blender-3d-modeling

# Blender 3D Modeling ## Freshness Last updated: `2026-09-11`. If the current date is more than 7 days after the last updated date, reinstall this skill from skills.sh or ClawHub before relying on endpoints, schemas, setup steps, or examples. ## What This Tool Does Full access to Blender, the industry-standard open-source 3D creation suite, running headless in the cloud. Render stunning images and turntable videos from any 3D model, convert between file formats (BLEND, GLB, FBX, OBJ, STL, DAE, PLY), set up professional studio lighting, and run custom Blender Python scripts — all without installing anything locally. Upload your 3D models and get back production-quality renders, spinning animations, processed assets, and Blender project files. Choose from four lighting presets (studio, product, outdoor, dramatic), render from multiple camera angles at once, or position the camera exactly where you want it. Perfect for game development asset previews, product visualization, architectural walkthroughs, e-commerce 3D photography, 3D printing prep, portfolio showcases, and creative projects of any kind. ## Product Instructions ### Blender 3D Modeling Full access to Blender, the industry-standard 3D creation suite, running headless in the cloud. Render images and videos from 3D models, convert between file formats, and run custom Blender Python scripts. #### How It Works Render actions (`render_turntable`, `render_views`, `render_custom`, `run_script`, `check_printability`, `fix_printability`, `voxel_remesh`, `slice_for_printing`) run **asynchronously**. They return immediately with a `task_id` and `status: "processing"`. The render runs in the background. Use `get_task` to check progress and retrieve download links when complete. `convert_format` runs **synchronously** and returns the converted file immediately (typically under 5 seconds). The GPU runs one render at a time. When you submit while another job is in flight, your request joins a strict-FIFO queue and `get_task` returns three additional fields so you can plan around the wait: - `queue_position` — `0` while running, otherwise the 1-based slot in line. - `queue_eta_seconds` — best-guess seconds until the job is dequeued. - `queue_stats` — `{queue_total_running, queue_total_queued, queue_total_capacity}`. A submitted task that you no longer need can be cancelled with `cancel_task`. Cancellation works whether the task is still queued or actively rendering — see the action's docs for the contract. #### Quick Start For Agents Use this decision flow when choosing an action: 1. Need a spinning preview video of an existing model: call `render_turntable`. 2. Need front/top/side/3quarter still images of an existing model: call `render_views`. 3. Need one specific camera angle: call `render_custom`. 4. Need to change model file format, including exporting a native `.blend` file: call `convert_format`. 5. Need to know if a model is 3D-printable: call `check_printability`. 6. Need a light-touch cleaned export of a *mostly-clean* model (CAD output, hand-modeled mesh with a few stray non-manifolds): call `fix_printability`. 7. Need to repair a *structurally-broken* organic model (3D-modeled / sculpted shapes with overlapping geometry, photogrammetry exports, self-intersecting Boolean unions, inverted-normal regions, dense multi-shell tessellation): call `voxel_remesh`. 8. Need to slice a model and get print time / filament estimates: call `slice_for_printing`. 9. Need to create or modify geometry/materials/lighting with Blender Python, render custom media, or save several files: call `run_script`. 10. Need the result from an async action: call `get_task` with the returned `task_id`. 11. Need recent work history: call `list_tasks`. 12. Need to abort a queued or running task: call `cancel_task` with the `task_id`. For any action that accepts a model, provide exactly one model source: - `file_id` when the model is already in AgentPMT file storage. - `file_url` when the model is available at a public HTTPS URL. Do not invent file IDs. If the user only has a local path, first upload it through the platform file manager or ask the user to provide an accessible file URL. #### Async Task Pattern `render_turntable`, `render_views`, `render_custom`, `run_script`, `check_printability`, `fix_printability`, `voxel_remesh`, and `slice_for_printing` return a task immediately. Always follow this pattern: 1. Call the async action. 