# ComfyUI-BrainDead

A comprehensive collection of ComfyUI custom nodes for caching, character consistency, prompt iteration, **3D mesh / PBR baking, character segmentation pipelines, image-to-3D generation, and game-asset preparation**. [![BrainDeadGuild](https://img.shields.io/badge/BrainDeadGuild-Community-purple)](https://BrainDeadGuild.com/discord) [![BrainDead.TV](https://img.shields.io/badge/BrainDead.TV-Lore-red)](https://BrainDead.TV) [![ComfyUI V3](https://img.shields.io/badge/ComfyUI-V3%20API-blue)](https://docs.comfy.org/custom-nodes/v3_migration) [![~110 nodes](https://img.shields.io/badge/nodes-~110-green)](nodes/)

--- ## 🎯 About BrainDead Nodes AN internally-used node pack for Biloxi Studios designed to help with character, story, 3D, and audio generation for next-generation UGC game and TV pipelines. All nodes use the **ComfyUI V3 API** and appear under the `🧠BrainDead` category in the node browser. --- ## Workflow Templates Eighteen ready-to-use workflows ship in [`example_workflows/`](example_workflows/) and appear in ComfyUI under **Workflow → Browse Templates → ComfyUI-BrainDead**, each with a thumbnail and an in-canvas note.

Pixal3D Image to 3D _{Image → MoGe FOV → 3-stage pipeline → mesh + voxelgrid → PBR bake}	CubePart Part Decomposition _{Mesh → open-vocab part names → TRIMESH per part → preview + export}	SAM3 Parts Segmentation _{Image → SAM3 multi-prompt → Parts Refine → Batch Edit → PSD export}
Lotus-2 Depth & Normal _{FLUX-based diffusion depth — far higher quality than feedforward estimators}	GLSL Skin Tinting _{4-output GPU shader: ILM / SR+Parts / Unity / Unreal skin-tone pipeline}	OVoxel PBR Bake _{TRELLIS2 voxelgrid → simplify → UV unwrap → albedo/normal/roughness/metallic}
TRELLIS2 Shape to Texture _{Voxelgrid → textured mesh with full PBR material bake via BD OVoxel Bake}	FaceWrap Pipeline _{MediaPipe landmarks → SAM3 guided face masks → socket infill → UV export}	Character Consistency _{Qwen-Image multi-view edit with identity lock, prompt iteration, save context}
Background Removal _{SAM3 + pymatting alpha matting → RGBA + white/black composites}	Face Segmentation _{MediaPipe + SAM3 → 25+ anatomy masks → UV-ready socket fill}	Channel Operations _{Pack / unpack / merge image channels (R/G/B/A round-trip)}
Mask Tools _{Luminance mask, flatten, crop-to-mask, fill-holes}	Image → Full PBR _{Remove BG → Lotus-2 depth+normal → SAM3 metal → Derive PBR (albedo/normal/rough/metal/AO + ORM/ARM)}	Game-Engine Packing _{Isolate parts → balance luma → crop & center → R/G/B channel pack (shared pivot, no bleed)}
Atlas / Flipbook _{Tile frames / packed textures into a grid sheet or sprite strip + per-cell UV layout}	TRELLIS2 → Unreal FBX _{Image → low-poly textured mesh + detail normal → single game-ready FBX (textures + vertex colors) via Blender}

