Pipeline Guide

360 Hextile V2 supports seven AI processing pipelines, each optimized for different use cases. This guide helps you choose the right pipeline and understand how they work.

Available Pipelines

1. Stable Diffusion 1.5

Best for: Lower VRAM systems and extensive ControlNet/LoRA ecosystem

Technical Details: - Model: Stable Diffusion 1.5 Inpainting - Type: Tiled (uses hexagonal projection) - VRAM: ~4-6GB - Native tile size: 512×512 - Processing: Sequential or batch modes

Features: - ✅ Positive and negative prompts - ✅ Extensive ControlNet support (8 types: Canny, Depth, OpenPose, Scribble, MLSD, Seg, Normal, Lineart) - ✅ LoRA support (largest ecosystem) - ✅ Cardinal direction prompts (6 directions) - ✅ Template-based hexagonal tiling - ✅ Low VRAM requirements

When to Use: - Have limited GPU memory (4-6GB) - Need ControlNet for structural guidance - Want access to vast LoRA library - Processing on older or mid-range GPUs

Example Configuration:

{
  "pipeline": "sd15_inpaint",
  "model": "stable-diffusion-v1-5/stable-diffusion-inpainting",
  "template": "Hextile_32_3K",
  "global_prompt": "fantasy landscape with magical elements",
  "negative_prompt": "blurry, low quality, distorted",
  "strength": 0.8,
  "guidance_scale": 7.5,
  "num_inference_steps": 30,
  "controlnet": {
    "enabled": true,
    "type": "canny",
    "strength": 0.85
  }
}

2. Stable Diffusion 2.1

Best for: Higher resolution tiles with ControlNet support

Technical Details: - Model: Stable Diffusion 2.1 Inpainting - Type: Tiled (uses hexagonal projection) - VRAM: ~4-6GB (512px) or ~6-8GB (768px) - Native tile size: 512×512 (Inpainting) or 768×768 (Base) - Processing: Sequential or batch modes

Features: - ✅ Positive and negative prompts - ✅ ControlNet support (4 types: Canny, Depth, OpenPose, Scribble) - ✅ LoRA support (SD2.1-specific) - ✅ Cardinal direction prompts (6 directions) - ✅ 768px native resolution option - ✅ Template-based hexagonal tiling

When to Use: - Need higher resolution tiles than SD1.5 - Want improved image quality over SD1.5 - Need ControlNet with moderate VRAM - Using SD2.1-specific LoRAs

Model Variants: - SD 2 Inpainting (512px): Best for hexagonal tiles, lower VRAM - SD 2.1 Base (768px): Higher quality, more VRAM required

Example Configuration:

{
  "pipeline": "sd21_inpaint",
  "model": "sd2-community/stable-diffusion-2-inpainting",
  "template": "Hextile_32_3K",
  "global_prompt": "realistic landscape with mountains",
  "negative_prompt": "blurry, low quality, distorted",
  "strength": 0.8,
  "guidance_scale": 7.5,
  "num_inference_steps": 30,
  "controlnet": {
    "enabled": true,
    "type": "depth",
    "strength": 0.85,
    "preprocessor": {
      "depth": {
        "boost": 1.5
      }
    }
  }
}

3. SDXL Inpainting

Best for: General-purpose 360° AI generation with proven quality, no ControlNet needed

Technical Details: - Model: Stable Diffusion XL 1.0 or SDXL-Turbo - Type: Tiled (uses hexagonal projection) - VRAM: ~12GB (both models) - Processing: Sequential or batch modes - Speed: Moderate (SDXL) or Fast (Turbo)

Features: - ✅ Positive and negative prompts - ✅ LoRA support - ✅ Cardinal direction prompts (6 directions) - ✅ Strength, guidance scale, inference steps controls - ✅ Template-based hexagonal tiling - ✅ NEW: Model selection (SDXL 1.0 or SDXL-Turbo) - ❌ No ControlNet (not in current version)

When to Use: - Standard 360° image enhancement - Want negative prompt control - Familiar with Stable Diffusion parameters - Need proven, stable results - Want fast iteration (use Turbo model)

Example Configuration (Standard SDXL 1.0):

