Generating high-resolution images with Tiled Diffusion
Your GPU’s VRAM caps the resolution you can generate in a single pass — but Tiled Diffusion removes that ceiling by splitting the canvas into overlapping tiles, denoising each one independently, and fusing the results. This tool calculates a tile size that fits your VRAM, works out how many tiles your target resolution needs, and tells you whether MultiDiffusion or Mixture of Diffusers is the better method for that canvas.
How it works
You enter the target resolution, your available VRAM, and a preferred tile overlap. The tool maps your VRAM to a safe latent tile size (smaller tiles for 4-6GB cards, up to roughly 1920px for 24GB cards). It then computes the tile grid from your target dimensions and the effective stride (tile size minus overlap), giving you the total number of tiles per image and an estimate of generation time relative cost. Larger canvases trigger a Mixture-of-Diffusers recommendation for cleaner seams.
MultiDiffusion vs Mixture of Diffusers
Both methods tile the canvas and fuse the denoised results, but they differ in how the fusion is handled:
MultiDiffusion merges tiles by averaging the predicted noise in overlapping regions at each denoising step. It is faster because the averaging is simple, and it works well for moderate resolutions where the overlap is sufficient to blend naturally. For resolutions up to roughly 2048 × 2048, MultiDiffusion is the faster choice.
Mixture of Diffusers uses a Gaussian-weighted blend in the overlap zones, giving a smoother transition between tiles. It handles seams better on very large canvases where a sharp tile edge would be visible — typically above 2048 on a side, or when the prompt content changes significantly across the image (for example a wide landscape scene). The trade-off is slightly more computation per step.
This tool recommends one method based on your target megapixels as a starting point. You can override it if you see quality differences in your specific use case.
The VAE decode problem
Tiled Diffusion handles the UNet denoising — the computationally expensive part — in tiles. But the VAE decoder that turns the latent into a pixel image runs on the entire full-resolution latent by default, which can spike VRAM above the same limit Tiled Diffusion was circumventing. On a 6GB card trying to generate a 2048 × 2048 image, an out-of-memory crash at the final VAE decode step is common.
The fix is to enable Tiled VAE (available in the same A1111 extension) alongside Tiled Diffusion. Tiled VAE decodes the latent in overlapping tiles too, preventing the spike. Always enable both.
Overlap and seam artefacts
Tile overlap controls how much neighbouring tiles share before fusion. Too little overlap and you see visible grid lines or hard seams, especially over smooth gradients (sky, skin, flat colour). Too much overlap slows generation because more pixels are processed twice.
A starting overlap of 48–64 px works well for most content. If you see seams:
- Increase overlap to 96 px and re-generate.
- Switch from MultiDiffusion to Mixture of Diffusers for the blending advantage.
- Check that batch tile count is set to 1 — processing more than one tile at a time can sometimes produce inconsistent results.
Using Tiled Diffusion for upscaling
Beyond generating new images, Tiled Diffusion paired with a ControlNet tile model is one of the most effective high-res upscaling workflows available in Stable Diffusion. The typical process: generate at a lower resolution to get the composition right, then upscale 2–4× using img2img with Tiled Diffusion and a tile-model ControlNet to add fine detail without changing the overall structure. This approach can produce 4K+ output from a 512 × 512 base on a mid-range GPU.