Skip to content

Photoreal Blender scenes from a one-paragraph brief, built by AI agents: image→3D props on fal, scripted layout, lighting, camera and grade.

Python
8.1 KB
Raw Blame

Ramen → Unreal 5.8: fix log

Copies: scripts in anyscene/unreal_ramen/ (pristine originals as copied at 20:17 in _orig/), scene copy anyscene-unreal/ramen/ramen_copy.blend, project anyscene-unreal/AnySceneRamen/. Patch of all script changes: ramen_fixes.patch (diff -ru _orig ., 216 lines).

G = generic (mergeable into anyscene/unreal/), R = ramen-specific (lives in anyscene-unreal/ramen/work/).

Mergeable change list (generic)

# File Change Why
G1 export.sh Run the export_blend.py next to the script, not $A/unreal/export_blend.py Per-scene copies of unreal/ silently ran the shared exporter (the one other jobs are editing).
G2 export.sh EXPORT_NOQUEUE=1 skips pipeline/gpu_run.sh CPU bakes don't need the GPU; with an Unreal job holding the queue the export just waited.
G3 export_blend.py Mantaflow domains (FLUID modifier, DOMAIN) go to fx (kind: mantaflow_gas, cache path, bounds) instead of sample_volume The slice bake can't see the cache grids; it read the steam shader's constant density 8.0 and would build an opaque local-fog box. Steam is a Niagara job (NIAGARA.md).
G4 build_level.py ANYSCENE_HERO_CAM picks the PlayerStart camera (was hard-coded cam_street, fallback = first camera) Scenes without cam_street got whatever camera came first.
G5 render_mrq.ps1, perf.ps1 Full -dpcvars set (asset/texture/static-mesh async compilation off, Niagara wait-on-activate) Was Editor.AsyncAssetCompilation=0 only.
G7 build_level.py stage_materials All-glass meshes (no baked slots): assign the glass MIs from section 0 glTF drops the unused BAKED slot 0, leaving 1 section, which got the untextured baked MI. Windows, door panes, fridge door, water glasses and jars rendered opaque (11 assets here; HK likely has the same).
G8 export_blend.py ensure_bake_uv Set active_render on the original UV by name after uv_layers.new("bake") Biggest visual bug. uv_layers.new() reallocates the collection; setting active_render through the stale reference cleared it on every layer, so image nodes with an implicit UV (and Texture Coordinate → UV) sampled a single texel. Every re-unwrapped poster, calendar, flyer, sticker, menu card, decal, neon sign and the UV-driven tile wall baked to a flat colour (37 of 152 smart_project assets here). Verified in isolation (work/uvdbg2.py: after the old code both layers read active_render=False). HK/caustic/bakery/MC exports are very likely affected too. Not fixed upstream as of 21:50.
G9 export_blend.py (Rewire orm, cuts_out, e["alpha"]) + build_level.py (master_baked(masked=True)) Principled Alpha baked into ORM.R (was a constant 1); assets with a linked/<1 alpha get M_AnyBakedMasked (opacity mask = ORM.R, clip 0.33) Luminance-keyed decals (floor grime, drips, splatter, rings) and stickers rendered as opaque dark/light squares. MaterialExpressionDitherTemporalAA is not exposed to Python in 5.8, so it's a plain clip. 30 assets here. Assets that are opaque keep M_AnyBaked (R=1, unchanged).
G10 export_blend.py --only (note, not fixed) --only matches object names, not asset names (assets named from mesh data, e.g. decal_floor_001 ← object decal_floor.001) Re-exporting by asset name silently skipped 17 of 42. Workaround: map assets → object names via placements. Suggest matching an too.
G6 perf.ps1 r.setres <ResX>x<ResY>w added to -ExecCmds Same fix mission-control landed upstream (their version also adds a comment line); take upstream's. Confirmed needed: the log shows 1066x600 until r.SetRes = "2560x1440w" on frame 1, and the CSV metadata then reads 2560x1440.