{
  "pipeline": "sdxl_inpaint",
  "model": "diffusers/stable-diffusion-xl-1.0-inpainting-0.1",
  "template": "Hextile_32_3K",
  "global_prompt": "vibrant landscape with mountains and lake",
  "negative_prompt": "blurry, low quality, distorted",
  "strength": 0.66,
  "guidance_scale": 7.5,
  "num_inference_steps": 50,
  "interpolation": "6CardinalDirections",
  "cardinal_prompts": {
    "front": "mountain peaks",
    "back": "serene lake",
    "left": "pine forest",
    "right": "rocky cliffs",
    "above": "clear blue sky",
    "below": "lush grass"
  }
}

Example Configuration (SDXL-Turbo for Speed):

{
  "pipeline": "sdxl_inpaint",
  "model": "local/sdxl-turbo",
  "template": "Hextile_32_3K",
  "global_prompt": "vibrant landscape with mountains and lake",
  "negative_prompt": "blurry, low quality, distorted",
  "strength": 0.66,
  "guidance_scale": 0.0,
  "num_inference_steps": 1,
  "interpolation": "6CardinalDirections"
}

4. SD3.5 Inpainting

Best for: Latest Stable Diffusion technology with quality/speed modes

Technical Details: - Model: Stable Diffusion 3.5 Large or Large-Turbo - Type: Tiled (uses hexagonal projection) - VRAM: ~12-16GB - Processing: Sequential or batch modes - Speed: Turbo is very fast, Large is moderate

Features: - ✅ Model selection (Turbo for speed, Large for quality) - ✅ LoRA support (SD3-specific LoRAs) - ✅ Cardinal direction prompts (6 directions) - ✅ Native 1024×1024 resolution support - ✅ Template-based hexagonal tiling - ✅ Auto-scheduler switching (Turbo vs Standard) - ❌ No negative prompt (SD3 doesn't use them effectively) - ❌ No ControlNet (not supported)

When to Use: - Want latest SD technology (MMDiT architecture) - Need fast iteration with Turbo mode - Want quality mode for final renders - Don't need negative prompts - Have 12-16GB VRAM available

Example Configuration (SD3.5-Turbo for Speed):

{
  "pipeline": "sd35_inpaint",
  "model": "stabilityai/stable-diffusion-3.5-large-turbo",
  "template": "Hextile_32_3K",
  "global_prompt": "vibrant landscape with mountains and lake",
  "strength": 0.66,
  "guidance_scale": 1.0,
  "num_inference_steps": 4,
  "interpolation": "6CardinalDirections",
  "cardinal_prompts": {
    "front": "mountain peaks",
    "back": "serene lake",
    "left": "pine forest",
    "right": "rocky cliffs",
    "above": "clear blue sky",
    "below": "lush grass"
  }
}

Example Configuration (SD3.5-Large for Quality):

{
  "pipeline": "sd35_inpaint",
  "model": "stabilityai/stable-diffusion-3.5-large",
  "template": "Hextile_32_3K",
  "global_prompt": "vibrant landscape with mountains and lake",
  "strength": 0.66,
  "guidance_scale": 5.0,
  "num_inference_steps": 40,
  "interpolation": "6CardinalDirections"
}

Model Selection: - SD3.5-Large-Turbo: 1-8 steps (optimal 4), guidance 0-3.5 (optimal 1.0) - SD3.5-Large: 20-50 steps (optimal 40), guidance 3.5-7.5 (optimal 5.0)

Key Differences from SDXL: - ⚡ Turbo mode is 10-15x faster - 🎯 No negative prompt field (SD3 architecture doesn't use it) - 🔧 Auto-scheduler switching (EulerAncestral for Turbo, DDIM for Large) - 📁 Separate LoRA directory (sd3-loras/ vs stable-diffusion-loras/)

5. SD3.5 TensorRT (Optimized)

Best for: Maximum speed with NVIDIA GPU optimization

Technical Details: - Model: SD 3.5 with TensorRT optimization - Type: Tiled (uses hexagonal projection) - VRAM: ~10-16GB - Processing: Sequential or batch modes - Speed: 2-5x faster than standard SD3.5

