1. Workflow Overview

This workflow, named "Crystal Palace Stylization", transforms input images into fantasy visuals with crystal-like aesthetics, emphasizing translucent materials, light refraction, and intricate details. Key features:

• Depth map generation via Depth Anything V2

• Crystal texture enhancement using LoRAs (e.g., 剪纸8.safetensors)

• Two-stage sampling (base + refinement) for quality

• Outputs high-resolution (e.g., 1536x1536) fantasy-style images

Core Models:

• Stable Diffusion: Base model (flux1-depth-dev.safetensors)

• Depth Anything V2: Depth estimation

• Custom LoRA: 剪纸8.safetensors (weight=0.8)

• VAE: ae.sft (enhances details/colors)

2. Node Details

Key Nodes:

1. DepthAnything_V2: Generates depth maps

   • Install: Via ComfyUI Manager (search "Depth Anything")

   • Model: depth_anything_v2_vitl_fp32.safetensors (auto-downloaded)

2. InstructPixToPixConditioning: Merges depth maps with text prompts

   • Input: Depth map + prompts (e.g., "Crystal World, 4K details")

   • Output: Conditioned latent data

3. LoraLoaderModelOnly: Loads crystal-style LoRA

   • Path: models/loras/

   • Manual setup required for LoRA files

4. KSamplerAdvanced: Advanced sampler

   • Parameters: Euler, 20 steps, CFG=14 (high-detail refinement)

5. FluxGuidance: Dynamic guidance scaling (guidance_scale=30)

3. Workflow Structure

Group 1: Fixed Process

• Input: Source image (e.g., b463ddb8...jpg)

• Process:

  1. Load image via LoadImage

  2. Generate depth map with DepthAnything_V2

  3. Resize to 1536x1536 (ImageResizeKJ)

• Output: Depth map + normalized image

Group 2: Conditioning

• Prompts:

  • Positive: "Crystal Palace theme, glitter, 4K resolution"

  • Negative: Default (cleared via ConditioningZeroOut)

• Process: CLIP text encoding

Group 3: Generation & Refinement

• Base Sampling: KSampler (15 steps, CFG=7.0)

• Refinement: KSamplerAdvanced (20 steps, CFG=14)

• Output: VAEDecode converts latent to final image

4. Inputs & Outputs

• Inputs:

  • Image: Any size (recommended ≥512x512)

  • Prompts: Describe crystal style (e.g., "Fantasy, light refraction")

  • Seed: Optional fixed value (e.g., 939252204466143)

• Output:

  • Path: ComfyUI/output/

  • Format: PNG (HD crystal-themed image)

5. Notes

1. VRAM: ≥10GB GPU (tiling recommended for 1536x1536)

2. Dependencies:

   • Verify LoRA and Depth Anything models are installed

3. Troubleshooting:

   • Missing models: Check models/loras/ and models/depth_anything/

   • OOM errors: Reduce output resolution

4. Optimization:

   • Enable xformers

   • Decrease KSamplerAdvanced steps for speed