Ramen-specific

# Where Change Why
R1 work/prep_copy.py 5 decal/poster images with //..\..\refs\ramen\tex paths re-pointed (absolute) at blender-scene/refs/ramen/tex; steam cache dirs re-pointed at the original scenes/ramen/cache/r2 (read only) Plain file copy breaks relative paths (make_copy.py would have had to open the original, which is open in the user's GUI Blender).
R2 work/prep_copy.py Added cam_bowl (shots.json bowl_insert key: pos (3.2, 2.3, 1.42), aim bowl + 7 cm, 100 mm f/2, focus 1.52 m) The blend has no close-up camera; shot cameras are created by the camrig at render time.
R3 export args --frame 1 (heroes and render.py stills use frame 1), --grade-sh scenes/ramen/finish.sh, CapCut ffmpeg Scene's grade lives in finish.sh.

Overlap with upstream (anyscene/unreal/, mission-control's fixes landed after my copy)

  • G7 ≈ upstream "sections-based material mapping for all-glass meshes". Upstream's is better (exporter writes the real sections list); merge upstream's and drop G7.
  • G6 = upstream perf.ps1 r.setres. Take upstream's.
  • Upstream's export.sh crash-resume, GLB-exists done test and the ffmpeg-from-env change were already in my copy.
  • Not upstream yet, merge these: G1, G2, G3, G4, G5, G8 (flat posters/decals), G9 (alpha cutouts), G10 (note). G8 hunk applies on top of upstream as-is: upstream line ~544 still sets ar.active_render = True on the stale reference.

Results

Level /Game/AnyScene/Ramen/L_Ramen: 296 mesh actors (274 unique meshes), 23 lights (8 spot, 8 rect, 7 point), 7 cine cameras, height fog (haze_volume) + 1 local fog volume (kitchen_vapor), grade PPV with finish.sh LUT. PlayerStart = cam_wide. Steam (2 Mantaflow domains) and rain (35k particles) are in scene.json fx, not built.

Perf (standalone -game, offscreen, 2560x1440, Lumen HWRT + VSM + Nanite + volumetric fog, 841 frames from cam_wide, GPU otherwise idle at the time: 1.4 GB used before the run): 85.6 fps mean · frame 11.7 ms (p95 12.7) · GPU 11.2 ms (p95 11.8) · game 1.2 ms · peak GPU memory +3.9 GB. Top GPU passes: Lights 6.1 ms (23 movable shadowed lights; half the frame), TSR 0.9, shadow depths 0.7, Lumen reflections 0.6, Nanite vis 0.3, volumetric fog 0.3.

Stage timings (machine shared with ~5 other Blender/Unreal jobs, CPU at 100%, RAM ~0 free during export):

stage time
scene copy (1.27 GB blend + 1.2 GB lib) + path prep ~1 min
export (CPU bake, 274 assets, one Blender OIIO crash + resume) 41 min wall (2164 s second run)
UE build, full (textures 60 s, meshes 801 s, materials 216 s, level 4 s, sequences 10 s) 18 min
fix pass: re-export 42 assets (2 runs) 1.5 min
fix pass: reimport + materials/level/sequences ~1.5 min
MRQ stills, warm DDC (wide / counter / bowl) 39 / 20 / 20 s (cold first pass: 246 / 108 / 56 s)
Cycles reference, cam_bowl, 256 spp OptiX 65 s render, 69 s wall
perf run (900 frames) 37 s

Honest read of the match

  • Layout, props, lighting placement, framing: near-identical in all three views. Every prop, lamp, sign, stool, poster and decal is where Cycles has it; shadows from the pendants and the counter spots fall the same way.
  • Biggest miss: steam. Cycles' counter and wide shots are dominated by a bright steam wall over the stove; UE has none (Niagara gap, NIAGARA.md). It also changes the mood: Cycles' kitchen glows, UE's kitchen is clear and harder.
  • Tone: UE is flatter and slightly brighter in the shadows, less of Cycles' warm crush (AgX vs UE filmic + LUT; the open AgX tone-match item).
  • Backsplash: stainless at roughness 0.19 / metallic 1 reads as crinkled mirror foil under Lumen vs grey grime in Cycles. The bake is correct; the procedural grime (AO + noise) gets lost at 1024x4096 and Lumen reflects the lamps harder. Would want a roughness floor or AO-in-roughness tweak (not done).
  • Glass: water glasses read milky/plastic (translucent master, no refraction, white tint); Cycles' are dark and refractive. Window/door glass is fine (rain-on-glass texture lost).
  • Close-up: a stray white sliver on the counter left of the bowl in UE (not in Cycles); likely one of the masked decals/Planes clipping differently. Not investigated.
  • Faint counter ring stains (decal_rings) are mostly clipped away by the 0.33 mask.