[bug] Alternating lighting on rotated aircraft segments: axis normal used for dynamic models #123

New Issue

2026-11-04T13:19:12Z

chexit commented

2026-11-04 13:19:12 +00:00

🤖 Filed by Codex, chexit's AI agent, with GPT-6.1 Terra.

During #115 we captured Lancet through the real render.set-models path with glow=false: before/after and reproducible preview, commit 914f723. Adjacent segments of equally inclined wings have alternating brightness. These are actual dynamic model instances, not the opaque-block placement test whose black surfaces were a separate test artifact.

Code inspection: horizon-assets/src/bake.rs::facing quantizes the quad normal to one of six directions; the dynamic block-model vertex path in horizon-client/src/game/minecarts.rs derives lighting normals from q.facing. Agent's float32 analysis of 32 faces at +/-45 degrees selected Up 20 times and Side 12 times due to rounding near equal components. Code plus native images support the diagnosis, but no engine fix or comparative GPU validation has been done.

Proposed fix: retain the geometric normal, or reconstruct it from non-degenerate quad positions in the dynamic model path, with correct winding and degenerate handling. Do not mask the effect with fullbright/light_emission or incorrect cullface. Acceptance: consistent shading across equal slopes, checks for both signs of 45 degrees and other legal rotations, and no regression of axis-aligned models/night/HD. Related #115. Leave issue open for review.

🤖 Filed by Codex, chexit's AI agent, with GPT-6.1 Terra. During #115 we captured Lancet through the real render.set-models path with glow=false: [before/after and reproducible preview](https://gitea.maoniao.team/chexit/airdef/src/branch/kimi-work/docs/AIRCRAFT-MODELS.md), commit 914f723. Adjacent segments of equally inclined wings have alternating brightness. These are actual dynamic model instances, not the opaque-block placement test whose black surfaces were a separate test artifact. Code inspection: horizon-assets/src/bake.rs::facing quantizes the quad normal to one of six directions; the dynamic block-model vertex path in horizon-client/src/game/minecarts.rs derives lighting normals from q.facing. Agent's float32 analysis of 32 faces at +/-45 degrees selected Up 20 times and Side 12 times due to rounding near equal components. Code plus native images support the diagnosis, but no engine fix or comparative GPU validation has been done. Proposed fix: retain the geometric normal, or reconstruct it from non-degenerate quad positions in the dynamic model path, with correct winding and degenerate handling. Do not mask the effect with fullbright/light_emission or incorrect cullface. Acceptance: consistent shading across equal slopes, checks for both signs of 45 degrees and other legal rotations, and no regression of axis-aligned models/night/HD. Related #115. Leave issue open for review.

chexit referenced this issue

2026-11-04 13:21:56 +00:00

[enhancement] Повторный художественный проход моделей: текущий ремейк не достиг нужного качества #115

chexit referenced this issue from a commit

2026-11-04 18:15:37 +00:00

Fix dynamic model lighting and document GPU comparison (#123)

chexit commented

2026-11-04 18:16:26 +00:00

🤖 Agent: Kimi (primary).

Реализовано и опубликовано в kimi-work (503e6e29a), engine 45b43f00d входит в итоговый 278f356aa ветки airdef.

Dynamic block models теперь используют нормаль из неколлинеарных вершин квада, с прежним осевым fallback для вырожденной геометрии. Unit-тест покрывает ±45°, другие наклоны, осевые и вырожденные квады. На реальном Lancet через set-models, glow=false и HD получены парные кадры до/после в одном мире/ракурсе, вид сверху и ночной кадр: чередующиеся полосы устранены. Галерея.

Массовый FPS и полёт этим не измерялись. Issue оставляю открытым для ревью.

🤖 Agent: Kimi (primary). Реализовано и опубликовано в `kimi-work` (`503e6e29a`), engine `45b43f00d` входит в итоговый `278f356aa` ветки `airdef`. Dynamic block models теперь используют нормаль из неколлинеарных вершин квада, с прежним осевым fallback для вырожденной геометрии. Unit-тест покрывает `±45°`, другие наклоны, осевые и вырожденные квады. На реальном Lancet через `set-models`, `glow=false` и HD получены парные кадры до/после в одном мире/ракурсе, вид сверху и ночной кадр: чередующиеся полосы устранены. [Галерея](https://gitea.maoniao.team/chexit/airdef/src/branch/kimi-work/docs/AIRCRAFT-MODELS.md). Массовый FPS и полёт этим не измерялись. Issue оставляю открытым для ревью.

chexit referenced this issue

2026-11-05 21:29:06 +00:00

[enhancement] Отдельный RT-пресет из этапа 10 HD-плана: тени и отражения, затем GI #125

chexit referenced this issue

2026-11-05 21:32:16 +00:00

[performance] Архитектурный аудит FPS и микрофризов: GPU instancing, объёмные эффекты и разделение simulation/render #126

chexit closed this issue

2026-11-06 11:45:30 +00:00

Sign in to join this conversation.