Features: - ✅ All SD3.5 features (prompts, LoRA, cardinal directions) - ✅ NVIDIA TensorRT acceleration - ✅ Optimized for RTX 20xx/30xx/40xx GPUs - ✅ Same quality as standard SD3.5 - ❌ Requires NVIDIA GPU with CUDA ≥ 7.0 - ❌ May require architecture-specific engine builds

Requirements: - GPU: NVIDIA Turing architecture or newer (RTX 20xx+) - Compute Capability: ≥ 7.0 - VRAM: 10GB minimum, 16GB recommended - CUDA: Supported version installed

When to Use: - Have a modern NVIDIA GPU (RTX 20xx or newer) - Processing many images/sequences - Want fastest possible SD3.5 inference - Don't need ControlNet

Architecture Compatibility: | GPU Series | Architecture | Compute | Compatible | |------------|--------------|---------|------------| | RTX 20xx | Turing | 7.5 | ✅ Yes | | RTX 30xx | Ampere | 8.0/8.6 | ✅ Yes | | RTX 40xx | Ada Lovelace | 8.9 | ✅ Yes | | GTX 10xx | Pascal | 6.1 | ❌ No | | GTX 16xx | Turing | 7.5 | ✅ Yes |

Example Configuration:

{
  "pipeline": "sd35_tensorrt_inpaint",
  "model": "stabilityai/stable-diffusion-3.5-large",
  "template": "Hextile_32_3K",
  "global_prompt": "vibrant landscape with mountains",
  "strength": 0.66,
  "guidance_scale": 5.0,
  "num_inference_steps": 40
}

Note: TensorRT engines are architecture-specific. An engine built for RTX 4090 (sm_89) may not work on RTX 3080 (sm_86). The system validates compatibility on startup.

6. Flux Inpainting

Best for: Advanced users wanting cutting-edge diffusion quality

Technical Details: - Model: Flux.1 Dev / Flux.2 - Type: Tiled (uses hexagonal projection) - VRAM: ~16GB - Processing: Sequential or batch modes - Speed: Similar to SDXL

Features: - ✅ Advanced prompt understanding - ✅ LoRA support - ✅ Cardinal direction prompts (6 directions) - ✅ Lower guidance scale (3.5 typical vs 7.5) - ✅ Fewer steps needed (28 vs 50) - ✅ Template-based hexagonal tiling - ❌ No negative prompt (Flux doesn't use them)

When to Use: - Want latest diffusion technology - Don't need negative prompts - Have 16GB+ VRAM available - Exploring unique Flux aesthetics

Example Configuration:

{
  "pipeline": "flux_inpaint",
  "template": "Hextile_32_3K",
  "global_prompt": "surreal dreamscape with floating islands",
  "strength": 0.8,
  "guidance_scale": 3.5,
  "num_inference_steps": 28,
  "interpolation": "6CardinalDirections",
  "cardinal_prompts": {
    "front": "floating crystal structures",
    "back": "ethereal waterfalls",
    "left": "bioluminescent plants",
    "right": "ancient ruins",
    "above": "aurora borealis",
    "below": "misty void"
  }
}

7. Real-ESRGAN Upscaler

Best for: Fast upscaling without AI generation

Technical Details: - Model: Real-ESRGAN (ncnn/Vulkan) - Type: Direct (no tiling, processes full image) - VRAM: ~2GB - Processing: Single-pass upscaling - Speed: Very fast

Features: - ✅ 2x, 3x, or 4x upscaling (2K → 4K, 6K, or 8K) - ✅ Three model variants (general, anime, sharper) - ✅ Tile size control for VRAM management - ✅ No prompts needed - ✅ Can run alongside diffusion models - ❌ No hexagonal tiling - ❌ No prompt control - ❌ No LoRA support

When to Use: - Just need upscaling, no AI generation - Want fast processing - Limited VRAM (works with 2GB) - Post-process after SDXL/Flux render

Example Configuration:

{
  "pipeline": "realesrgan",
  "upscaling": {
    "model": "realesrgan-x4plus",
    "scale_factor": 4,
    "tile_size": 0
  },
  "output": {
    "width": 8192,
    "height": 4096,
    "file_format": "PNG"
  }
}

Model Variants: - realesrgan-x4plus: General purpose (best for photos) - realesrgan-x4plus-anime: Optimized for anime/illustrations - realesrnet-x4plus: Sharper details, may introduce artifacts

Scale Factor Options: - 2: 2x upscaling (2K → 4K, 4K → 8K) - 3: 3x upscaling (2K → 6K, 3K → 9K) - 4: 4x upscaling (2K → 8K, 1K → 4K)

Tile Size (VRAM Management): - 0: Auto (Full Image) - Best quality, most VRAM - 256: 256px tiles - Less VRAM (4-6GB GPU) - 512: 512px tiles - Balanced (8GB GPU) - 1024: 1024px tiles - More VRAM (12GB+ GPU)

Use smaller tile sizes if you encounter out-of-memory errors.

