The Ultimate ComfyUI Empty Latent Resolution Guide (1MP Aspect Ratio Buckets)

One of the most frequent mistakes beginners make when transitioning from Midjourney or WebUI to ComfyUI is typing arbitrary video dimensions like 1920 (Width) and 1080 (Height) directly into the Empty Latent Image node.

The generation completes, but the image is plagued with deformed figures, dual torsos, or duplicate heads. In this guide, we break down why this happens and give you the exact resolution lookup table trained into modern foundation models like Flux.1 and SDXL.

The Science Behind "Aspect Ratio Bucketing"

During pre-training, diffusion models are not fed arbitrary screen sizes. Black Forest Labs (for Flux.1) and Stability AI (for SDXL) trained their models using aspect ratio bucketing.

Each bucket is strictly calibrated to maintain a total area of approximately 1,048,576 pixels (1.0 Megapixel), and both dimensions must be divisible by 64 (or 16) to align with the VAE latent compression factor.

When you feed an uncalibrated resolution like 1920x1080 (which is 2.07 Megapixels, double the native capacity), the model's self-attention layers perceive excessive canvas space. Having no concept of a 2MP single scene, the diffusion process attempts to tile two 1MP concepts side-by-side—spawning duplicate bodies or repeating horizon lines.

Complete 1MP Resolution Lookup Table

How to Upscale to 4K Properly

If your end goal is a crisp 4K wallpaper (3840×2160), do not generate at 4K in Empty Latent. The correct ComfyUI workflow pipeline is:

1. Generate the base composition at 1344×768 in Empty Latent Image.
2. Pass the decoded VAE output through an Upscale Model Loader (using models like 4x-UltraSharp or 4x_NMKD-Superscale).
3. (Optional) Perform a gentle second-pass KSampler (Inpainting/Hires fix) with a low denoise value between 0.25 and 0.35 to sharpen fine textures without altering the core scene composition.