Scenario Maya Fx

Maya FX (Nucleus, Bifrost, MASH, caches)

An expert FX TD fixes scale and inputs before touching a knob, changes few settings one at a time, never judges a live solve, and proves the result on the cache in numbers (penetration, stretch, pops, convergence) and in renders from the shot camera. 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-maya-expert (execution channel, review loop, 2027 version traps).

Status (2026-09-24): Maya 2027 is not installed. The pure-Python layer of scripts/mx_fx.py passed offline (tests/code/maya-fx/test_fx_offline.py); every Maya call is not yet run in Maya; names marked [verify] are answered by job_00_fx_probe.py (python3 tests/code/maya-fx/run_all.py).

Stance (the expert delta)

• Scale before anything. Nucleus and Bifrost compute in meters whatever Maya's unit: Space Scale and scene_units_in_meters 0.01 in a cm scene, or a 182 cm character is 182 m to the solver (2027 Help; SARKAMARI [00:29:22]; Stamatelos [00:30:21]). MPM detail_size is in meters, so a unit change silently changes resolution (Gast [00:07:29]).
• Never simulate the production mesh. Duplicate the skinned body and garment, drive the duplicates with a blend shape at weight 1, hide the originals (SARKAMARI [00:10:04]).
• Give the solver a clean start. Hold 25 frames, blend 25 frames into the first pose, action at 0, Nucleus start -50 (SARKAMARI [00:05:26]); Julan: A-pose at 0, walk by 24.
• Substeps before stiffness. Collision iterations below the substep count do nothing; high stretch resistance at low substeps cannot converge (2027 Help, Tips). SARKAMARI's remaining penetration vanished at 30 substeps and 35 iterations, with Space Scale moved to 0.25 in the same pass [00:29:22] [00:29:57].
• Cache, judge the cache, render the cache. nCache before every playblast; the render scene holds Alembics only (SARKAMARI [00:19:32], [00:40:07]).
• Bifrost fails silently. Unpulled outputs are never computed, empty arrays are valid data, mismatched arrays truncate (Jason E3aOt5wuvoo [00:22:49]; Nordenstam [00:38:11]; graph doc). A sim solves only forward from its own start frame: default 1, and it must exist in the timeline (sim guide; Nordenstam [00:56:23]).
• Do not simulate what you do not need. Start a snowball at the collider with its impact velocity (Gast [00:55:18]); layer garments through Alembic colliders (SARKAMARI [00:32:06]); coarse background sources (Stamatelos [00:21:25]).
• One change per run, logged (Jason cd3HgvM8sXg [00:21:43]); mx_fx.set_attrs returns the log.

Establish first

Workflow

1. Scale and inputs. mx_fx.scale_report(meshes, expected_size_m=...); if the model is not real size, fix the asset (Julan [00:02:02]). Record any deliberate non-physical scale (SARKAMARI's 0.25 on a 182 cm character) with art_choice=True. GATE: no error line.

2. Pick the system (options keyed to the brief):

   Effect	Default	Alternative and when
   Cape, garment on a character	nCloth	Bifrost MPM cloth when it must tear, live in a Bifrost graph, or go to a game via USD (Jason hmuWr2nWSms, yo6y7cANfX0)
   Dust puff, smoke, mist	Bifrost Aero	add MPM sand or snow for dirt clumps; nParticles for grit (doc-level only)
   Snow, sand, mud impact	MPM granular	none in the sources
   Scatter, crowds of copies, motion graphics	MASH (Repro)	Instancer for speed; Bifrost scatter_points for USD point instancers
   Liquids	basic_liquid_graph	legacy Bifrost Fluids is gone

3. Prepare cloth inputs. Topology gate mx_fx.cloth_topology_verdict(mx_audit.audit(garment), expected_border_loops=1): evenly spaced quads (cross links need quads), manifold, single-sided, holes sealed (SARKAMARI [00:03:05]). Pre-roll mx_fx.add_preroll(ctrls, action_start, rest=bind_pose); duplicates at the bind frame with mx_fx.sim_duplicate. GATE: verdict clean, the duplicate follows the original within 0.01 cm.

4. Build.

   • nCloth: mx_fx.build_cloth_rig(garment_SIM, body_SIM, action_start, preset=...): one nucleus, Space Scale for the unit, start at the pre-roll, first pass substeps 10 / iterations 12, preset values, thin cloth and collider. Pin a cape collar with an input-attract lock map (value 1 plus Input Attract Method "Lock values of 1.0 or greater", Collide Last Threshold 1.0, [added] recipe from the doc) or Point to Surface with Exclude Collisions (SARKAMARI [00:23:40]), members on the input mesh (output-mesh members risk a DG loop, doc; point_to_surface(..., cloth=) rewrites them). GATE: rig["verdict"] has no error.
   • Bifrost: start from basic_*_graph and Explode (Stamatelos [00:39:07]); script it with mx_fx.BifrostGraph (logs every VNN call) or reuse a GUI Echo All Commands log. Hide what you bring in (Jason). Open or flat emitters (ring, ground patch) need Shell, closed ones Solid (Stamatelos [00:10:40]). Set the sim's start frame. GATE: aero_source_verdict(...) on emitter_facts(mesh) and sim_range_verdict(...) clean; every output pulled; point or voxel count above 0 on every frame.
   • MASH: mx_fx.mash_network(sources, geometry="Repro") [verify MASH.api]; unit-size sources (Waters). GATE: mx_fx.mash_cache_reasons(mx_fx.mash_inventory(waiter)) read.

5. First pass cheap, then cache. Low resolution (Aero detail 0.1, cloth first pass), frames evaluated in order from the start frame (mx_fx.run_sim, step_frames), every frame played in the GUI. After a geo_detail_size or time_step_size change, or a compound created or exploded, re-run from the start (sim guide). GATE: nothing non-finite, no empty frame.

6. Measure and look. rep = mx_fx.sim_report(cloth_out, body_SIM, start, end, action_start=..., thickness=...), rep["convergence"] = mx_fx.convergence_test(...), mx_fx.sim_verdict(rep, budget_spf=...) (adds the doc's Timing Output gate: solve seconds per frame against the brief). Then look: headless mx_review.review([cloth_out, body_SIM], out, views=("side","back","threequarter"), modes=("clay",), focus=fixed) at the action start, the peak and the last frame; GUI mx_review.playblast from the shot camera; volumes with mx_fx.volume_review (Arnold, tricubic). Judge with references/critique.md.

7. Triage in this order (never stiffen first): scale; start state (pre-roll, Push Out, Crossover Push with Trapped Check); substeps with iterations above; thickness down, Self Collide Width Scale up; stretch and compression together (SARKAMARI [00:24:50]); Bend Solver High Quality for flips; Collide Last Threshold 1.0 when attract or rigidity drags cloth through; lower Strength on popping constraints; "frame change too large" means play from the start or cache (2027 Help). Leftover pokes: a sculpted corrective on the cached mesh, or accepted under blur (SARKAMARI [00:37:12]). Aero detail: 2027 sharpening (aero_refinement_settings, low amount), adaptivity bounds, or post_refine_aero on the cache (frames in order, original sources and colliders) before lowering detail size everywhere (sim guide).

8. Final cache and deliver. Final pass settings, nCache from the nucleus start (ncache_verdict: mcc stops at 2 GB, use mcx; positions only, so the cloth transform must stay still), disable the nucleus, Alembic of cloth and body over the shot range with UVs and color sets (mx_fx.alembic_export); Bifrost file_cache with only the properties the renderer needs; a cache-only render scene. GATE: mx_fx.alembic_check and file_sequence_report clean, a look at the cache.