# 🚀 SD Ultimate Upscaler with Nunchaku Setup Guide

This guide covers the complete setup for the enhanced SD Ultimate Upscaler workflow with Nunchaku, including Florence2 automatic captioning for optimal results.

## 🔧 Prerequisites

- ComfyUI installed and working

- NVIDIA GPU with 8GB+ VRAM (12GB+ recommended)

- CUDA 12.6+ (12.9 recommended)

## 📦 Custom Nodes Installation

### Required Nodes

- [ComfyUI_UltimateSDUpscale](https://github.com/ssitu/ComfyUI_UltimateSDUpscale)

Enables tiled upscaling for high-resolution images

- [ComfyUI-nunchaku](https://github.com/nunchaku-tech/ComfyUI-nunchaku)

Quantized Flux models for blazing speed

- [ComfyUI-Florence2](https://github.com/spacepxl/ComfyUI-Florence2)

Automatic image captioning for better upscaling prompts

- [rgthree-comfy](https://github.com/rgthree/rgthree-comfy)

Image comparison tools

- [was-node-suite-comfyui](https://github.com/WASasquatch/was-node-suite-comfyui)

Advanced save options

### Installation

- Clone all repositories into ComfyUI/custom_nodes/

- Restart ComfyUI after installation

## 📁 File Structure & Model Downloads

### Directory Layout

```

ComfyUI/

├── models/

│ ├── checkpoints/ # Nunchaku models

│ ├── clip/ # Text encoders

│ ├── vae/ # VAE models

│ ├── upscale_models/ # AI upscalers

│ └── loras/ # LoRA models (optional)

```

### 🤖 Nunchaku Models

Source: mit-han-lab/nunchaku-flux.1-dev

- RTX 20s/30s/40s: svdq-int4_r32-flux.1-kontext-dev.safetensors

- RTX 50 series: svdq-fp4_r32-flux.1-kontext-dev.safetensors

- Place in: models/checkpoints/

### 📝 Text Encoders

Source: comfyanonymous/flux_text_encoders

- t5xxl_fp16.safetensors (~9.79GB)

- ViT-L-14-TEXT-detail-improved-hiT-GmP-TE-only-HF.safetensors (~246MB)

- Place in: models/clip/

### 🎨 VAE Model

Source: black-forest-labs/FLUX.1-schnell

- flux_vae.safetensors (~335MB)

- Place in: models/vae/

### 🔍 Upscaler Models

Recommended Options:

- 4x_NMKD-Siax_200k.pth

- 4x-UltraSharp.pth

- RealESRGAN_x4plus.pth

- Place in: models/upscale_models/

### 🎯 LoRA Models (Optional)

- Photorealistic LoRAs for enhanced quality

- Place in: models/loras/Flux/

## 🧠 Florence2 Setup

### Model Download

Florence2 will auto-download the first time you run the workflow:

- Model: microsoft/Florence-2-large

- Size: ~1.5GB

- Purpose: Automatic image analysis and prompt generation

## 📋 Complete Workflow Features

### 🎨 Workflow Components

| Node | Purpose | Key Settings |

|------|---------|--------------|

| LoadImage | Input your image | Support for various formats |

| Florence2 | Auto-generate prompts | Analyzes image content automatically |

| NunchakuFluxDiTLoader | Load quantized model | INT4/FP4 variants available |

| NunchakuFluxLoraLoader | Apply LoRA enhancements | Optional for style improvement |

| UltimateSDUpscale | Main upscaling engine | Tiled processing with seam fixing |

| Image Comparer | Before/after comparison | Side-by-side visual comparison |

## ⚙️ Ultimate SD Upscaler Settings (From Actual Workflow)

### 🎯 Current Workflow Settings

| Parameter | Current Value | Purpose | Alternative Options |

|-----------|---------------|---------|-------------------|

| upscale_by | 2.0 | Upscaling factor | 4.0, 6.0, 8.0 for higher scaling |

| steps | 20 | Quality steps | 12-16 (faster), 26+ (max quality) |

| cfg | 1.0 | Must stay 1.0 | Never change for Flux models |

| denoise | 0.1 | Conservative enhancement | 0.15-0.30 for more changes |

| sampler | "euler" | Sampling method | dpmpp_2m, dpmpp_sde alternatives |

| scheduler | "normal" | Noise scheduling | karras, exponential options |

### 🧩 Tile Processing Settings

| Parameter | Current Value | Purpose | VRAM Optimization |

|-----------|---------------|---------|-------------------|

| tile_width/height | 1024 | Processing chunks | 768 (8-10GB), 512 (6-8GB) |

| mask_blur | 8 | Blend tile edges | 12-16 for softer blending |

| tile_padding | 32 | Tile overlap | 48-64 for better seams |

### 🔧 Seam Fixing Settings

| Parameter | Current Value | Purpose | When to Adjust |

|-----------|---------------|---------|----------------|

| seam_fix_mode | "Band Pass" | Fix tile boundaries | "None" for speed, "Half Tile" alternative |

