Wide shots work best. The model needs background and mid-ground to infer what surrounds you.
Describe what's around you, not what's in the photo. The model can already see the photo — what it can't see is the 270° behind you.
Low numbers make it match your photo's colour, light and mood closely but wander from the words. High numbers obey the prompt literally and drift from the photo's look. If the result feels like a different room, come down.
Skips generation entirely. Drop one you already like and go straight to seam repair, upscaling, or HDRI export — nothing regenerates, so the image you liked stays exactly as it is. Only the seam pass costs anything; upscaling is optional and the HDRI conversion is free.
Rolls the panorama half a width so the wrap sits in the middle of the frame, masks a band over it, and has an inpainting model paint across the join with real context on both sides — then rolls it back. A generator can never do this, because it never sees the two ends at the same time. Do this before upscaling.
Wraps a strip from each edge around to the other side before upscaling, so the upscaler sees continuous content across the join instead of a border, then crops the padding back off. Without this the 8K pass puts a fresh seam into an image you already repaired.
This upscaler redraws rather than sharpens, so it needs to know how much licence it has. Low sticks to what's there and just cleans it up. High invents stone grain, brass patina and steam structure — better on soft sources like this one, but it will change things. 0.4–0.6 is the useful range.
The upscaler's own tiling option. It wraps both axes though, and an equirectangular should only wrap horizontally — so it may do something strange at the poles. Worth racing against the padding above to see which scores better.
Composites your untouched upload back over its own patch of the finished panorama at full resolution. Your photo gets shrunk to a few hundred pixels to fit the plate, so without this the upscaler is inventing detail for the one region where real detail already exists. Off when you sent the raw photo, since there's no known patch to restore.
A PNG tops out at pure white, so your lamps carry the same brightness as a grey wall and Blender lights the scene flat. This finds the pixels that are obviously emitting and pushes them far past white into real radiance, then writes a Radiance .hdr you can drop straight into an Environment Texture node.
Everything brighter than this is treated as a light source. The preview marks what's selected — aim for lamps and shafts only. Too low and your floor starts glowing.
How much brighter than white the light sources become. This is what produces hard shadows and specular highlights. Around 50× for soft interior lamps, 300×+ for anything standing in for sunlight.
Lifts the whole image, shadows included. Use this if the scene is lit but too dim; use boost instead if it's bright enough but looks flat.
Drop either one into the Meridian workbench to score the wrap seam. Camera captures sit at 1.0–1.4×.