1. Workflow Overview

This workflow specializes in style transfer and content reconstruction using:

• Inverse Sampling: Encodes reference images to latent space via VAE

• Dynamic Noise Control: DisableNoise + FlipSigmas for noise inversion

• Flux Stack: Bidirectional sampling with ODE solvers

• Prompt Guidance: Modify content via text (e.g., replace "cat" with "elephant")

2. Core Models

3. Key Nodes

4. Pipeline Stages

Stage 1: Reference Processing

• Input: Upload image (e.g., da5452ac-33d5-4d88-b08e-7f8a6eb327dc.png)

• Key Steps:

  • ImageResize+ to 1024x1024

  • VAEEncode to latent space

Stage 2: Inverse Sampling Core

• Tech Stack:

  • FlipSigmas noise inversion

  • FluxReverseODESampler (CFG=3.5, linear decay)

  • BasicGuider for conditioning

Stage 3: Content Generation

• Prompt Editing: Modify keywords in CLIPTextEncode

• Output: Decode via VAEDecode

5. Inputs & Outputs

• Required Inputs:

  • Reference image (auto-resized)

  • Custom prompts (bilingual support)

• Outputs:

  • 1024x1024 generated image

  • Real-time preview

6. Critical Notes

1. Dependencies:

   • Download FP8 BaseAlgo_F.1 and ae.sft

   • Flux requires Python 3.10+

2. Hardware:

   • Minimum: 8GB VRAM

   • Recommended: RTX 3060+

3. Troubleshooting:

   • FP8 not supported → Upgrade CUDA to ≥11.8

   • Blurry results → Adjust CFG in FluxDeGuidance (2.5-4.0)