| seam_fix_denoise | 1.0 | Seam blending strength | Lower if artifacts appear |

| seam_fix_width | 64 | Seam processing width | 32-128 based on image complexity |

| seam_fix_padding | 16 | Seam overlap area | Increase for complex images |

## 🎨 Florence2 Auto-Prompting

### How It Works

1. Automatic Analysis: Florence2 analyzes your input image

2. Caption Generation: Creates detailed description

3. Prompt Enhancement: Uses description for better upscaling

4. Manual Override: You can still edit the generated prompt

### Benefits

- Better Results: AI understands your image content

- Consistency: Maintains image characteristics during upscaling

- Time Saving: No need to manually describe images

- Accuracy: Preserves important visual elements

## 🚦 Optimization Presets

### ⚡ Speed Preset (Quick Preview)

```json

{

"upscale_by": 2.0,

"steps": 12,

"denoise": 0.1,

"seam_fix_mode": "None"

}

```

### ⚖️ Balanced Preset (Current Workflow)

```json

{

"upscale_by": 2.0,

"steps": 20,

"denoise": 0.1,

"seam_fix_mode": "Band Pass"

}

```

### 💎 Quality Preset (Maximum Detail)

```json

{

"upscale_by": 4.0,

"steps": 26,

"denoise": 0.2,

"seam_fix_mode": "Band Pass"

}

```

### 🔋 Low VRAM Preset (8GB GPUs)

```json

{

"tile_width": 768,

"tile_height": 768,

"steps": 16,

"seam_fix_mode": "None"

}

```

## 📱 Using the Workflow

### Step-by-Step Process

1. Load Image: Drop your image into the LoadImage node

2. Auto-Analysis: Florence2 automatically describes the image

3. Review Prompt: Check the generated caption in ShowText node

4. Adjust Settings: Modify upscaling parameters if needed

5. Run Workflow: Execute the complete pipeline

6. Compare Results: Use Image Comparer to see before/after

7. Save Output: Images automatically saved with timestamps

### 🎯 Prompt Customization

The workflow uses automatic prompting, but you can enhance it:

For Photos:

```

[Florence2 caption] + highly detailed, sharp focus, professional photography, enhanced details

```

For Art:

```

[Florence2 caption] + vibrant colors, sharp lines, professional quality, enhanced details

```

## 🔧 Troubleshooting

### 💥 Memory Issues

- Reduce tile size to 768 or 512

- Lower steps to 12-16

- Disable seam fixing temporarily

- Use smaller upscale factor (2x instead of 4x)

### 🧩 Visible Seams

- Keep "Band Pass" seam fixing enabled

- Increase tile_padding to 48-64

- Increase mask_blur to 12-16

- Check seam_fix_width is adequate (64-128)

### 🐌 Slow Performance

- Reduce steps to 12-16

- Set seam_fix_mode to "None"

- Use 2x upscaling instead of 4x+

- Increase tile size if VRAM allows

### 🎭 Florence2 Issues

- First run takes longer (model download)

- Restart ComfyUI if Florence2 fails to load

- Check internet connection for initial download

## ⏱️ Performance Benchmarks

| GPU | VRAM | Settings | Processing Time |

|-----|------|----------|-----------------|

| RTX 4090 | 24GB | 2x upscale, 1024px tiles, 20 steps | 15-30s |

| RTX 4080 | 16GB | 2x upscale, 1024px tiles, 16 steps | 25-45s |

| RTX 3080 | 10GB | 2x upscale, 768px tiles, 16 steps | 45-75s |

| RTX 3070 | 8GB | 2x upscale, 768px tiles, 12 steps | 60-90s |

## 🎁 Workflow Advantages

### 🚀 Speed Benefits

- Nunchaku quantization: 3-5x faster than standard Flux

- Smart tiling: Efficient memory usage

- LoRA integration: Enhanced quality without speed loss

### 🎯 Quality Features

- Florence2 analysis: Intelligent prompt generation

- Seam fixing: Professional tile blending

- Comparison tools: Easy before/after evaluation

- Flexible scaling: 2x to 8x upscaling options

### 🛠️ User Experience

- Automatic workflow: Minimal manual input required

- Visual feedback: Image comparison built-in

- Organized output: Timestamp-based file naming

- Memory efficient: Works on consumer GPUs

***

🎯 This enhanced workflow combines Nunchaku's speed, Florence2's intelligence, and Ultimate SD Upscaler's quality for professional results with minimal setup