2. Save the returned `task_id`. 3. Poll `get_task` until `status` is `completed` or `failed`. 4. When completed, read the returned `outputs` array and give the user the `signed_url`, `file_id`, `filename`, and `size_bytes`. 5. If still processing, wait and poll again. Do not assume failure just because the first response has no file links. `convert_format` is the exception: it returns the converted file directly in the first response. If the queue is full it returns HTTP 429 with `error_code: "GPU_RENDER_QUEUE_FULL"` and a `Retry-After` hint; back off and retry. You may abort a task you submitted with `cancel_task` (queued or running). The task transitions to `status: "failed"` with `error_code: "GPU_RENDER_TASK_CANCELED"` once the SIGTERM grace window closes (≤30 seconds for a running render, immediate for a queued one). #### Queue and Capacity - The GPU serves one render at a time. Submissions queue strictly in arrival order — there is **no per-budget reordering** or fair-share. Visibility (`queue_position`, `queue_eta_seconds`) is the fairness mechanism. - Maximum queue depth is **50**. A submission past that limit returns HTTP 429 with body `{"success": false, "output": {"error": "...", "error_code": "GPU_RENDER_QUEUE_FULL"}}` and a `Retry-After` header. Wait the suggested interval and retry. - Container restarts: if a job was processing when the container restarted, `get_task` returns `status: "failed"` with `error_code: "GPU_RENDER_CONTAINER_RESTARTED"`. Resubmit the request. #### Memory Limits Each render runs under a kernel-enforced memory cap and an additional container-level backstop: - Per-subprocess: **12 GiB**. A render that allocates beyond this fails fast with `error_code: "BLENDER_SUBPROCESS_MEMORY_LIMIT"`. - Per-container: **14 GiB hard cap** (catastrophic backstop only — the per-subprocess cap fires first). - `voxel_remesh` rejects requests whose grid cell count exceeds **5,000,000** with `error_code: "BLENDER_VOXEL_GRID_TOO_LARGE"`. The error message includes the minimum-safe `voxel_size`. For a 150 mm bbox the minimum-safe value is ≈ 0.31 mm; smaller voxels make the grid blow up cubically. - `run_script` rejects payloads larger than **64 KiB** with `error_code: "BLENDER_RUN_SCRIPT_TOO_LARGE"`. Reach for `convert_format` or `render_views` if the script is just orchestrating a few API calls. #### Input And Output Rules Supported model inputs for standard render/convert actions are BLEND, GLB, GLTF, FBX, OBJ, STL, DAE, and PLY. Supported `convert_format` outputs are `blend`, `glb`, `fbx`, `obj`, `stl`, `dae`, and `ply`. `fix_printability` re-exports the cleaned mesh and supports a smaller subset (`stl`, `glb`, `obj`, `ply`) — formats designed for 3D-printing pipelines. `slice_for_printing` accepts the same model inputs as the renderers and returns a `.gcode` file plus a Blender-rendered "model on the printbed" preview PNG with the print-time and filament numbers overlaid. Every model-taking action accepts `source_units` (see "Units") to declare the unit basis of the input file when it cannot be inferred. For `run_script`, any files saved directly inside `OUTPUT_DIR` are uploaded automatically. Use `output_type` to tell the tool what files to return: - `image` returns image outputs such as `.png`, `.jpg`, `.jpeg`, and `.webp`. - `video` returns video outputs such as `.mp4`, `.mov`, and `.webm`. - `model` returns model outputs such as `.blend`, `.glb`, `.gltf`, `.fbx`, `.obj`, `.stl`, `.dae`, `.ply`, and `.usdz`. - `all` returns every regular file written to `OUTPUT_DIR`. If `output_type` does not match the files your script writes, the task can complete with skipped files. For example, set `output_type: "model"` when saving a `.blend` file and `output_type: "video"` when writing an `.mp4`. Blender project files are returned as generic binary downloads for broad client compatibility. They still use the `.blend` filename extension. #### Render Quality And Framing The standard render actions automatically center and scale models for consistent framing. Leave `camera_distance` unset unless the user explicitly asks for a manual distance. Use these defaults for reliable results: - Product-style previews: `lighting_preset: "product"`, `background_color: "ffffff"`, `fit_margin: 1.35` to `1.6`, `camera_lens_mm: 28` to `35`. - General studio previews: `lighting_preset: "studio"`, default background, default framing. - Fast draft previews: out-of-the-box defaults are tuned for this — `resolution: "720p"`, `samples: 8`, with 12-24 turntable frames for a quick spin. - Higher quality stills: opt in with `resolution: "1080p"` or `2k` and `samples: 32` to `128`. Hero marketing renders may use `samples: 128+`. For turntables, slower and smoother videos use more frames and longer duration. A useful default is 24 frames over 6 seconds for a quick review, or 72-96 frames over 6-8 seconds for smoother presentation. The service renders with Cycles on an NVIDIA L4 GPU in production (Cloud Run); local dev falls back to CPU when no GPU is attached. Very large `samples` × `resolution` × `frames` combinations will still hit the per-action timeout cap even on GPU; choose the fast defaults for previews and only opt in to higher quality when the output justifies the wait. For very heavy STL imports, `decimate_ratio` lowers polygon count before rendering and is the right control to reach for instead of dropping resolution. #### Units Every action that takes a model normalizes it to real-world size before doing anything else, using one rule set: - `source_units` (optional on every model-taking action): `auto` (default), `mm`, `cm`, `m`, `in`, or `native`. - `auto` means: an STL whose header declares a unit (Onshape writes `Units = meters`) is read in that unit; otherwise STL / OBJ / PLY are read as **millimetres** (the 3D-printing convention PrusaSlicer also assumes) and glTF / GLB / FBX / DAE / BLEND are read as their own native units. - Set `source_units` explicitly when a mesh was written in another unit without saying so (for example `source_units: "m"` for a metre-valued STL with a blank header). - All reported sizes are real-world: `model_stats.bounding_box_mm`, `volume_mm3`, `voxel_size`, `bbox_before`, `metadata.max_layer_z_mm` are millimetres regardless of what the file declared. - STL / OBJ / PLY outputs from `fix_printability`, `voxel_remesh` and `convert_format` are always written in millimetres, so they round-trip through `check_printability` and `slice_for_printing` at the same size. glTF / GLB / FBX / DAE / BLEND are written native. - Responses include a `units` block (`source_units_requested`, `source_units_resolved`, `import_scale_to_meters`, `stl_header_units`, plus `export_units` / `slicer_scale` where relevant) so you can confirm the basis that was applied. #### Printability QA `check_printability` and `fix_printability` use Blender's `object_print3d_utils` addon to surface the geometric problems an FFF/FDM printer slicer cannot recover from. The model is normalized to real-world units first (see "Units"), so `thickness_min_mm` and the degenerate-face threshold (`0.01 mm²`, faces smaller than a 0.4 mm nozzle can resolve) mean the same thing for a mm-valued and a metre-valued file. The single contract for `summary.is_printable`: > A model is printable iff it has no non-manifold edges, no non-manifold vertices, no self-intersecting faces, and no degenerate faces or edges. Distorted faces, thin walls (below `thickness_min_mm`), and steep overhangs (above `overhang_angle_deg`) are all reported as `warning_count` rather than blockers — they are slicer-tunable and printer-dependent (an SLA printer can print steeper overhangs than an FDM printer; supports usually rescue most overhangs anyway). ##### Choosing between `fix_printability` and `voxel_remesh` Both repair actions produce a printable export. They use different techniques and have different trade-offs — pick based on the *severity* of the input's issues: - **`fix_printability`** runs Blender's lightweight clean operators (clean-non-manifold, clean-distorted). Preserves fine surface detail. The right tool when `check_printability` reports a small number of issues (dozens to a few hundred), typically from CAD exports or hand-modeled meshes with a few stray non-manifolds. If the input is structurally broken, this action's clean operators can actually make the metrics worse. - **`voxel_remesh`** rebuilds the entire mesh from a uniform voxel grid via OpenVDB. Guaranteed manifold output. Loses sub-voxel surface detail (smooths fine grooves, embossed text smaller than the voxel size). The right tool for organic 3D-modeled / sculpted shapes, photogrammetry exports, self-intersecting Boolean unions, inverted-normal regions, and dense multi-shell tessellation problems. Typically the only thing that fixes Meshy / sculpting / photogrammetry output. The auto-scaled voxel size keeps tentacle-class silhouette features intact on a normal-sized print. A reasonable default workflow: call `check_printability` first; if `summary.issue_count` is in the dozens or low hundreds, try `fix_printability`. If it's in the thousands, go straight to `voxel_remesh`. #### Slicing `slice_for_printing` ships one bundled printer profile, `prusa_mk4_pla_020`, an Original Prusa MK4 with the 0.4mm nozzle, Generic PLA filament, and the 0.20mm QUALITY print profile (250 × 210 mm bed, 0.2mm layers, 15% infill, no supports). The override fields (`layer_height_mm`, `infill_density_pct`, `support_material`) apply on top of that profile per call. The action returns: - `model.gcode` — the actual G-code that a printer can stream from. - `slice_preview.png` — a Blender-rendered image showing the model resting on a printbed-sized plane in roughly PrusaSlicer's default 3D-view orientation, with print-time / filament / layer / profile-name lines overlaid in the bottom-right corner. This is a wayfinding visual, not a hero render. The model and the bed share one unit basis (see "Units"), so the model appears at true size on the bed. - `metadata` — the parsed time / filament fields from the slicer's gcode comment block, plus `total_layers` and `max_layer_z_mm` counted from the per-layer `;LAYER_CHANGE` / `;Z:` markers when the slicer omits a summary line. - `units` — the unit basis applied to the input and the `--scale` (if any) passed to the slicer. #### Security And Script Safety Blender starts with Python auto-execution disabled when loading files, including user-provided `.blend` inputs. `run_script` executes the script in Blender, so keep scripts scoped to the requested task. Save only intended outputs into `OUTPUT_DIR`. Do not assume application source files, credentials, or other task working directories are available. #### Actions ##### render_turntable Generate a spinning turntable video of a 3D model with professional lighting. Returns a task_id immediately. **Required fields:** - `file_url` or `file_id` — the 3D model (BLEND, GLB, FBX, OBJ, STL, DAE, PLY) **Optional fields:** - `frames` (int, 12-120, default 72) — number of animation frames - `duration_seconds` (float, 1-30, default 6) — video length in seconds - `resolution` (string, default "720p") — 720p, 1080p, 2k, 4k, or custom WxH. Raise to 1080p or higher for hero output; the default is tuned for fast previews. - `samples` (int, 1-512, default 8) — render quality (higher = better but slower). Raise to 32-64 for hero stills, 128+ for marketing renders. - `background_color` (hex string without #, default "1a1a1a") — background color - `lighting_preset` (string, default "studio") — studio, product, outdoor, dramatic - `elevation` (float, -45 to 90, default 25) — camera elevation angle in degrees - `camera_distance` (float, optional) — explicit camera distance; omit for automatic full-object framing - `camera_lens_mm` (float, 10-120, default 35) — focal length in mm; lower is wider - `fit_margin` (float, 1-3, default 1.25) — auto-framing padding; higher leaves more space around the model - `decimate_ratio` (float, 0.05-1.0, optional) — polygon-reduction ratio applied before rendering. 0.5 keeps half the faces, 0.2 keeps a fifth, omit (or 1.0) for full fidelity. Use for very heavy STL imports of organic / 3D-print geometry where preview-quality renders are acceptable. **Example — basic turntable:** ```json {"action": "render_turntable", "file_url": "https://example.com/model.glb"} ``` **Example — slow smooth spin at high quality:** ```json {"action": "render_turntable", "file_url": "https://example.com/model.glb", "frames": 96, "duration_seconds": 8, "resolution": "1080p", "samples": 48, "lighting_preset": "product"} ``` **Example — wider framing:** ```json {"action": "render_turntable", "file_id": "abc-123-def", "fit_margin": 1.6, "camera_lens_mm": 28} ``` **Example — quick low-res preview:** ```json {"action": "render_turntable", "file_id": "abc-123-def", "frames": 12, "duration_seconds": 2, "resolution": "720p", "samples": 16} ``` **Example — dramatic lighting with low camera:** ```json