MIT JavaScript DSH plugin that automates a text → concept image → Blender grey-box pipeline, with AABB-validated layout plans, engine-agnostic `.blend` / `.fbx` / `.glb` + `layout.json` output and an optional Unity batchmode adapter.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-11-16
License
MIT
01What can it help you accomplish?
Turn one sentence into an editable grey-box scene skeleton inside a DeepSeek Harness session: concept image first, then a Blender scene with lighting, camera and props
An editable grey-box scene (hierarchy + lighting + camera + parameterised props) exported as `scene.blend` / `scene.fbx` / `scene.glb` plus `layout.json` under `<outputDir>/jobs/<jobId>/` — a blockout you keep editing, not a playable level
Solo and offline game developers who need a layout blockout before any art exists, and want it produced from a prompt in the same session they are already working in
Make the layout follow the written description instead of the fixed courtyard composition — a fountain on the east side, four stone columns in the corners
A plan JSON with a unique `id`, a `type`, `position` and `size` per object, where `position` is the bottom centre of the object and `size` the full extent in metres (Unity Y-up); Blender then assembles simple primitives at exactly those coordinates and sizes
Level and technical designers who need exact coordinates rather than an auto-composed arrangement, and can inspect the saved plan before assembly
Hand the generated scene to your own engine, or let the optional Unity adapter assemble it for you
The same FBX/GLB plus `layout.json` (`groups[].objects[]` with `instanceName`, position, rotation and scale) for Unreal or Godot, or an assembled `Assets/Scenes/SceneSkeleton.unity` with lighting, camera and optional prefab when `enableUnityAdapter` is on
Unity teams that want the scene assembled in batchmode, and Unreal / Godot users who import the mesh themselves and place it from the layout file
02How to install into DeepSeek Harness
Prerequisites
- Node.js 20 or newer — the plugin uses ESM and `AbortSignal.timeout`; the runtime DSH ships with also works
- A DSH build that supports bundles (`dsh.bundle.patch` + `dsh.client`); the README reports testing against 0.2.1-rc.5
- ComfyUI started by you and listening on `http://127.0.0.1:8188` with at least one loadable checkpoint — the plugin does not install, start or stop it
- Blender 4.x driven as `--background --factory-startup --python`; the default `scenic` mode also needs `scenic-bpy` installed into Blender's own Python
- About 25 GB of disk for ComfyUI, a checkpoint, Blender and an optional Unity project; PBR texturing adds about 6.4 GB, and 8 GB of VRAM is enough but slow (12 GB or more is recommended)
- Only for the optional Unity adapter: Unity 2022.3 LTS or newer with working batchmode and an activated licence
- No textures and no models are bundled — `assetLibraryPath` and `assetModelPath` point at files you supply, and anything missing falls back to flat colour or a procedural primitive
Installation steps
- 01
From a local checkout run `dsh plugin --profile web add /absolute/path/to/dsh-scene-skeleton-pipeline`, then start DSH with `dsh web`
$ dsh plugin --profile web add /absolute/path/to/dsh-scene-skeleton-pipeline
- 02
No YAML edit is needed: the profile `package.json` gains the dependency, `dsh.profile.bundles` gains `dsh-scene-skeleton-pipeline`, and `dsh.bundle.patch` mounts it
- 03
Under Settings → scene skeleton pipeline set the three fields the core path needs: `comfyUrl` (default `http://127.0.0.1:8188`), `blenderExe` (empty = auto-detect), `outputDir` (empty = `$DSH_HOME/scene-skeleton-pipeline/outputs`)
- 04
For the default `scenic` mode install scenic-bpy into Blender's own Python: `"/Applications/Blender.app/Contents/Resources/4.5/python/bin/python3.11" -m pip install --user "scenic-bpy>=1.19.0"`
- 05
Optional extras: `unityProjectPath` + `enableUnityAdapter` for Unity assembly, and `enableTextureGeneration` after cloning ComfyUI-SuperPBR and placing the SuperMat / MoGe / VAE models
Verify the integration
- Read the load log: `[dsh-scene-skeleton-pipeline] loaded`, `service registered: scenePipeline`, `tool registered: check_environment`, `settings routes registered`, `pipeline engine ready: comfy -> blender`, `pipeline tool registered: run_scene_pipeline`, `plan tool registered: plan_scene_layout`
- Run a dry run first — the panel button 「运行 dry-run」 or `run_scene_pipeline({ prompt: "一座带台阶的日式庭院,四角有石柱", dryRun: true })`; it only checks dependencies and wires nodes, so a failing environment shows the node as `blocked` and the rest as `skipped`
