Chapter 3: LOD, Performance & Memory Guardrails

Created by Sarah Choi (prompt writer using ChatGPT)

LOD, Performance & Memory Guardrails for Character Concept Artists

Why LOD and memory should live in your concept

Level of Detail (LOD), performance, and memory are often treated as downstream problems, but concept decisions set ~80% of the cost envelope before modeling begins. Silhouette complexity, material variety, transparency usage, and decal density all compound across mix‑and‑match systems. By embedding guardrails—clear budgets, fallback states, and atlas strategies—directly in your boards, you protect readability, frame rate, and the ability to ship many variants without content explosions.

Foundations: the three budgets you influence

Concept artists shape three budgets:

  1. Geometry budget (triangles, silhouette breaks, separate pieces).
  2. Material budget (shader complexity, number of materials per asset, transparency).
  3. Texture/memory budget (texture count/size, atlas packing, mips, compression). Every slot (Head, Hair, Outfit parts, Palettes, Decals) must carry its own micro‑budget and degrade gracefully across LODs and camera contexts.

Screen‑size thinking: design for the pixel, not the inch

LOD is fundamentally about projected screen size. Paint tiny callouts showing the character at common distances (e.g., 25 px face, 120 px torso, 500 px hero close‑up) and annotate which details must persist vs. which can collapse. Translate filigree and micro‑greebles into contrast carriers (bigger shapes, simplified trims) that still read when mips blur.

Geometry guardrails

Design shapes that simplify predictably:

  • Prefer stepable silhouettes—forms that can lose bevels and vents at LOD2–3 without breaking the style language.
  • Cluster small ornaments into mergeable islands so lower LODs can weld them into the base without popping.
  • Avoid knife‑edge overlaps that require high geo to avoid z‑fighting at distance; propose chamfers and recessed beds instead.
  • For modular kits, define slot shells (consistent under‑geometry) so swappable parts don’t require unique base meshes per variant.

Material/shader guardrails

Expensive materials accumulate. In art, tag materials with complexity bands:

  • Band A (cheap): opaque, single‑layer PBR, no subsurface.
  • Band B (moderate): masked transparency, simple emissive, detail normal.
  • Band C (costly): full transparency, anisotropy, clearcoat, SSS, transmission. Constrain costly bands to small, hero‑critical zones and provide LOD shader fallbacks (e.g., clearcoat → baked highlights; transmission visor → opaque tint at distance). Where possible, move look into textures and curvature masks instead of layered shaders.

Texture/memory guardrails

  • Author concepts with atlas groupings in mind: “Head Core,” “Hair Shared,” “Soft Kit,” “Hard Kit,” “Decal Sheet,” each with indicative texel density.
  • Prefer channel packing (R: Roughness, G: Metalness, B: AO) to reduce texture count; show this intent in notes so texturing follows suit.
  • For mix‑and‑match, push small accessories to shared accessory atlases; reserve unique 2K–4K sheets for hero chest/helm only when justified by camera usage.
  • Call out mip‑safe patterns; avoid high‑frequency tartans and tight herringbones unless they’re baked into a lower‑frequency version for distant LODs.

Heads: LOD, expression fidelity, and material choices

  • Geo: Concentrate head topology where expressions need it (lip roll, nasolabial folds, eyelids). For LODs, specify where loops can collapse while preserving the eye/mouth ellipses.
  • Eyes: Favor opaque cornea + faked caustics at mid LODs; reserve refractive cornea only for cinematics. Provide a LOD table: LOD0 tearline card present; LOD1 baked spec highlight; LOD2 single sphere with baked AO.
  • Facial hair & brows: Prefer masked over fully transparent cards; design shapes that can switch to textured geo blocks at distance.
  • Makeup/facepaint: Author two density versions (close & far) to avoid aliasing and moiré on thin liners.

Hair: alpha, ribbons, and impostors

Hair dominates overdraw and memory. In concept:

  • Provide three representations per style: Cards (LOD0), Ribbons/Chunk Strips (LOD1–2), and Cap Texture (LOD3 impostor). Paint each explicitly.
  • Limit see‑through density by clustering strands into readable clumps that survive alpha‑to‑coverage and mip blur.
  • Avoid long, thin flyaways that won’t survive LOD; concentrate expressiveness in silhouette bangs and tip shapes that convert cleanly to ribbons.
  • For dyed tips/emissive streaks, keep accents thicker than minimum pixel widths at gameplay camera to avoid shimmer.

Outfits: layering without material inflation

  • Establish a material cap per layer: e.g., undersuit (1), mid garment (1), outer armor (2 incl. emissive), cape (1 masked). If the cape is present, budget steal from the outer layer to hold total steady.
  • Design trim zones that carry style at all LODs so that the LOD2 jacket (fewer seams) still reads as the same set via consistent trims.
  • Hard/soft junctions: specify baked shadow bands and paintwear so LODs without geo recess still separate visually.

Palettes: LUTs, dye rails, and shader stability

  • Treat palettes as data‑light: prefer palette LUT or small mask textures over unique albedos per colorway.
  • Define value rails for each dye channel across palettes so LOD mips don’t merge channels into a mud mid‑value.
  • Separate emissive intensity maps from dye channels to avoid reauthoring materials for every palette.

