Rate splat fidelity
Help train fidelity-ml v0.4. Below you'll see a reference render and two compressed candidates of the same scene. Pick the candidate that looks closer to the reference, or hit tie if you genuinely can't tell. Skip the pair if it's ambiguous or rendered badly — skipped pairs are dropped, not forced into the training set.
0 scenes available Reference preset: lossless-repack Target: 1,000 ratings/scene minimum
Rating panel
No rendered orbit frames are available in this build yet.
The collection page is wired and ready — once the bench
render step ships frames into benches/reports/frames/
they'll appear here automatically.
Thanks — you've hit the hourly cap
You've submitted the maximum 100 ratings this hour. The cap exists to keep one browser from flooding the corpus; it resets automatically. Come back in an hour, or share the page with a friend.
Something went wrong
Privacy
We do not collect personally identifying information. No accounts, no cookies for tracking, no analytics tags attached to your votes. Here's exactly what each rating row contains in our database:
-
scene_id,left_preset,right_preset,winner— the public identifiers of what you compared and your verdict. -
respondent_hash— a one-way SHA-256 of your IP address concatenated with your User-Agent string. Neither the raw IP nor the UA is stored. The hash is non-reversible: we can use it to detect a single browser flooding the page (and cap them at 100 ratings/hour), but we cannot recover who you are or where you came from from a row in our table. -
created_at— server-side RFC-3339 timestamp.
Source: see
apps/api/migrations/0003_ratings.sql
and
docs/fidelity-ml-v0.4-collection.md .
Methodology
Each pair shows one of 0 SplatBench scenes
rendered at a fixed orbit-8 camera position. The reference frame is
lossless-repack — the format-round-trip-only preset that
adds zero perceptible loss. The two candidates are different lossy
presets (e.g. web-mobile vs size-min). The
reference + the pair are all the same orbit position from the same
scene, so the only thing changing is the encoder.
Aggregated ratings get fed through a Bradley-Terry model that converts pairwise outcomes into per-frame absolute scalar scores; those scores become the supervision signal for the v0.4 MLP. Per the doc the training kicks off at 1,000 ratings per scene minimum; 10,000 is the desirable mark.