- Without launching ComfyUI or Blender: `bash scripts/verify/all.sh --skip-real` verifies only host wiring, layout and export
03DSH integration and capability boundaries
DSH bundle plugin — installed into a profile with `dsh plugin --profile web add /absolute/path/to/dsh-scene-skeleton-pipeline` and auto-mounted through `dsh.bundle.patch` + `dsh.client`; it registers the `scenePipeline` service, the `check_environment` / `run_scene_pipeline` / `plan_scene_layout` model tools and a settings panel at `/plugins/dsh-scene-skeleton-pipeline/*`, with config kept in `$DSH_HOME/scene-skeleton-pipeline/config.json`
Environment detection
no arguments — it only probes the machine→`{ checkedAt, environment: { blender, comfy, unity } }`, each item reporting `ok` / `missing` / `error` with a detail and an optional path
probe only — no software is installed, started or stoppedeach of the three dependencies is reported separately, so a single failing one can be fixed on its ownConcept image through ComfyUI
the prompt, plus an optional API-format workflow JSON (`comfyWorkflowPath`), checkpoint name (`comfyCheckpoint`) and negative prompt (`comfyNegativePrompt`)→a rendered concept image written into `comfy/` inside the job folder
a real run launches ComfyUI (about 1–1.5 minutes) and Blender headless (about 2–5 seconds)only API-format workflow JSON is accepted — a UI-format ComfyUI workflow is not supportedBlender headless grey-box skeleton
`blenderSceneMode`: `scenic` (default) with `scenicWidth` / `scenicDepth` / `scenicWallHeight`, or `legacy`→a parameterised courtyard — terrain, columns, walls, archways, trees, benches, lanterns — grouped as Terrain / Buildings / Props, or fixed blocks plus a reference-image plane in `legacy` mode
`scenic` mode needs `scenic-bpy` installed in Blender's own Python; `legacy` mode works without itthe result stays a grey box — replacing it with real art assets is your jobModel-written layout plan with AABB validation
`sceneDescription` via `plan_scene_layout`, with optional `width` / `depth` and a `savePath` to write the plan to disk→`{ ok, plan, errors, model, savedPath? }` — objects drawn from the 18-type asset-library whitelist, every violation listed at once with rule, message and affected objects
structure, scene bounds, ground penetration, floating, collision and support are all checked at AABB levelBlender validates the plan once more before assembly: an invalid plan aborts the run instead of being silently acceptedEngine-agnostic export and the `layout.json` contract
`layoutAxis` = `z-up` (default, Blender native) or `y-up` (values converted for Unity)→`scene.blend` + `scene.fbx` + `scene.glb` + `layout.json` carrying the hierarchy, ambient and directional lighting, camera and per-object transforms
downstream parsers must read `axis` first — files without that field are treated as `y-up`the layout `scale` is an extra per-instance scale, constant `[1,1,1]` from Blender; editing it in the file still scales that instanceOptional PBR texturing with automatic lite fallback
`enableTextureGeneration`, `textureResolution` (256–2048) and `textureObjects` = `main` (ground plus building objects) or `all`→per-object `_albedo` / `_normal` / `_roughness` / `_metallic` / `_ao` / `_height` maps in `textures/` plus a `LLM_<id>_pbr` material that overrides the flat colour
each selected object costs one ComfyUI inference, so `all` is slow and a single job textures at most 12 objectsmissing SuperPBR / MoGe nodes or models skip the whole stage — the main pipeline still finishes
04Who is it for? When not to use it?
Good for
- Solo and offline game developers who need a layout blockout before any art exists, and want it produced from a prompt in the same session they are already working in
- Level and technical designers who need exact coordinates rather than an auto-composed arrangement, and can inspect the saved plan before assembly
- Unity teams that want the scene assembled in batchmode, and Unreal / Godot users who import the mesh themselves and place it from the layout file
Not for
- The pipeline does not take over ComfyUI, Blender or Unity: it neither installs nor stops them, it offers no MCP bridge, and ComfyUI only accepts API-format workflows on a single-machine `127.0.0.1` endpoint (`comfyUrl` can be changed).
- No collision bodies, no rigid-body physics, no navigation and no walkability guarantee — validation is AABB-level static rules only (bounds, ground penetration, floating, collision, support).
