Scenario Unity 2D

Unity 2D (developer and artist)

Expert 2D work in Unity 6.3 is a chain of coupled settings proven by numbers and a frame: one PPU from art to camera, atlases judged by what they save, tiles and lights authored as data and costed, a camera that moves in whole pixels, and a controller whose feel is measured, not asserted. Everything below was run in Unity 6000.3.21f1 through batch jobs and PlayMode tests (references/procedures.md). If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

REQUIRED BACKGROUND: scenario-unity-expert (channels, review loop, 6.3 traps).

Stance (the expert delta)

1. Pixel perfect is a pipeline, and URP already ships the camera. One PPU on every sprite (tile px = 1 unit), Point, Compression None, whole-pixel pivots, Assets PPU = sprite PPU. Never install com.unity.2d.pixel-perfect in URP: it is the Built-in component (6.3 Manual). Measured: 100% uniform 6 x 6 blocks at 1080p.
2. Pick Grid Snapping by what must be pixelated, and judge pixels in motion. Pixel Snapping snaps sprites only; lights, bloom, particles stay at screen resolution (2qeNu2QApAM [00:01:47]); Upscale Render Texture pixelates everything. Retro AA vs Point is shimmer vs blur (Manual), and in URP 17.3 it only acts under Stretch Fill. Measured over a sub-pixel pan: integer framing moves in whole art pixels (rigid 0.9997); Stretch Fill Point leaves 0.6% of pixels shimmering per frame, Retro AA 0.3% at twice the softness; no Pixel Perfect Camera, 2.8%.
3. Under the SRP Batcher an atlas cuts textures, not Batches or SetPass. Measured 21/21 batches, 2/2 SetPass, bound textures 20 to 1 (Unity hXlpnwD-TgY [00:05:55], [00:07:32]); memory only drops when usage is high. Compress once, on the atlas, except pixel-art pages stay None: compression shifts pixel colors (2qeNu2QApAM [00:02:24]). AuditAtlases reproduces the 6.3 Sprite Atlas Analyzer reports plus Read/Write; grouping by co-visibility is the step no tool does.
4. The controller owns the velocity; forgiveness is design, each trick behind a toggle. Latch input in Update; collisions, jump, horizontal, gravity, apply in FixedUpdate; coyote and buffer are timestamp windows (Tarodev source). "Working on the player's intent" (Thorson, GMTK yorTG9at90g [00:11:03]), so coyote, buffer, variable jump, fall clamp, apex hang and corner correction each get an on/off toggle for A/B play (Tarodev 3sWTzMsmdx8 [00:02:24]), and hazard hitboxes are smaller than the art (GMTK [00:09:28]). Two choices are feel decisions, never defaults: the variable-height method (velocity cut 50% precise, gravity x3 smooth, sustain while held long; Dawnosaur 2S3g8CgBG1g [00:00:36]) and the air friction (Tarodev AirDeceleration 30 drifts 3.1 units, 5.7 body widths, after release; Celeste's "massive air friction" drops almost straight down, GMTK [00:04:14]; 240 stops in 3 steps).
5. Lights are routed by sorting layers and paid per light batch. Consecutive sorting layers share a light batch only if every visible light lights both or neither, and the normal, shadow and light passes run once per batch (URP 17.3 source). A player fill light that skips the Characters layer (Sasquatch 1h-hSlffawM [00:01:35]) split the demo level from 1 batch into 4 and light fill from 0.99 to 3.64 screens. Cost is fill rate (e-book p. 131): bench at 4K over 6 runs, shadows x5.4 to x5.9, the batch split +72 to +101%, normal maps +15 to +36%, large vs small lights +8 to +43% (noisy). Never a pitch-black area (Global 0.25, GDC YhrwKF_i-BI [00:11:43]); enable normal maps on the light (GDC [00:12:31]); Fast on low tiers.
6. Sprite animation never blends and follows the controller. Exit time off, duration 0, Any State self-transition off; parameters from what the controller did, not from input (Brackeys hkaysu1Z-N8 [00:16:17], [00:18:57]).
7. The camera and the juice are part of feel. Loose vertical damping rising, tight once falling past a speed threshold, back at vertical speed 0 (hysteresis), lerped (Sasquatch 9dzBrLUIF8g [00:05:37], [00:07:14]): in a 40-unit drop a fixed 0.6 s damping let the feet reach the bottom edge, the switch kept 3.2 to 3.4 units below in view. Swap room cameras on trigger EXIT by exit direction ([00:14:43]); bounds end exactly at a secret room ([00:02:10]). Hit stop 20 ms (Nijman AJdEqssNZ-U [00:18:22]) to 67 ms (Celeste, GMTK [00:08:40]) on unscaled time, extend not stack; shake with an off option (Nijman [00:32:14]).
8. Be honest about GUI-only editors. Tile Palette, Rule Tile inspector, AutoTile masks, Sprite Editor, Skinning Editor and the Animator window are mouse tools. Agents author through Tilemap.SetTiles, RuleTile.m_TilingRules, RuleOverrideTile.ApplyOverrides, ISpriteEditorDataProvider, AnimatorController; rigging and AutoTile masks stay human (references/gui-paths.md).

