Seamless Latent Tiling — ComfyUI Custom Node

Generates perfectly tileable/repeating patterns by circular-padding the latent tensor at every denoising step.

How It Works

Normal generation:          With this node:

 ┌──────────┐               ┌──┬──────────┬──┐
 │          │               │ R│          │ L│  ← right edge wraps to left
 │  Latent  │    ──►        │  │  Latent  │  │
 │          │               │ R│          │ L│
 └──────────┘               └──┴──────────┴──┘
                                    ↑
  Edges are independent       UNet sees wrapped context,
  → seams when tiled          so predictions at edges account
                              for what's on the other side
                              → seamless tiling

At each denoising step:

1. Pad the latent with circular copies (right edge wraps to left, bottom to top)
2. Run the UNet on this larger wrapped tensor
3. Crop the output back to the original size
4. The sampler continues as normal

The result tiles perfectly because the model "sees" across the seam boundaries during generation.

Installation

Copy the entire seamless_latent_tiling folder into:

ComfyUI/custom_nodes/seamless_latent_tiling/

Restart ComfyUI. The node appears under sampling/seamless.

Workflow

Model Loader  →  [Seamless Latent Tiling ✦]  →  KSampler  →  VAE Decode  →  Save
                        ↑
                   Connect MODEL
                   output/input

Simply insert the node between your model loader and sampler. Everything else stays the same.

Parameters

| Parameter | Default | Description | |-----------|---------|-------------| | padding | 64 | Overlap in image-space pixels (divided by 8 internally for latent space). Controls how much context the UNet sees across edges. | | tile_x | True | Enable horizontal seamless tiling (left ↔ right) | | tile_y | True | Enable vertical seamless tiling (top ↔ bottom) |

Padding Guidelines

• 64 px (8 latent px) — Fast, works for simple patterns
• 128 px (16 latent px) — Good balance, recommended starting point
• 256 px (32 latent px) — Maximum context, best for complex scenes, slower

Tips for Best Results

1. Use pattern-oriented prompts: Include words like "seamless pattern", "tileable texture", "repeating design" in your prompt.

2. Square images work best: Use 512×512 or 768×768 for most reliable tiling.

3. Increase padding if you see faint seams: If the edges almost-but-not-quite match, increase the padding value.

4. tile_x / tile_y independently: If you only need horizontal tiling (e.g., a border strip), disable tile_y to save memory and get better vertical composition.

5. Verify your result: After generating, tile the image 2×2 or 3×3 in any image editor to visually confirm seamlessness.

Known Limitations

• ControlNet: ControlNet injects spatial features at intermediate UNet layers that are NOT padded by this node. ControlNet guidance may break seamlessness near edges. Use without ControlNet for guaranteed results.

• Memory overhead: The UNet processes a slightly larger tensor. With 128px padding on a 512×512 image, the latent goes from 64×64 to 96×96 (~2.25× area). Plan GPU memory accordingly.

• Inpainting models: Concat conditioning (c_concat) IS padded when it matches the latent spatial dimensions, so inpaint models should work. However, this is less tested.

Compatibility

• Works with SD 1.5, SDXL, and turbo/lightning variants (including z-image-turbo)
• Works with any sampler/scheduler in ComfyUI
• Works with LoRAs and textual inversions
• Does NOT require model modifications — purely a sampling-time wrapper