360° VAE Decode

After a good sampler pass your latent is about as seamless as latent space gets. But VAE decode is its own opportunity to wreck the wrap: the decoder's convolutions look at a flat rectangle and have no idea the left edge is supposed to meet the right edge. Equirect360VAEDecode is a VAEDecode replacement that gives the decoder that context by circular-padding the latent before decode and cropping the padding back off after.

Three inputs, two of which you'll actually set:

• samples - the latent from 360° KSampler (or a standard KSampler).
• vae - the VAE matching your model, from a VAELoader.
• circular_padding (default 16, 0–128) - how much X circular padding to apply in latent space. 0 is a fine baseline; go 8–16 if you see a seam after decode; and set it to 0 if you're using Apply Circular Padding VAE.

The mechanism is clean: it pads the latent with a wrapped copy on the left/right, decodes the padded latent, then crops the padding off. Because the FLUX VAE is 8× compression, 16 latent pixels become 128 image pixels - so the padding values look small but the effect is generous. Output is a regular IMAGE, ready for 360° Edge Blender or a Save Image node.

The one footgun to remember: double padding. If you've patched the VAE with Apply Circular Padding VAE (which makes the decoder's own convolutions circular), you should not also pad here - set circular_padding to 0, or you get mushy, doubled edges. The tooltip says exactly this, the README repeats it, and it's the most common mistake with the two VAE-side nodes in this pack.

Install

Same as the whole pack - Manager (search "ComfyUI-DiT360"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/cedarconnor/ComfyUI-DiT360
cd ComfyUI-DiT360
pip install -r requirements.txt

then restart. No heavy dependencies - numpy and Pillow, both already in ComfyUI. Models: FLUX.1-dev in models/checkpoints, DiT360 LoRA (~2–5GB, Insta360-Research on Hugging Face) in models/loras at strength 1.0.

Troubleshooting

• Edges look soft or doubled after decode → you're double-padding. Check whether Apply Circular Padding VAE is in the graph and zero this node's circular_padding.
• Seam still visible after decode → raise circular_padding to 8–16, then lean on the Edge Blender for the rest.
• Everything looks fine → leave it at 0 and let the Edge Blender handle the seam. Less work, same result.

Honest take: this is insurance, not the main event. The sampler's circular padding does the heavy lifting, decode padding catches what leaks through, and the Edge Blender sweeps the rest. On a 16GB card at 2048×1024 it's cheap enough to leave on - just never at the same time as the VAE patch.