Establish first

Ask once: pixel art or HD (defaults: pixel, 16 PPU, 320 x 180); target screens including 21:9 and phones; camera model (static rooms, follow with bias and fall damping, lead toward action); real-time 2D lights or painted light, and the lowest tier; controller profile (precision: own the velocity, high air friction; momentum: low air friction; physics-rich: Dynamic body with forces; top-down: velocity); jump height in tiles and the variable-height method; hazards and one-way platforms; input devices. Then ut_2d.project(P).

Workflow

1. Project. TwoDSetup.ProjectSetup: sorting layers back to front (few: they drive light batching), Player and Ground layers. GATE: order read back, no pixel-perfect package, Input System, fixed step 0.02 s.
2. Import. SpriteJobs.ImportPixelArt (slicing, pixel pivots, _NormalMap) or ImportHD (PPU = height / (2 x ortho), e-book p. 21); shaped tiles get SetPhysicsOutline (16 px sprites generate a 4-point rectangle). GATE: AuditSprites 0 errors; one PPU.
3. Atlases. AtlasJobs.BuildAtlas per co-visible group, lit and plain apart, pixel pages uncompressed. GATE: 1 page, 0 findings, grouping rule written.
4. Level. TilemapJobs.CreateRuleTile + BuildLevel; reskins with CreateRuleOverrideTile; Composite Merge, Outlines, Static; per-tile colliders only for runtime edits, then Use Delaunay Mesh. GATE: rule_mismatches == 0 (reskin read-back too), drop probe within 0.05 units, capture shows no seams.
5. Actors and animation. AnimJobs.BuildSpriteAnimator, LevelJobs.Populate (player, Hazard2D spikes with inset hitboxes, one-way platforms). GATE: LintAnimator 0 errors; hitbox-to-art ratios below 1.
6. Lights. LightJobs.LightRig, then LightJobs.LightCost (tier; engine check once the camera exists). GATE: Audit2D 0 errors; replica light batches = URP's own count; every batch break and every light over half the view justified; normals on/off capture pair differs.
7. Camera. CameraJobs.SetupPixelPerfect, CapturePixelPerfect; a sub-pixel pan (ut_2d.pan_shots) scored by ut_2d.shimmer_check; follow cameras through CinemachineJobs.Setup (CM 3.1.7 pinned) plus CinemachineFallDamping and CinemachineRoomSwap. GATE: pixel_block_check >= 0.999 at 1080p and 720p, pan min rigid >= 0.99, void_columns 0 at 21:9, bounds_hide_secrets ok, contact sheet looked at.
8. Feel. PlayMode tests with simulated input (ut_run.run_tests(P, "PlayMode", graphics=True)). GATE: apex = 50 Hz prediction, windows within one step, one press = one jump at 30/60/144 fps, each toggle A/B, method and air-drift numbers stated. Feel itself is a human verdict: ship toggles and numbers.
9. Deliver. Audit2D.Scene clean; AgentProfile.PlayModeTimings budget pass; report each step with its Unity call, then Verified and Assumed.