Comparison Table

Pipeline Switching

Automatic VRAM Management

The system intelligently manages GPU memory:

Switching Between Diffusion Models:

Currently using: SDXL (12GB in VRAM)
User switches to: Flux

System automatically:
1. Unload SDXL model
2. Run garbage collection
3. Clear CUDA cache
4. Load Flux model (16GB)

Result: No OOM crash!

Using Real-ESRGAN Alongside:

Currently using: SDXL (12GB in VRAM)
User also uses: Real-ESRGAN

System keeps both:
- SDXL stays loaded (12GB)
- Real-ESRGAN loads (2GB)
- Total: 14GB (fits in 24GB GPU)

Real-ESRGAN is a "direct processor" so doesn't trigger unloading.

In the UI

1. Click pipeline dropdown in header
2. Select desired pipeline
3. UI updates with appropriate form
4. Previous pipeline automatically unloaded (if diffusion)
5. Config updated with new pipeline

Choosing the Right Pipeline

Decision Tree

Do you just need upscaling?
├─ YES → Use Real-ESRGAN
└─ NO → Continue...

Do you need ControlNet?
├─ YES → Use SD 1.5 (8 types) or SD 2.1 (4 types)
└─ NO → Continue...

How much VRAM do you have?
├─ 4-6 GB → SD 1.5 or SD 2.1
├─ 8-12 GB → SDXL
├─ 12-16 GB → SD 3.5 or SD 3.5 TensorRT
└─ 16+ GB → Flux

Need negative prompts?
├─ YES → SD 1.5, SD 2.1, or SDXL
└─ NO → Any pipeline works

Want maximum speed?
├─ YES → SD 3.5 TensorRT (2-5x faster)
└─ NO → Choose based on quality needs

Use Case Examples

Scenario 1: Enhance a 360° Photo (Quality) - Pipeline: SDXL Inpaint - Model: SDXL 1.0 Inpainting (Default) - Why: Need subtle enhancement with negative prompt control - Settings: Low strength (0.3-0.5), high steps (50-70), guidance 7.5

Scenario 2: Fast Iteration / Previewing - Pipeline: SDXL Inpaint - Model: SDXL-Turbo (Fast) - Why: Quick iterations to test prompts and settings - Settings: Medium strength (0.5-0.8), 1-4 steps, guidance 0.0 - Tip: Iterate with Turbo, then final render with SDXL 1.0

Scenario 3: Create Artistic 360° Scene - Pipeline: Flux Inpaint - Why: Want unique Flux aesthetics and style - Settings: Medium strength (0.6-0.8), moderate steps (28-35)

Scenario 4: Upscale Existing 360° Image - Pipeline: Real-ESRGAN - Why: Just need 4x resolution increase - Settings: Choose model variant, adjust tile size if needed

Scenario 5: Two-Stage Processing - Step 1: Flux Inpaint (create artistic version) - Step 2: Real-ESRGAN (upscale to 8K) - Why: Combine creative AI with resolution boost

Scenario 6: Rapid Prototyping Workflow - Step 1: SDXL-Turbo (test 5-10 prompt variations, 1 step each) - Step 2: SDXL 1.0 (final render with best prompt, 50 steps) - Why: 10-30x faster iteration before final quality render

Parameter Guidelines

SDXL Inpaint

Model Selection: - SDXL 1.0 Inpainting (Default): Best quality, standard settings - SDXL-Turbo (Fast): 10-30x faster, requires different settings

Strength (0.0 - 1.0): - 0.0-0.3: Subtle enhancement, keep original details - 0.4-0.7: Moderate transformation - 0.8-1.0: Heavy transformation, more creativity

