Universal Seamless Tiles

Universal Seamless Tiles

Model-agnostic seamless / tileable image generation for ComfyUI.

The classic seamless-tiling trick (switching every Conv2d to circular padding) only works on convolutional U-Nets like SD1.5 and SDXL. Modern transformer / DiT models — Flux, Wan / Anima, and similar — patchify with a Linear or a non-overlapping Conv, so that trick has nothing to grab and produces no tiling at all.

This pack takes a different, architecture-independent approach that works across SD1.5, SDXL, Flux, and Wan/Anima:

• Latent rolling during sampling makes the denoiser treat the canvas as periodic, so no fixed seam can form.
• Circular VAE decode removes the pixel-level seam at the true edges, for both 2D (SD/Flux) and 3D (Wan) autoencoders.

Neither node mutates ComfyUI's shared/cached model — they clone/copy first, so tiling can never "leak" into your other generations.

Nodes

Seamless Tile Model (DiT) — MODEL → MODEL

Installs a UNet function wrapper that rolls the latent by a per-step offset (and un-rolls the result), so the model can never lock onto a fixed edge. The roll magnitude is tapered with the noise level: large structural rolls early, near-zero rolls on the final detail steps — this keeps the tiling artifact on the true edge (where the circular VAE handles it) instead of leaving a faint line inside the image.

| Input | Meaning | |-------|---------| | model | Any diffusion model (SD/SDXL/Flux/Wan/…). | | tiling | enable (both axes), x_only, y_only, disable. | | seed | Makes the per-step offsets reproducible. |

Make Circular VAE (DiT) — VAE → VAE

Applies circular padding to the VAE decoder so the decoded edges wrap. Handles both Conv2d (SD/Flux AutoencoderKL) and Conv3d / CausalConv3d (Wan VAE), preserving Wan's causal temporal padding and single-frame fast path. Feed the result into a normal VAE Decode.

| Input | Meaning | |-------|---------| | vae | The VAE to make tileable. | | tiling | enable, x_only, y_only, disable. | | copy_vae | Make a copy (recommended) or Modify in place. |

Usage

Load Checkpoint ─▶ Seamless Tile Model (DiT) ─▶ KSampler ─▶ VAE Decode ─▶ (image)
                                                              ▲
                        VAE ─▶ Make Circular VAE (DiT) ───────┘

Set both nodes to the same tiling mode. To verify the result tiles, offset the output by 50% in x/y (e.g. an image-offset node) and check that no seam appears.

Notes & limitations

• Latent rolling is very good but not conv-perfect. Quality scales with step count because the method averages across steps. On turbo / few-step models, use more steps (e.g. 15–20) for tighter seams, or expect a slightly softer result.
• Rolling does not roll ControlNet hints or other spatial transformer_options data, so ControlNet/inpaint spatial alignment with tiling is out of scope for now.
• For single-image use of video models (Wan/Anima), only the spatial (H/W) axes are rolled.

Compatibility

• ComfyUI (recent builds with set_model_unet_function_wrapper).
• Tested against SDXL, Flux, and Wan/Anima model + VAE code paths.

License

MIT — see LICENSE.