diff --git a/README.md b/README.md index 7df9a18..9bcc08f 100644 --- a/README.md +++ b/README.md @@ -361,7 +361,7 @@ The .NET GPU worker that shells out to art_pipe. Three mutually exclusive execut | **Resident** | `ArtPipe:ResidentModel` | Keeps one model in VRAM for the container lifetime; consumes `art.model..tasks`; recycles after N jobs (e.g. SDXLLightning 500) | | **Model-queue** | `ArtPipe:ModelQueue` | Same per-model queue, but a short-lived subprocess + per-job lease — an **idle container holds zero VRAM**. This is the production doctrine | -Routing is API-side via `ArtStudio:StageModelMap` (production defaults: `modeling → Pixal3D`, `texturing → Hunyuan3DPaint21`) with worker-side `ArtPipe:PublishToModelQueues`. Resilience pieces: a stage-run watchdog, capped failover down the per-stage method list (`ArtStudioStageFailoverOptions` — `PinnedOnly` methods like the edit models and material variation are excluded from failover chains and generation pickers), and display-derivative generation (400 px AVIF thumbnail + display AVIF via the shared `cavif` wrapper). A separate single-replica `ArtStudioMetrics` handler folds lifecycle events into the metrics read model. +Routing is API-side via `ArtStudio:StageModelMap` (production defaults: `concept → Flux2Klein9B`, `modeling → Pixal3D`, `texturing → Hunyuan3DPaint`) with worker-side `ArtPipe:PublishToModelQueues`. Resilience pieces: a stage-run watchdog, capped failover down the per-stage method list (`ArtStudioStageFailoverOptions` — `PinnedOnly` methods like the edit models and material variation are excluded from failover chains and generation pickers), and display-derivative generation (400 px AVIF thumbnail + display AVIF via the shared `cavif` wrapper). A separate single-replica `ArtStudioMetrics` handler folds lifecycle events into the metrics read model. ### GPU coordinator & VRAM leases @@ -380,7 +380,9 @@ Each model runs as its own Swarm stack with fully **baked weights** (`HF_HUB_OFF | Stack (`spikersoft-artpipe-…`) | Stage / role | VRAM (MB) | |---|---|---| | `modeling` (prodstages monolith) | all 7 stages serialized, per-task leases | per-stage | -| `-model-sdxl` | concept (SDXL Lightning, resident) — currently stood down (replicas 0) | — | +| `-model-sdxl` | concept text-only, inpaint, background removal, and upscale (SDXL Lightning) | 9,000 | +| `-model-flux2klein9b` | concept default (FLUX.2 Klein 9B) | 18,000 | +| `-model-flux2dev` | concept quality option (FLUX.2 dev, method-pinned) | 20,000 | | `-model-flux2klein` | concept `edit_image` (FLUX.2 Klein) | 13,000 | | `-model-qwenedit` | concept `edit_image` (Qwen-Image-Edit; also needs ~15 GB host RAM for the offloaded text encoder) | 16,000 | | `-model-triposr` | modeling `image_to_3d` | 3,000 | @@ -391,8 +393,8 @@ Each model runs as its own Swarm stack with fully **baked weights** (`HF_HUB_OFF | `-model-hunyuanomni` | modeling `image_to_3d` with **bbox proportion control** (Hunyuan3D-Omni) | 11,000 | | `-model-textto3d` | modeling (Hunyuan3D text/image→3D) | 14,000 | | `-model-pixal3d` | modeling `image_to_3d`, PBR output — **production default** | 18,000 | -| `-model-hunyuan` | texturing `texture_mesh` (Hunyuan3D-Paint 2.0, diffuse) | 14,000 | -| `-model-hunyuan21` | texturing `texture_mesh` (Hunyuan3D-Paint **2.1**, full PBR maps) — **production default** | 21,000 | +| `-model-hunyuan` | texturing `texture_mesh` (Hunyuan3D-Paint 2.0, diffuse) — **production default** | 14,000 | +| `-model-hunyuan21` | texturing `texture_mesh` (Hunyuan3D-Paint **2.1**, full PBR maps) | 21,000 | | `-model-blender` | resident Blender (rig/animate/export/enrich) | ~2,048 | | `-model-safety` | resident SafetyCheck gate | ~1,024 | | `-model-qrmonster` | QR Art lane (on-demand, replicas 0 when idle) | 9,500 |