[ArtStudio] ProArt capability parity — user/predefined workflows, multi-method stages, full PBR map sets, rig-method chain, curation (epic #346 phase 2) #378
Assignees
spikerj (Joseph Spiker)
Clear assignees
spikerj (Joseph Spiker)
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: spikerj/spikersoft-issues#378
Reference in New Issue
Block a user
Joey (2026-07-05): Eric's original was more capable than what the platform exposes — multiple model-generation methods, texturing with map production, multiple rigging methods via user-defined or predefined workflows, each operation an eventhandler (.py where best). Correct. The engine parity is nearly complete; the platform exposure was deliberately narrowed for the student MVP. This ticket restores the breadth deliberately.
Honest parity matrix (engine vs platform-exposed today)
spikersoft-artpipe, works via its own server/batch UI today)rig_modelconfig = the "user-defined workflow" seam)Phases
allowed_models(backend plumbing largely exists via StageSettings/model queues); composer "advanced" panel (staff first, students later behind a flag). Server-side failover chain per stage (ProArt's concept-failover pattern generalized).rig_model(auto/unirig/rignet/mesh2rig/smart_rig) through StageSettings params; deploy UniRig/RigNet residents on CUDA (#368 R4 pattern); record which rigger succeeded (P4 ticket already asked for this).Trellis3D.Pythoneventhandler pattern) instead of .NET-wrapping, where that's simpler — evaluate against #368's supervisor (which already speaks the worker protocol; the .NET layer buys leases/queues/telemetry, so likely keep it, but decide explicitly).Interim note: Eric's full standalone experience (batch UI :5000, Blender addon, all methods) still runs from the
spikersoft-artpipecheckout today — nothing was deleted; the platform simply hasn't exposed it all yet.W6 (staff curation) backend slice merged in spikersoft-backend PR #88. Additive-only (embedded
CurationRatingson ArtAsset mirroringQuarantinedArtifacts, #369 legacy-safe): staff quality ratings (POST gallery/{id}/rating, 1–5 + note, upsert-by-staff) and a curation query (GET gallery/curation— paged Approved assets with derived metrics/artifact sizes/chosen methods + ratings, sortable/filterable). 36 new tests, 437 ArtStudio suite green. Found no pre-existing student engagement to duplicate; enrichment metrics aren't stored as queryable fields (only a metadata-json GridFS blob) so the query surfaces queryable proxies — deep blob-parse deferred.#378 status — where the phases stand:
Ticket stays open for W3/W4/W5 + the W6 remainders.
W4 (rig-method chain) domain slice merged in spikersoft-backend PR #97. Restores rig-method breadth + records which method actually produced each result — additive, builds on W1/W1.5.
auto/unirig/rignet/mesh2rigalongsidesmart_rig(kept default → no-selection byte-identical); extras carry the newrig_model=<key>param seam (correction:rig_modeldidn't previously exist — the op isBlender/rig_mesh). Reaches the worker via the unchanged W1StageMethodspath; unknown rig key fails fast at startup.ArtAsset.StageMethodProvenance, #369-guarded likeCurationRatings): records the actual method per stage, capturing the failover winner (the orchestrator rewrites the selection to the succeeding alternative). Optionalmethod_usedworker result field is additive/back-compat — worker NDJSON contract unchanged. Exposed on asset reads + staff curation projection (chosen-vs-actual).ArtStudio:StageFailover:Stages:riggingwhen residents land).#378 status: W1 ✅ · W1.5 ✅ · W2 ✅ · W4 ✅ · W6 (curation) ✅ backend. Remaining: W3 (full PBR/texture map sets — domain slice self-mergeable, actual maps CUDA-gated) · W5 (.py-native eventhandler decision) · CUDA execution for W3/W4 (UniRig/RigNet/Hunyuan3DPaint residents on SERVER, #368 R4) · Angular composer/curation UIs. Ticket stays open.
W3 (full PBR map sets) domain slice merged in spikersoft-backend PR #98. The platform can now carry/store/expose/export full PBR map sets — additive, byte-identical for today's single-texture assets.
PbrTextureMaps.cs): base-color/normal/orm/metallic/roughness/ao/emissive kinds → glTF-2.0 viewer channels + export params + include-flags + filename tokens. Additive to the existingtexture-setfallback.Artifacts; #369-trivial since zero new fields; legacy → empty set).*_path(as before) + turns oninclude_normal/orm/emissive— byte-identical when absent (unclassifiable single texture → no flags, proven).Textures({kind, channel, url}— viewer channel-switcher contract) + curation projection (PbrMapCount/PbrChannels); safety-gate thresholds extended to new kinds. 40 new tests, 661 ArtStudio+ArtPipe green.#378 status: W1 ✅ · W1.5 ✅ · W2 ✅ · W3 ✅ · W4 ✅ · W6 curation ✅ backend. The ProArt parity DOMAIN layer is now substantially restored. Remaining: W5 (.py-native eventhandler decision) · CUDA/Metal model execution for W3/W4 (Hunyuan3DPaint full-PBR, UniRig/RigNet residents — SERVER #368 / Mac #358) · Angular composer/viewer/curation UIs · W6 multi-take. Ticket stays open.
W5 — Python-native eventhandlers: recommendation (for your sign-off)
Researched the
.py-native-vs-.NET-wrapper question against the actual code. Recommendation: confirm your hypothesis — keep the .NET wrapper (status quo, formalized by #368). Do not build a pure-.pyresident path. Evidence, specific to this stack:The
.py/.NET boundary is already drawn exactly at the model. The.pyworker is pure inference over the stdin/stdout NDJSON protocol (ready/started/progress/result/error/pong); it knows nothing about RabbitMQ, GPU leases, GridFS, Jaeger, the safety gate, or SignalR. The .NET host (ArtPipeProcessor) is the entire distributed-systems plane around it. Eric's design intent is already fully expressed — byte-preserved worker protocol, per-model venv isolation, one always-loaded model per service (#368, Mac-validated). "Python-native where best" is satisfied today: the.pyis the eventhandler body; .NET is only the transport shell.What the .NET layer buys is all load-bearing here, and Option B (pure
.py) doesn't remove a layer — it re-implements 6+ wire-compatible protocol clients in a second language: GPU leases (SharedGpuLeaseServiceRPC — there's aGpuLeaseWireCompatTestsguard because this wire format is a forever-contract), per-model queues, OTLP+W3C trace-context (house rule: wrong key = silent no-exporter, #307 found 7 invisible services), DLQ/retry, the fail-closed child-safety gate (least-desirable code to fork into a 2nd language), progress events, and 700+ lines of orchestration. No process hop is actually eliminated — the model runs in Python either way; the.NET↔.pyhop is a local pipe with the model already warm (#368: sdxl 4.2s warm).Criterion if hybrid is ever revisited (door explicit, not just shut): a stage may be
.py-native only if it simultaneously needs no GPU lease, produces no servable image (no safety gate), and needs no trace-continuity/DLQ/progress. For ProArt model stages that set is empty. So: ship Option A, don't hold W1–W4 breadth waiting on a.pypath.W5 disposition: decided → keep .NET wrapper;
.py-native rejected. No implementation ticket needed (it's the status quo).BUT the analysis surfaced a real orthogonal bug → filing separately (see next). The safety gate routes
SafetyCheckthroughIArtPipeStageExecutor, which in resident mode is theResidentArtPipeStageExecutorthat throws on any non-resident model — so a resident image-gen deployment would fail-closed-quarantine every output. Must be fixed before the #368 SERVER cutover.