Unlock Stunning Portraits: Advanced AI Workflow Revealed
1. Workflow Overview
Purpose: Generate hyper-realistic portraits with optimized lighting and skin details, suitable for photography and advertising.
Core Models:
Base Model:
基础算法_F.1(FP8 optimized).LoRAs:
F.1-光影氛围-摄影光影-写真-氛围_极致的光影效果_2.0(strength=0.8 for lighting).仟士-真实人像-眸渊·光漪 - Portraiture_v1.0(strength=0.4 for skin texture).
VAE:
ae.sft(color enhancement).CLIP:
t5xxl_fp8_e4m3fn(multimodal text encoder).
Key Features:
Auto-Captioning: Uses
Joy_caption_twoto generate prompts from input images.Multi-Stage Processing: Initial sampling → Upscaling → Film simulation → Color grading.
2. Key Nodes
Node | Function | Installation |
|---|---|---|
| Loads captioning model ( | Install |
| Tile-based 2x upscaling (uses | Via |
| Adds film grain and dynamic range (adjustable parameters). | Requires |
| Generates depth maps for post-processing DOF effects. | Install |
Dependencies:
LoRA Models: Download and place in
models/loras.Captioning Model: Download
Meta-Llama-3.1-8B-Instruct-bnb-4bit(~4GB).
3. Workflow Groups
Group | Function | Inputs/Outputs |
|---|---|---|
1-FLUX Model Setup | Loads base model, LoRAs, and VAE; configures Flux sampling. | Input: Model files → Output: Initialized model. |
2-Image Basics | Sets resolution (768×1280) and prompts (manual/auto-generated). | Input: Text/image → Output: Conditioned tensors. |
3-Auto-Captioning | Generates descriptions from input images (e.g., clothing, pose). | Input: Image → Output: Text description. |
4-Initial Sampling | Generates low-res images (CFG=3.5, steps=20). | Input: Latent → Output: Low-res image. |
5-SD Upscaling | Tile-based 2x upscaling to avoid VRAM overload. | Input: Low-res image → Output: 2x image. |
6-Post-Processing | Sharpening, color correction, and depth compositing. | Input: Upscaled image → Final output. |
7-Film Simulation | Applies grain and dynamic range (adjustable). | Input: Image → Output: Film-style image. |
4. Inputs & Outputs
Input Parameters:
Image (optional): Upload via
LoadImagenode (e.g.,27775044-882e3cdb...png).Prompt (optional): Manually entered or auto-generated.
Resolution: Default
768×1280, upscaled to1536px.
Output:
HD portrait (with lighting/skin enhancements), saved as
MoYou.jpg.Intermediate results previewable via
PreviewImagenodes.
5. Tips & Warnings
VRAM: ≥12GB GPU recommended for upscaling.
Common Errors:
Captioning fails → Verify
Meta-Llamamodel path.Overly strong grain → Adjust
grain_strengthinLayerFilter: FilmV2(default=0.05).
Optimizations:
Reduce
FluxGuidanceCFG (e.g., 3.5→2.5) if oversharpened.Disable upscaling during testing for faster iterations.