← Demos Server Connect ?

✕

JPIP Gigapixel Viewer

Only the precincts that fall inside the current viewport are fetched and decoded, so pan-and-zoom on a 400+ MP image stays interactive.

URL parameters

Param	Default	Effect
?server=	—	JPIP server (host[:port]).
?debounce=	60	Trailing-debounce window in ms; 0 disables. Coalesces rapid wheel / trackpad events into one fetch.
?debug	off	Per-frame timing dump to the browser console.
?variant=	auto	mt or st — force multi- or single-threaded WASM.
?maxSize=	1920x1080	Cap the WebGL render target to WxH. Bounds per-frame precinct fetch + decode work on 4K / ultrawide displays. Pass ?maxSize=window to disable the cap.
?fit=	stretch	stretch — CSS scales the render target up to fill the window, GPU upsamples via GL_LINEAR. contain — CSS size == render size, canvas is shown at native pixel scale centered in the window (like the foveation demo).
?precinctCacheMB=	64	LRU precinct cache budget in megabytes. Received precincts are tracked in the &model= cache-model advertisement so the server skips ones already in the viewer's cache on subsequent pans (70–95% byte reduction typical). Set 0 to disable.
?prefetchMargin=	128	After each successful render, fetch an N-canvas-pixel halo of adjacent precincts into the LRU cache so short follow-up pans hit locally. Auto-cancelled when the user resumes interacting. Set 0 to disable.
?prefetchDelayMs=	150	Idle-delay before the halo prefetch fires (in milliseconds). Short enough to overlap typical pause-and-examine behaviour; long enough that continuous panning never triggers it.
?midPaintMs=	200	Show a partial render (whatever precincts have arrived) if a fetch takes longer than N ms. Gives feedback on slow first-time loads; costs one extra decoder run. Fast (LAN / cached) fetches never trigger it. Set 0 to disable.
?thumbnail=	1	Small bottom-right overview showing the full image (fetched once at connect time from the coarsest reduce level) with a rectangle marking the current viewport. Rectangle updates locally on every pan/zoom — no extra server round-trips. Set 0 to hide.
?thumbnailSize=	160	CSS-pixel size of the thumbnail on its long side. Decoder output resolution is independent — only the displayed size changes.
?cb=	off	Cache-buster for dev iteration. auto or 1 appends ?v=<Date.now()> to the WASM/JS fetches so the browser refetches after every rebuild. Any other value is used verbatim (snapshot at a build hash). Default (no ?cb): full HTTP caching, identical to before.
?carryForward=	1	While a JPIP fetch is in flight after a pan or zoom, redraw the previous frame's pixels at their image-coord position with the thumbnail filling any uncovered area — the image scrolls smoothly under your gesture instead of freezing until the new bytes arrive. Set 0 to disable (canvas stays as last drawn until the new fetch completes). Auto-suppressed on fast networks via ?carryForwardMinMs=.
?carryForwardMinMs=	100	Minimum median fetch+decode latency (ms) before carry-forward redraws fire. On localhost / LAN where fetches typically finish in <50 ms, the previous-frame translate would be visible for ≤2 frames before the sharp final paint replaces it — and the brief blurry-stretched intermediate looks worse than waiting. Above this threshold the smoothing kicks in. Set 0 to force-enable regardless of measured latency.

JPIP query (§C.4)

Each pan fires one GET:

/jpip?fsiz=W,H&roff=X,Y&rsiz=W,H&type=jpp-stream&model=…

Field	Meaning
fsiz	Resolution frame size — picks the discard level.
roff	Region offset within the frame.
rsiz	Region size (viewport at the active reduce level).
model	Cache-model advertisement (see below).

Cache-model format (§C.9)

After frame 1 the viewer appends &model=Hm,Ht0,M0 (single-tile) or &model=Hm,Ht0-N,M0 (multi-tile) so the server skips header data-bins already delivered.

Class	Prefix	Example
Main header	Hm	Hm
Tile header	Ht	Ht0, Ht1-5
Metadata bin	M	M0
Precinct	P	not sent by this demo