Apply Circular Panorama (All-in-One)

If Apply Circular RoPE is the focused experiment and Apply Circular Padding VAE is the decoder patch, Apply Circular Panorama is where the pack throws both model-level tricks into one node and hands you the switches. It patches your diffusion model's Conv2d layers for X-only circular padding (where applicable) and applies circular RoPE (same shift/angle options), all in a single model-in, model-out wrapper. Convenient for testing; not something to leave on out of curiosity.

Inputs:

• model - your FLUX + DiT360 LoRA model.
• patch_conv2d (default true) - model-internal Conv2d circular padding on X. The tooltip is blunt: for FLUX this can reduce stability and quality, and the pack recommends preferring the 360° KSampler/VAE padding or Apply Circular Padding VAE instead. Enable only if you know you need padding inside the model.
• patch_rope (default true) - attention-level wrap. Tooltip recommendation: OFF by default; enable only for testing RoPE seam handling. (Note the quirk: the defaults are true, but the pack's own advice is to turn them off. Treat this node as "everything toggled, your job to decide.")
• rope_mode (shift default / angle) - shift is the safe one; angle is aggressive and experimental.
• rope_seam_width (default 0 = auto) - shift only: token columns near the right edge to wrap. 0 or 4–16 recommended; larger values distort.
• rope_verbose (default false) - prints diagnostic info about the RoPE patch.

Output: a patched model into your sampler. If that sounds like a lot of knobs for one node - it is. This is the "try everything at once" node, which is exactly why the defaults deserve skepticism: both toggles default on, but the documentation recommends both off in favor of the sampler/VAE/decode padding path.

How it works: it reuses the same two mechanisms as its siblings - the Conv2d forward patch (circular X, constant Y, the approach ComfyUI_pytorch360convert also uses) and the RoPE wrapper (shift vs angle). It's not new machinery; it's a convenience wrapper with every switch exposed.

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; models are FLUX.1-dev in models/checkpoints and the DiT360 LoRA (~2–5GB, Insta360-Research on Hugging Face) in models/loras at strength 1.0.

Troubleshooting

• Washed out or broken output → turn patch_rope off; prefer full-precision dev over fp8; use shift with rope_seam_width 0–16.
• Quality dropped but no seam improvement → that's patch_conv2d on FLUX doing exactly what the tooltip warned about; disable it and use sampler/decode padding instead.
• Don't stack it with Apply Circular RoPE or a second RoPE patch - you'll double-patch, and double-patching is how you get garbage.

Honest take: this is the node you reach for when you want one quick A/B test of whether the model-level fix helps on your model at all. Run it with everything on, compare against the padding + blend baseline, then go back to the baseline - which, per the pack's own troubleshooting, is the reliable path. The kitchen sink is best used one drain at a time.