1. Workflow Overview

• Purpose:
  A high-speed image imitation & optimization workflow based on Nunchaku Flux, featuring:

  • Image-to-prompt (via JoyCaption).

  • Fast image generation (using quantized FluxDiT).

  • Image scaling & post-processing (e.g., super-resolution).

  • GPU cleanup (easy cleanGpuUsed).

• Core Models:

  • Nunchaku FluxDiT: Quantized model (svdq-int4-flux) for speed.

  • CLIP Encoder: t5xxl_fp8_e4m3fn.safetensors + clip_l.safetensors.

  • Upscaler: 4x-UltraSharp.pth.

  • VAE: ae.safetensors.

2. Key Nodes

• JoyCaption: Uses Meta-Llama-3.1-8B to generate prompts from images.

• FluxGuidance: Controls conditioning strength (default 3.5).

• ModelSamplingFlux: Configures resolution (1024x1024) and quantization.

• KSampler: Generates images in 20 steps with euler sampler.

• ImageScaleByAspectRatio: Resizes images with lanczos interpolation.

3. Workflow Structure

1. Group 1: Nunchaku Text-to-Image

   • Input: Prompt (auto-generated or manual).

   • Process: CLIP encode → FluxDiT → VAE decode → Save.

   • Output: Final image (e.g., cyberpunk cat).

2. Group 2: Image Loading & Prompt Reverse

   • Input: Uploaded image (e.g., 2794c263...jpg).

   • Process: Upscale → Caption → Pass to Group 1.

   • Output: Text description (e.g., "a neon cat").

4. Inputs & Outputs

• Inputs: Image (optional), resolution (1024x1024), steps (20).

• Outputs: Image saved to ComfyUI/output.

5. Notes

• Dependencies:

  • Install Nunchaku UI and Meta-Llama-3.1-8B.

• Errors:

  • Missing model files: Check paths for t5xxl_fp8_e4m3fn.safetensors.

• Optimization:

  • Use svdq-int4 for lower VRAM usage.