- Budget about 25 GB of disk for ComfyUI, a checkpoint, Blender and an optional Unity project, plus about 6.4 GB when PBR texturing is on; 8 GB of VRAM works but is slow, 12 GB or more is recommended.
05Compatibility, maintenance and safety notes
- The pipeline does not take over ComfyUI, Blender or Unity: it neither installs nor stops them, it offers no MCP bridge, and ComfyUI only accepts API-format workflows on a single-machine `127.0.0.1` endpoint (`comfyUrl` can be changed).
- No collision bodies, no rigid-body physics, no navigation and no walkability guarantee — validation is AABB-level static rules only (bounds, ground penetration, floating, collision, support).
- Budget about 25 GB of disk for ComfyUI, a checkpoint, Blender and an optional Unity project, plus about 6.4 GB when PBR texturing is on; 8 GB of VRAM works but is slow, 12 GB or more is recommended.
- The geometry is procedural or declarative, so a concept image's architecture is not reproduced, and the model's layout succeeds only most of the time — JSON extraction, structural/physical validation and 3 retries reduce but do not remove bad plans.
MIT · no release yet · package version 0.1.0, verified with the in-repo `scripts/verify` suite
06Frequently asked questions
How does dsh-scene-skeleton-pipeline install and register inside DeepSeek Harness?
Install it from a local checkout with `dsh plugin --profile web add /absolute/path/to/dsh-scene-skeleton-pipeline` and start DSH with `dsh web`. The profile `package.json` gains the dependency, `dsh.profile.bundles` gains the package and `dsh.bundle.patch` mounts it, so no YAML edit is required. On load it registers the `scenePipeline` service, the `check_environment` / `run_scene_pipeline` / `plan_scene_layout` model tools and its settings routes; the README lists the exact log lines to compare against.
What does a session actually do, and how do I try it without launching ComfyUI or Blender?
Call `run_scene_pipeline({ prompt, dryRun: true })`, or press the panel's 「运行 dry-run」 button: a dry run only checks the dependencies and wires the pipeline nodes — a failing environment marks the node `blocked` and the following nodes `skipped`, and the artifacts are placeholders. With `dryRun: false` the run launches ComfyUI (about 1–1.5 minutes) and Blender headless (about 2–5 seconds) and writes `comfy/`, `blender/` and optionally `textures/` / `unity/` under `<outputDir>/jobs/<jobId>/`; the tool result's `artifacts` holds the absolute path of every output.
What has to be installed on the machine, and what does the plugin touch?
Node 20+, a DSH build with bundle support (tested against 0.2.1-rc.5), a ComfyUI you started yourself on `http://127.0.0.1:8188` with at least one loadable checkpoint, Blender 4.x (plus `scenic-bpy` in Blender's Python for the default `scenic` mode), about 25 GB of disk, and — only when `enableUnityAdapter` is on — Unity 2022.3 LTS+ with batchmode and an activated licence. The README states the plugin does not take over software lifecycles: it neither installs nor stops ComfyUI, Blender or Unity, and it does not bridge MCP.
What can the generated scene not do?
It is a grey box, not a playable level: the README lists no collision bodies or rigid-body physics (validation is AABB-level static checks), no navigation, triggers or interaction logic, no geometry reconstructed from the concept image, and image-to-3D is still on the roadmap. Textures and GLB models are not bundled either — without `assetLibraryPath` / `assetModelPath` the run falls back to flat colours and procedural primitives, and the scene still completes.
07Related DSH workflows
dsh-our-free-model
by ebony-vinyl
Adds model choices, availability checks, usage panels, and a local OpenAI-compatible forwarder.
dsh-routing-suite
by yjh051108
dsh-routing-suite — injector + router-standard kit: install the runtime injector first, then the task-aware reasoning-mode router preset (measured P1-P23).
dsh-ops
by t-auto
Run Bash and PowerShell commands, search and edit local files, and manage long-running background jobs from DSH.
dsh-connect-workbuddy
by dingminhua
Connect locally signed-in WorkBuddy models to DeepSeek Harness with a read-only credits overview and model management.
08Data and sources
**DSH 编排插件:一段文字 → ComfyUI 出概念图 → Blender 生成引擎无关的场景骨架(`scene.blend` / `.fbx` / `.glb` + `layout.json`),可选 Unity 装配适配器。**
[dsh-scene-skeleton-pipeline] tool registered: check_environment
安装后 profile 的 `package.json` 会多出依赖,`dsh.profile.bundles` 里出现 `dsh-scene-skeleton-pipeline`,
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-16. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.