Guidance Scale (0.0 - 20.0): - For SDXL 1.0: - 5.0-7.0: More creative, less prompt adherence - 7.5-10.0: Balanced (recommended) - 10.0-15.0: Strict prompt following - 15.0+: Very strict, may reduce quality - For SDXL-Turbo: - 0.0: Required! (Turbo is trained without guidance) - Any value > 0 will degrade quality

Inference Steps: - For SDXL 1.0: - 20-30: Fast, lower quality - 40-60: Balanced (recommended) - 70-100: Highest quality, slow - For SDXL-Turbo: - 1-4: Optimal range (start with 1) - Higher values don't improve quality

Flux Inpaint

Strength (0.0 - 1.0): - 0.6-0.8: Recommended range for Flux - Flux tends to need higher strength than SDXL

Guidance Scale (1.0 - 10.0): - 2.0-3.0: Very creative - 3.5-4.5: Balanced (recommended) - 5.0+: Strict adherence

Inference Steps: - 20-25: Fast - 28-35: Balanced (recommended) - 40-50: Highest quality

Real-ESRGAN

Model Selection: - Photos/Realistic: realesrgan-x4plus - Anime/Art: realesrgan-x4plus-anime - Speed priority: realesrnet-x4plus

Denoise Strength (0.0 - 1.0): - 0.0: No denoising - 0.3-0.5: Light cleanup (recommended) - 0.6-0.8: Strong denoising - 0.9-1.0: Maximum denoising (may blur)

Tile Size: - 0: Auto (recommended) - 128-256: Lower VRAM usage - 512-1024: Faster, needs more VRAM

Resolution Controls

Available for SDXL, SD3.5, and Flux tiled pipelines.

Output Resolution

Control the final equirectangular output resolution. All presets maintain the 2:1 aspect ratio required for equirectangular images:

In the UI: Output Tab → Output Resolution (dropdown + manual input)

Tile Resolution Controls

Located in the Hextile Tab, these control how tiles are extracted and stitched:

Tile Input Resolution

Size at which tiles are extracted from the equirectangular image and sent to the diffusion model.

Tile Output Resolution

Size at which processed tiles are resized before stitching back into the equirectangular image.

• Typically matches Tile Input Resolution
• Can be increased for upscaling effect (e.g., input 512 → output 1024)
• Uses LANCZOS resampling for quality

Model-Sensitive Tile Sizes

The UI shows model-specific native tile sizes: - SDXL 1.0: 1024×1024 (native) - SDXL-Turbo: 512×512 (native) - SD3.5 Turbo/Large: 1024×1024 (native) - Flux Schnell/Dev: 1024×1024 (native)

Tip: Using non-native tile sizes shows a warning. Processing still works but may produce different results.

Example Configurations

Standard Quality (Default):

{
  "hextile": {
    "template": "Hextile_44_4K",
    "tile_input_width": 1024,
    "tile_input_height": 1024,
    "tile_output_width": 1024,
    "tile_output_height": 1024
  },
  "output": {
    "width": 8192,
    "height": 4096
  }
}

SDXL-Turbo with Native 512 Tiles:

{
  "pipeline": "sdxl_inpaint",
  "model": "local/sdxl-turbo",
  "hextile": {
    "tile_input_width": 512,
    "tile_input_height": 512,
    "tile_output_width": 512,
    "tile_output_height": 512
  }
}

Tile Upscaling (input small, output larger):

{
  "hextile": {
    "tile_input_width": 512,
    "tile_input_height": 512,
    "tile_output_width": 1024,
    "tile_output_height": 1024
  }
}

VRAM Considerations

If encountering OOM errors: Try reducing tile input size to 512×512 with a compatible model.

Cardinal Prompt System

Available for SDXL and Flux pipelines (not Real-ESRGAN).

How It Works

Global Prompt: "fantasy landscape"

Cardinal Prompts:
├─ Front: "enchanted forest entrance"
├─ Right: "ancient stone bridge"
├─ Back: "mystical mountain peaks"
├─ Left: "crystal clear lake"
├─ Above: "starry night sky"
└─ Below: "mossy forest floor"

Result: Each tile gets interpolated prompt based on viewing direction

Interpolation Modes

None (Global Only): - All tiles use same prompt - Uniform look across 360° - Simpler, more predictable

6 Cardinal Directions (Recommended): - Smooth transitions between directions - More natural 360° scenes - Takes advantage of hextile geometry

LoRA Support

Available for SD 1.5, SD 2.1, SD 3.5, SDXL, and Flux pipelines.

