1. Workflow Overview

This workflow leverages FLUX architecture and inverse sampling for high-precision image reconstruction, with features:

• Inverse Sampling: Mathematically optimizes latent space to edit details while preserving structure.

• FLUX Model Control: Dynamic noise scheduling and conditioning for controllable outputs.

• Style Transfer & LoRA: Supports custom LoRAs (e.g., anthropomorphic cats) and CLIP text guidance.

• Three-Stage Process: Initial Sampling → Inverse Sampling → Regeneration for complex edits.

2. Core Models

3. Key Nodes

4. Workflow Structure

1. Initial Sampling

   • Input: Original image → ImageResize+ to 1024x1024.

   • VAEEncode converts image to latent space.

   • BasicScheduler sets initial noise (28 steps, simple mode).

2. Inverse Sampling

   • FluxReverseODESampler reconstructs latents (eta=0.84, higher = more original).

   • FluxDeGuidance adjusts style strength (default 3.5).

3. Regeneration

   • FluxForwardODESampler generates new image.

   • LoRA Loader injects custom features (e.g., cat anthropomorphism).

   • VAEDecode saves final output.

5. Input & Output

• Input Parameters:

  • Image: Upload any image (e.g., animal photo), auto-resized to 1024x1024.

  • Prompts: Use LibLibTranslate (e.g., "老虎" → "tiger").

  • Inverse Sampling: eta (0~1, similarity to original), steps (default 28).

• Output: One 1024x1024 edited image (PNG).

6. Notes

• Hardware: Minimum 16GB VRAM (inverse sampling is VRAM-intensive).

• Tuning Tips:

  • eta: Near 1 preserves original, near 0 enables creativity.

  • FluxDeGuidance: >3 for strong style transfer, <2 for prompt-driven results.

• Troubleshooting:

  • VRAM errors: Reduce steps or switch to fp16 precision.

  • Sampling fails: Ensure input resolution is divisible by 2 (e.g., 512x512).