Decals: legal zones, batching, and fallback

  • Group decals into few materials by placing them on shared sheets; keep vector‑friendly shapes so resolution scales.
  • Mark legal zones on relatively flat areas to minimize stretching and costly high‑order projections.
  • Provide fallback art for distant LODs (solid shape badge vs. typographic logotype) to avoid shimmer and z‑fight.

Cross‑slot strategy: combinations without combinatorial memory blow‑up

  • Adopt tiered heroism: per character, designate 1–2 hero slots that may have unique high‑res textures; all other slots must use shared atlases.
  • Ensure slot shells match across sets so a single base texture can support multiple geometry variants via masks and decals, not new sheets.
  • For seasonal/event looks, prioritize palette and decal swaps over new geo; when new geo is essential, limit it to overlay pieces that reuse existing atlases.

Camera‑aware LOD boards

Create a standard “camera ladder” board: portrait UI, dialogue close‑up, over‑shoulder, gameplay mid, crowd/raid distance. For each rung, provide thumbnails of the head, hair, and outfit indicating which features survive (✓) and which convert (→) or drop (×). This becomes the LOD contract for modeling and tech art.

Streaming and loading considerations (design levers)

  • Favor few, reusable atlases to cut streaming stalls when swapping outfits in UI.
  • Design capsule swaps (e.g., hair → tied variant; cape → short wedge) that change only a single slot rather than forcing a full set reload.
  • Keep texture page counts per character predictable to reduce cache thrash during party scenes.

Platform targets and compression hints

  • If targeting low‑end or Switch‑like hardware, note texture ceilings (e.g., 2×2K per hero set, 1×1K for hair, 1×512 decals) and favor BC1/BC3‑friendly albedo/alpha layouts in concepts (avoid subtle gradients dependent on BC7 for readability).
  • Where skin tones are critical, propose shared gradient ramps rather than per‑character 2K albedos.

Authoring LOD paintovers

For every hero concept plate, include a mini LOD paintover strip:

  • LOD0: full trims, fine stitching, multiple seam families, hair cards.
  • LOD1: remove minor seam families, merge ornaments, hair ribbons.
  • LOD2: collapse bevels, block trims, hair cap texture.
  • LOD3: billboard/impostor where valid (crowd scenes), emblem → color block. This visual contract helps modelers simplify without fear of style drift.

Budget boxes and acceptance tests

Attach a budget box to each sheet: expected tri range, materials per slot, texture sheets and sizes, overdraw risk notes (hair, capes). Add an acceptance test paragraph: “At gameplay camera, face landmark ellipse remains distinct; at LOD2, eye sockets still read; emissive trims do not bloom into skin; decals legible at ≥ 24 px height.”

Heads & hair example guardrails (ready to copy)

  • Head: 12–18k tris LOD0, 6–9k LOD1, 3–4k LOD2; 1 material (opaque skin); optional eye material.
  • Hair: 8–14k tris LOD0 cards, 3–6k ribbons LOD1, <1k cap LOD2; 1 material masked; no full‑screen transparency.
  • Eyewear/Visor: tinted opaque at LOD1+; transmission only in portrait/cinematic context. (Numbers are illustrative—replace with project targets.)

Outfits example guardrails (ready to copy)

  • Undersuit (soft): 1 material, lives on Soft Kit Atlas; trims carry identity at LOD2.
  • Armor (hard): 1–2 materials max (base + optional emissive); beveled motifs translate to baked normal at LOD1; no nested glass layers.
  • Cape: masked at LOD0, switches to ribboned wedge at LOD2; stays under 1 material.

Palettes & decals example guardrails (ready to copy)

  • Palettes: 1 LUT per faction; 2–3 mask textures (512–1k) shared across set; emissive map separate and small.
  • Decals: single shared 1k–2k sheet; legal‑zone map included in orthos; fallback solid badges for LOD2+.

Testing rituals you can run without a build

  • 50% zoom test: Downscale your plate to sim mip blur; anything that disappears should be demoted to trims, not core reads.
  • Grayscale/value test: Ensure silhouette and class reads survive without hue.
  • Alpha audit: Bucket everything into Opaque / Masked / Transparent; challenge any Transparent outside hair/visors.
  • Flipbook pop test: Thumbnail LOD0→LOD2 sequence and watch for popping; adjust which features drop per step.

Handoff package for production

Include with your concept:

  1. Camera ladder board with feature persistence marks.
  2. LOD paintover strip for head, hair, outfit.
  3. Budget boxes per slot and an at‑a‑glance material cap table.
  4. Atlas plan diagram showing which parts share which sheets.
  5. Decal legal zones and fallback icon set.
  6. Palette LUT + mask previews with value rails.

Common pitfalls (and fixes)

  • Alpha everywhere (lace, veils, tassels, feathers): Consolidate into masked chunks; provide opaque alternates for crowded scenes.
  • Ornament creep: Replace micro filigree with embossed fields that bake to normal.
  • Palette bloat: Too many unique albedos—move to LUT + masks.
  • Decal shimmer: Typography too fine—supply bold glyph versions for distant LODs.
  • LOD pop: Features drop abruptly—stage removals over two steps and align with natural seams.

Closing: style that scales

Performance is a design constraint, not a muzzle. When you specify how details degrade with dignity, which materials are allowed where, and how textures are shared instead of duplicated, you unlock bigger wardrobes, cleaner UI swaps, and smoother raids—without sacrificing identity. Make LOD and memory visible in your art, and your characters will ship both beautiful and fast.