LoRA Directories

Each pipeline family uses its own LoRA directory:

Using LoRAs

1. Place .safetensors files in the appropriate directory for your pipeline
2. Select the pipeline in the UI
3. Navigate to the LoRA tab
4. Click "Add LoRA"
5. Choose model and set strength (0.0-1.0)
6. Add multiple LoRAs (they stack)

Combine Modes

• Additive (default): Each LoRA strength is applied independently
• Normalized: LoRA weights are normalized to sum to 1.0

Tips

• Start with strength 0.5-0.7
• Too high strength can overwhelm prompts
• Use only compatible LoRAs (SDXL LoRAs for SDXL, etc.)
• SD 1.5 has the largest LoRA ecosystem
• System lists available LoRAs via API

Config Files

Pipeline selection is saved in config files:

{
  "pipeline": "flux_inpaint",  ← Saved with config!
  "hextile": { ... },
  "diffusion": { ... },
  "output": { ... }
}

Loading Configs: - Config file includes pipeline - System automatically switches to saved pipeline - UI updates to show correct form - All settings restored

Pipeline Management

The Pipeline Management page lets you install, repair, and manage AI models.

Accessing Pipeline Management

Navigate to Pipelines in the main navigation menu, or click the gear icon next to the pipeline dropdown.

Pipeline Status

Each pipeline shows its current status:

Installing Pipelines

1. Click Install on any uninstalled pipeline
2. For gated models (SD 3.5, Flux), authenticate first:
3. Click Authenticate with HuggingFace
4. Enter your access token
5. Wait for validation
6. Monitor download progress via the progress bar
7. Once complete, the pipeline is ready to use

Download Sizes

Repair/Reinstall

If a pipeline isn't working correctly: 1. Click Repair to validate and fix the installation 2. If repair fails, click Uninstall then Install again

Health Check

Click Validate to run a health check on any installed pipeline. This verifies: - Model files are present and complete - Configuration is valid - Pipeline can be loaded

Test Render

Click Test to run a quick test render with default settings. This confirms the pipeline works end-to-end.

Performance Tips

VRAM Optimization

If OOM (Out of Memory) Errors: 1. Switch from batch to sequential mode 2. Use smaller template (Hextile_20_2K vs 44_4K) 3. Reduce inference steps 4. Lower output resolution 5. For Flux: Use SDXL instead (less VRAM)

Speed Optimization

Fastest Processing: 1. Use Real-ESRGAN for upscaling only 2. Use batch mode (if VRAM allows) 3. Reduce inference steps (but watch quality) 4. Use smaller templates 5. For Flux: 28 steps is good balance

Quality Optimization

Best Quality: 1. Use Hextile_44_4K template 2. Increase inference steps (70-100 SDXL, 40-50 Flux) 3. Use sequential mode (better tile blending) 4. Adjust strength carefully 5. Use cardinal prompts for better 360° coherence

Troubleshooting

"CUDA Out of Memory"

Solution: - Switch to sequential processing mode - Restart application (clears VRAM leaks) - Close other GPU applications - Use smaller template - Reduce batch size

"Pipeline Not Found"

Solution: - Check backend startup logs - Verify processors registered correctly - Restart backend server

"Model Loading Failed"

Solution: - Check model path in config - Verify HuggingFace model ID - Ensure internet connection (for downloading) - Check disk space for model cache

Tiles Don't Blend Well

Solution: - Use sequential mode (better blending) - Increase inference steps - Adjust strength (try lower values) - Use cardinal prompts for consistency - Ensure template is correct

Next Steps

• Getting Started - First render walkthrough
• Templates - Understanding hexagonal templates
• API Reference - Programmatic access
• Architecture - How the system works

Ready to start? Try your first render →

See also: - ControlNet Guide - Using ControlNet for structural guidance - Settings - Application settings and HuggingFace authentication

Last Updated: 2026-01-22 | V2.1 Pipeline Guide