strictly preserving MOSS's original composition.
### 3️⃣ Stage 3: Facial Identity Consistency (InstantID Hybrid)
- Crops the subject's face with precision keypoints.
- Injects pooled multi-angle facial embeddings from your reference folder.
- Uses **LAB skin color matching** and **High-Frequency Pore Transfer hf_pore_transfer)** to seamlessly blend the refined face back into the canvas with zero visible seams or "plastic skin".
---
## 🌟 Exclusive Features & Controls
- **📁 Native Folder Picker Button**: Click the button on the node to open Windows Explorer and select your photo reference folder directly.
- **🎯 Multi-Face & Gender Targeting**:
- Filter by gender: all, female_only, male_only, or couple_both (dual-reference mode: Ref 1 = Female, Ref 2 = Male).
- Target by spatial position: largest (foreground subject), center (middle person in a group), or left-to-right ordering.
- **👅 🚬 Anatomical Protection mouth_mask_keep & nose_mask_keep)**:
- Preserves sticking-out tongues, open-mouth laughing poses, cigarettes, straws, vapes, and nose piercings without identity corruption.
- **🎛️ Built-in Presets**:
- Instant presets: default (balanced), high_fidelity (max resemblance), strong_expression, mouth_object, and mask_keep.
- Save your custom configurations with the **"💾 Salvar Configuração Atual como Preset..."** button.
---
## 🧩 Required Custom Node
Install the **ComfyUI InstantID Hybrid Face Refiner** suite:
- **Via ComfyUI Manager (Recommended)**:
- Open ComfyUI Manager, click **"Install Missing Custom Nodes"** or search for:
ComfyUI InstantID Hybrid Face Refiner
- Registry URL: https://registry.comfy.org/publishers/pesis768/nodes/comfyui-instantid-hybridrefiner
- **Via Git Clone**:
```bash
cd ComfyUI/custom_nodes
git clone https://github.com/pesis768/ComfyUI-InstantID-HybridRefiner.git
📥 Required Models
1. MOSS-Image SFT (Stage 1)
Model:
mosskva...safetensorsinmodels/diffusion_models/Text Encoder:
qwen3.5_9b_w4a8_gptq.safetensorsinmodels/text_encoders/VAE:
flux2-vae.safetensorsinmodels/vae/
2. Krea 2 Refiner (Stage 2)
Model:
kreros1.0....safetensorsinmodels/diffusion_models/Text Encoder:
qwen3vl_4b_fp8_scaled.safetensorsinmodels/text_encoders/VAE:
qwen_image_vae.safetensorsinmodels/vae/
3. InstantID Hybrid (Stage 3)
SDXL Checkpoint: Juggernaut XL, RealVisXL, or any realistic SDXL model in
models/checkpoints/InstantID Model:
ip-adapter.bininmodels/instantid/InstantID ControlNet:
instantid-controlnet.safetensorsinmodels/controlnet/InsightFace:
antelopev2models inmodels/insightface/models/antelopev2/
⚙️ Recommended KSampler Parameters
StageDenoiseStepsCFGSampler / SchedulerStage 1 (MOSS)1.0012 - 161.0 - 1.5euler / simpleStage 2 (Krea 2 Refiner)0.30 - 0.458 - 121.0euler / simpleStage 3 (InstantID)0.55 - 0.6512 - 151.0lcm / exponential
💡 Tip: Set Krea 2 refinement denoise to
0.35for subtle texture enhancement, or0.45for stronger cinematic lighting transformation!



















