1. Workflow Overview

• Purpose: Generates high-fidelity images with minimal "AI-like" artifacts through multi-stage sampling, super-resolution, and progressive denoising.

• Key Techniques:

  • FP8 Low-Precision Model: Uses FLUX_FB8V1.safetensors (FP8 E4M3) for memory efficiency.

  • Two-Stage KSampler: Base generation → Progressive denoising refinement.

  • Hybrid Upscaling: Combines RealESRGAN and latent space scaling (LatentUpscaleBy).

2. Core Models & Nodes

3. Workflow Structure

Group 1: Model Loading & Text Encoding

• Input: Positive prompt (e.g., dragon girl...), negative prompt (auto-handled).

• Output: Text conditioning.

• Key Nodes:

  • DualCLIPLoader: Enhances prompt understanding.

  • LoraLoaderModelOnly: Applies LoRA 熬润.safetensors (weight=0.7).

Group 2: Base Image Generation

• Input: Resolution 768x1152, seed=randomize, sampler=DEIS.

• Output: Initial latent image.

• Key Nodes:

  • KSampler (Efficient): CFG=0.95, denoise=0.6.

  • VAEDecode: Decodes latent for preview.

Group 3: Super-Resolution & Refinement

1. ImageUpscaleWithModel: 4x upscale via RealESRGAN.

2. VAEEncodeTiled: Re-encodes upscaled image to latent space.

3. LatentUpscaleBy: Secondary latent scaling (nearest-exact).

4. Progressive KSampler: Denoise strength reduced (0.4→0.35).

Group 4: Result Comparison

• Tool: Image Comparer for before/after sliding.

4. Inputs & Outputs

• Inputs:

  • Required: Positive prompt, resolution (768x1152 recommended).

  • Optional: Seed (default random), LoRA weight (default=0.7).

• Outputs:

  • Final image (PNG with metadata).

  • Comparison slides (via Image Comparer).

5. Notes

1. VRAM: ≥10GB GPU (e.g., NVIDIA 30/40 series).

2. Compatibility: Disable xformers for FP8 models.

3. Troubleshooting:

   • CLIP encode failed → Verify t5xxl_fp8 model path.

4. Optimization: Use Tiled VAE plugin to reduce VRAM usage.