nishi code wiki / research / style lens

The style lens: one simulation, many art styles

SOTA census · compiled 2026-08-01 · 6 axes · ~45 sourced claims · 9 declared gaps

lineage: forked from research_rtrender parent domain: rendering goal: photoreal / anime-cel / western cartoon / painterly / pixel, switchable at runtime over ONE simulation

Claims are labelled SHIPPING / RESEARCH / ANNOUNCED / DEPRECATED. Gaps are declared UNVERIFIED rather than guessed. Engine constraint filter: WebGL2 / GLSL ES 3.0, vertex+fragment only (no compute guaranteed), integer-math core, no third-party libraries.

1. Anime / cel — the shipped technique stack

ArcSys (the reference cel recipe): a hard-step NdotL terminator instead of a smooth ramp; hand-edited vertex normals so the terminator falls where an animator would draw it, not where geometry says; the ilm control texture — per-channel masks for specular intensity, shadow bias (which pixels resist falling into shadow), highlight size and inner line art; and inverted-hull outlines — the mesh redrawn flipped and expanded along normals, with vertex colour driving per-region outline thickness and Z-offset. All from the 2015 talk; it remains the canon a decade later. Everything here is vertex+fragment only — no compute anywhere in the stack.

HoYo (anime faces at scale): a plain NdotL terminator looks wrong on flat anime faces, so Genshin ships pre-baked face SDF shadow maps: a texture encoding the shadow line's sweep as the light circles the head (R channel covers 0–180°, G the mirror), sampled by comparing against the facing-vs-light angle (FdotL) — the sun sweeps, and the face shadow animates along an ARTIST-authored path. Plus per-character shadow ramps (the shadow's colour is a lookup, not a multiply) and a fake-SSS warm band at the terminator that reads as subsurface scattering for free. [SECONDARY — all of this is community shader archaeology, flagged as needing verification; treat channel conventions as per-implementation.]

Hi-Fi Rush (the architectural lesson): toon rendering was decoupled from G-buffer rendering — the G-buffer carries semantics, and the style (comic shader, toon light, toon face shadows, outlines) is applied in deferred passes, with post-process volumes controlling where stylization applies. The whole world — not just characters — runs through it at 60 fps native res on UE4. This is the single most load-bearing precedent for a style LENS as opposed to a pile of per-material style shaders.

2. Western toon, watercolour, and the 1-bit dither class

Obra Dinn renders internally in 8-bit grayscale and converts to 1-bit in a single post pass with a tiling dither pattern. The detail that matters: a screen-space dither pattern swims the moment the camera moves. Pope's fix was to anchor the dither to the 3D camera — mapping the pattern onto a sphere around it — so rotating the view keeps dots pinned to the scene; translation still swims but is masked by full-screen change. He also moved to a blue-noise variant because it survives screen scaling and video compression far better than Bayer. The entire lens is one fragment pass plus a texture — the cheapest complete style in this report.

Watercolour has a real-time recipe with a paper trail: MNPR implements edge darkening, paper distortion/granulation, pigment turbulence and colour bleeding as viewport passes with painted per-object control parameters — i.e. watercolour is ALSO a post-stack over semantic inputs, same shape as the cel lens. Western-cartoon flat shading is the cel stack minus the anime face machinery, plus bolder outline width tables; The Plucky Squire ships full 2D-page↔3D-world switching as a core mechanic (style AND projection switch over one game state).

3. Style switching over one unchanged simulation — shipped precedents

Minecraft + Iris is the strongest precedent, because it runs on exactly this class of content. The shaderpack format is a fixed contract of named programs: gbuffers_* geometry passes plus deferred/composite/final fullscreen passes (up to 99 numbered composites) — and the geometry passes take vertex and fragment shaders only, no compute (shaders.properties program reference). The game engine never changes; the pack is pure data + GLSL against stable semantic buffers. Result: one simulation, an open-ended market of lenses. That IS the target architecture, proven at ecosystem scale. [synthesis: the “ABI” framing is this report's; Iris documents a format, not a contract promise.]

4. Real-time neural stylization — the honest verdict

Verdict: a diffusion-class anime lens at 60 fps in a 2026 browser is not feasible. The 60 fps budget is 16.7 ms/frame; the best measured in-browser single image is ~1 s with a 1-step distilled model on WebGPU — ~60x over budget before adding any temporal coherence, on an API a WebGL2-only engine does not even have. The systems that DO hit real-time (MirageLSD, StreamDiffusionV2) are datacenter products: the browser could only ever be a thin client to a rented H100, which violates the sovereignty constraint outright. The feasible neural unit in-engine remains the tiny-MLP class discussed in research_rtrender — nothing diffusion-shaped.

5. Playable generative world models — impressive, and not runtimes

Verdict: these are not runtimes. None offers determinism, none offers save states, and the frame-generating class holds a world together for minutes at best (Genie 3: “a few minutes”; Matrix-Game: minute-level; Oasis visibly dream-drifts within one). A game runtime's first obligation — the same state tomorrow, byte for byte — is absent by construction. [inference over absence: no capability list for any of these advertises determinism or save/load; none was tested here.] Their honest role in this stack is offline asset generation — Marble already IS that product (splat/mesh export into a real renderer), Muse says “ideation” on the tin — and never the runtime. Waypoint-1.5 running 720p on a local RTX 3090 is the number to watch, not to adopt.

6. What this means for the Nishi stack

Constraint filter: WebGL2 / GLSL ES 3.0, fragment+vertex only (no compute guaranteed — same constraint Iris gbuffers passes live under), integer-math core, no third-party libraries. Everything ranked below fits it; the two NOT-adoptable classes are named at the bottom.

Not adoptable as runtime: (a) diffusion-class neural lenses — ~60x over the 16.7 ms budget in-browser even 1-step distilled (axis 4); (b) generative world models — no determinism, no save states, minutes of persistence (axis 5); their role is offline asset generation feeding the real renderer, exactly where Marble already positions itself.

Declared UNVERIFIED — do not treat as measured