Workflow Overview

This workflow, titled “Aolun Red Carpet Workflow,” leverages the Flux.1 model to generate high-quality Chinese mythology-style artistic images. The theme centers on a young woman (inspired by “Aolun”), dressed in celestial Chinese fairy attire, set against a red carpet or moonlit scene. The workflow’s features and goals include:

• Theme Generation: Creates art based on detailed Chinese prompts, covering themes, scenes, attire, dynamics, and atmosphere.

• Style Enhancement: Uses multiple Lora models (e.g., “X敖闰西海龙王-国漫女神” and “猫爪_动漫国风美人颜”) to boost Chinese aesthetics and detail quality.

• Resolution Flexibility: Supports custom resolutions (e.g., 1352x2048), ideal for movie stills.

• Multilingual Support: Translates Chinese prompts to English via a Baidu Translate node for improved compatibility. This workflow suits Chinese mythology art, movie concept design, or high-resolution digital illustrations.

Core Models

1. Flux.1 (基础算法_F.1)

   • Function: An efficient text-to-image model for generating high-resolution artistic images.

   • Source: Download from Civitai (e.g., 基础算法_F.1_fp8_e4m3fn.safetensors), place in ComfyUI/models/checkpoints/.

2. Lora Models

   • X敖闰西海龙王-国漫女神-哪吒之魔童闹海_F.1_v0.6: Enhances national comic style and “Aolun” character traits, weight 0.8.

   • F.1--细节质感提升--F.1 DETAILER_V1.0: Boosts image detail and texture, weight 0.35.

   • 猫爪_动漫国风美人颜_FLUX_1.0: Adds anime Chinese beauty features, weight 0.35.

   • 【高定礼服】绪儿 时尚街拍滤镜V3_V3: Optimizes attire and fashion sense, weight 0.15.

   • Source: Download from Civitai, place in ComfyUI/models/loras/.

3. VAE Model (ae.sft)

   • Function: Decodes latent images into pixel images, ensuring high-resolution output.

   • Source: Download from Flux.1-related repositories, place in ComfyUI/models/vae/.

Component Explanation

Below are the key nodes in the workflow, their purposes, functions, and installation methods:

1. UNETLoader

   • Purpose: Loads the Flux.1 base model.

   • Function: Outputs a model object to drive generation.

   • Installation: Built into ComfyUI.

2. DualCLIPLoader

   • Purpose: Loads the CLIP model.

   • Function: Outputs a CLIP object to encode text prompts.

   • Installation: Built into ComfyUI.

   • Dependencies: Requires t5xxl_fp8_e4m3fn and clip_l files.

3. LoraLoader (Multiple)

   • Purpose: Loads multiple Lora models for fine-tuning.

   • Function: Sequentially enhances national comic style, detail, anime beauty, and fashion.

   • Installation: Built into ComfyUI.

4. CLIPTextEncode

   • Purpose: Encodes positive text prompts.

   • Function: Converts detailed prompts into conditioning data with a guidance scale of 3.5.

   • Installation: Built into ComfyUI.

5. BaiduTranslateNode

   • Purpose: Translates Chinese prompts to English.

   • Function: Optimizes prompt compatibility with English output.

   • Installation: Requires BaiduTranslate plugin, install via ComfyUI Manager or GitHub.

6. EmptyLatentImage

   • Purpose: Creates an initial latent image.

   • Function: Generates a custom-resolution latent space (e.g., 1352x2048).

   • Installation: Built into ComfyUI.

7. RandomNoise

   • Purpose: Generates random noise.

   • Function: Serves as the sampling input with a seed of 177677572491923.

   • Installation: Built into ComfyUI.

8. BasicScheduler

   • Purpose: Calculates sampling steps and timesteps.

   • Function: Uses a “simple” scheduler with 20 steps.

   • Installation: Built into ComfyUI.

9. KSamplerSelect

   • Purpose: Selects the sampler.

   • Function: Outputs the “euler” sampler.

   • Installation: Built into ComfyUI.

10. SamplerCustomAdvanced

    • Purpose: Performs advanced sampling.

    • Function: Combines noise, guidance, and latent image to generate results.

    • Installation: Built into ComfyUI.

11. BasicGuider

    • Purpose: Combines model and conditioning.

    • Function: Outputs a guider object.

    • Installation: Built into ComfyUI.

12. FluxGuidance

    • Purpose: Applies Flux guidance.

    • Function: Enhances conditioning effect with a guidance scale of 3.5.

    • Installation: Built into ComfyUI.

13. VAELoader

    • Purpose: Loads the VAE model.

    • Function: Outputs a VAE object.

    • Installation: Built into ComfyUI.

14. VAEDecode

    • Purpose: Decodes the latent image.

    • Function: Generates a pixel image.

    • Installation: Built into ComfyUI.

15. PreviewImage

    • Purpose: Previews the generated image.

    • Function: Displays the output in real-time.

    • Installation: Built into ComfyUI.

Workflow Structure

1. Model Loading and Fine-Tuning Group

   • Role: Loads the Flux.1 model and applies multiple Lora fine-tunings.

   • Input Parameters: Model file (基础算法_F.1), Lora files, and weights.

   • Output: Fine-tuned model and CLIP object.

2. Prompt Processing Group

   • Role: Translates and encodes prompts.

   • Input Parameters: Chinese prompt (translated to English via BaiduTranslate).

   • Output: Encoded conditioning data.

3. Generation Preparation Group

   • Role: Creates latent image and sampling parameters.

   • Input Parameters: Resolution (1352x2048), noise seed, sampling steps (20).

   • Output: Latent image and sampling conditions.

4. Sampling and Generation Group

   • Role: Executes image generation.

   • Input Parameters: Fine-tuned model, conditioning, noise, sampler (Euler).

   • Output: Latent image.

5. Decoding and Display Group

   • Role: Decodes and previews the image.

   • Input Parameters: Latent image, VAE model.

   • Output: High-resolution pixel image.

Inputs and Outputs

• Expected Inputs:

  • Resolution: 1352x2048 (adjustable to 1024x1536 or 1080x1920).

  • Seed: 177677572491923 (randomizable).

  • Prompt: Detailed Chinese description (provided, covering theme, scene, attire, etc.).

  • Negative Prompt: Not set (suggest adding “Worst quality, blurry”).

  • Sampling Steps: 20.

  • Guidance Scale: 3.5.

  • Batch Size: 1 (adjustable up to 4).

• Final Output:

  • High-quality artistic image (PNG format, 1352x2048 resolution, themed around a Chinese mythology woman).

Notes and Tips

1. Resource Requirements: High-resolution generation requires 16GB+ VRAM; an NVIDIA GPU is recommended.

2. Model Files: Ensure 基础算法_F.1_fp8_e4m3fn and ae.sft are in the correct paths.

3. Plugin Installation: Requires rgthree (includes Fast Groups Bypasser) and BaiduTranslate plugins.

4. Prompt Optimization: Replace or modify prompts as needed, maintaining structured descriptions.

5. Performance Optimization: Reduce resolution (1352x2048 → 1024x1536) or steps (20 → 10) for faster generation.

Example Illustration

• Input: Prompt (translated) “A high-quality digital painting movie still... wearing an ice-blue mermaid-style dress...”.

• Process:

  1. Load Flux.1 model and apply Lora.

  2. Translate and encode the prompt.

  3. Generate a 1352x2048 latent image.

  4. Decode into a pixel image.

  5. Preview and save.

• Output: A movie still depicting a Chinese mythology woman, with an ice-blue hanfu glowing under moonlight, rich in detail.