> Templates appear in **Workflow → Browse Templates → ComfyUI-BrainDead** and are prefixed **`BD-`** so they're easy to find and identify as BrainDead workflow templates. --- ## Installation ### ComfyUI Manager Search for **"BrainDead"** in ComfyUI Manager and install. Dependencies in `requirements.txt` are installed automatically. ### Manual Installation ```bash cd ComfyUI/custom_nodes git clone https://github.com/BizaNator/ComfyUI-BrainDead cd ComfyUI-BrainDead pip install -r requirements.txt ``` ### Dependencies `requirements.txt` is installed automatically by ComfyUI Manager. It covers: | Package | Purpose | |---------|---------| | `moge` (git) | MoGe-2 monocular geometry for Pixal3D FOV estimation | | `utils3d` (git, pinned) | Required by MoGe `.pt` submodule — PyPI version is incomplete | | `pipeline` (git, pinned) | Required by MoGe | | `natten` | Neighborhood Attention for Pixal3D NAF upsampler | > **Pixal3D model weights** (~10 GB) are downloaded automatically on first run. > `cube_part` is vendored under `nodes/cubepart/vendor/` — no separate install needed. --- ## Features ### Cache Nodes (`🧠BrainDead/Cache`) Smart caching with **lazy evaluation** — upstream nodes are completely SKIPPED when cache is valid. | Node | Description | |------|-------------| | **BD Cache Image** | Cache IMAGE tensors as PNG files | | **BD Cache Mask** | Cache MASK tensors as PNG files | | **BD Cache Latent** | Cache LATENT tensors as safetensors | | **BD Cache Audio** | Cache AUDIO data as WAV files | | **BD Cache String** | Cache STRING data as TXT files | | **BD Cache Any** | Cache any data type as pickle | | **BD Cache Mesh** | Cache TRIMESH objects as PLY files | | **BD Save File** | Save any data in native format, output file path. Optional `context_id` + `suffix` for template-based saves | | **BD Bulk Save** | Save N inputs in ONE Run using a save context (16 typed slots + parallel labels) | | **BD Save Context** | Define a path template + variables once, downstream save nodes auto-resolve | | **BD Get Context Path** | Resolve a context's template into a STRING — wire into ComfyUI's built-in SaveImage `filename_prefix` or any third-party save node | | **BD Load Image** | Load image from STRING path | | **BD Load Mesh** | Load 3D mesh from STRING path | | **BD Load Audio** | Load audio from STRING path | | **BD Load Text** | Load text from STRING path | | **BD Clear Cache** | Clear cached files by pattern | **How Caching Works:** 1. Place a cache node after an expensive operation 2. First run: generates data, saves to `output/BrainDead_Cache/` 3. Subsequent runs: loads from cache, **skips upstream generation entirely** 4. Change the `seed` to force regeneration **Save Context System** — replace ad-hoc concat-string filename chains: 1. Add ONE `BD Save Context` node with a template like `%project%/%character%/%character%_%name%_v%version%%suffix%` and set vars (character, name, version, project, etc.) 2. Downstream `BD Save File` / `BD Bulk Save` nodes auto-pick the context (or specify `context_id` if multiple are registered) 3. Each save just sets its `suffix` (e.g. `_albedo`, `_skin_mask`, `_head`) — context resolves the rest 4. Per-save `custom_vars` (multiline `key=value`) layer on top of the context for one-off variations like `subfolder=PBR` or `materials=metal` 5. `BD Get Context Path` outputs a STRING path for ANY save node — works with ComfyUI's built-in SaveImage and third-party packs ### Character Nodes (`🧠BrainDead/Character`) ![Character consistency workflow](docs/images/nodes/character.png) Advanced tools for maintaining character consistency with Qwen-Image models. | Node | Description | |------|-------------| | **BD Qwen Character Edit** | Enhanced image editing with character preservation | | **BD Qwen T2I Custom** | Text-to-image with custom system prompts | | **BD Qwen Multi-Image** | Multi-reference image editing | | **BD Qwen Identity Lock** | Strong identity preservation with weighted references | **Features:** - Customizable system prompts for character consistency - Multi-image support (up to 3 reference images) - Per-image weight and role control - Identity-focused templates that prioritize facial features ### Mesh Nodes (`🧠BrainDead/Mesh`) 3D mesh processing, color sampling, simplification, and PBR texture baking tools. | Node | Description | |------|-------------| | **BD Load Mesh** | Load a mesh (GLB/GLTF/OBJ/PLY/STL/FBX/…) as TRIMESH. File picker over the input folder + upload button, or a `file_path` override. Outputs mesh (TRIMESH) + resolved mesh_path. Feeds any BD mesh node (CuMesh, Blender, CubePart, Export). | | **BD Cache Mesh** | Cache TRIMESH objects as PLY files | | **BD Sample Voxelgrid Colors** | Sample vertex colors from TRELLIS2 voxelgrid | | **BD Sample Voxelgrid PBR** | Sample full PBR attributes from voxelgrid | | **BD Apply Color Field** | Apply COLOR_FIELD to any mesh (deferred color application) | | **BD Transfer Vertex Colors** | BVH-based vertex color transfer between meshes | | **BD Transfer Colors Pymeshlab** | Transfer colors using pymeshlab | | **BD Mesh Repair** | Repair mesh topology (holes, normals, duplicates) | | **BD Smart Decimate** | Edge-preserving decimation with pymeshlab | | **BD Export Mesh With Colors** | Export mesh with vertex colors to GLB/PLY/OBJ. Preserves the mesh's baked PBR material; optional `diffuse`/`normal` IMAGE inputs embed those maps into the GLB (so the file is textured, not a white placeholder). Optional `context_id` (BD Save Context) for template-based naming | | **BD Trimesh → MESH** | Convert TRIMESH → ComfyUI native MESH (geometry only) to feed built-in 3D nodes (Save 3D Model / SaveGLB) | | **BD MESH → Trimesh** | Convert native MESH → TRIMESH to pull built-in Hunyuan3D/voxel results into the BD pipeline | | **BD CuMesh Simplify** | GPU-accelerated mesh simplification with color preservation | | **BD CuMesh Quad Remesh** | GPU-accelerated quad remeshing from triangle meshes | | **BD MeshLib Fill Holes** | Fill mesh holes and open boundaries via pymeshlab | | **BD UV Unwrap** | UV unwrap with xatlas (GPU) or Blender Smart UV | | **BD Planar Grouping** | Group faces by normal direction with boundary straightening | | **BD Combine Edge Metadata** | Combine edge metadata from multiple sources | | **BD OVoxel Bake** | All-in-one PBR bake from voxelgrid (simplify + UV + bake) | | **BD OVoxel Texture Bake** | Bake-only: takes pre-processed mesh + voxelgrid | | **BD Mesh To OVoxel** | Convert textured mesh to VOXELGRID format | | **BD Export OVoxel** | Export VOXELGRID to .vxz compressed format | | **BD Load OVoxel** | Load VOXELGRID from .vxz + sidecar files | | **BD Fix Normals** | Fix face orientation (Python-only, fast) | | **BD Pack Bundle** | Pack mesh + textures + colors into MESH_BUNDLE | | **BD Unpack Bundle** | Unpack MESH_BUNDLE into individual components | | **BD Cache Bundle** | Cache MESH_BUNDLE for fast reload | | **BD Mesh Inspector** | Inspect mesh properties (verts, faces, UVs, colors) + interactive three.js viewer | | **BD Mesh Preview (Thumbnails)** | Render a TRIMESH_LIST (e.g. CubePart `parts`) or single TRIMESH to a labeled, color-coded contact-sheet IMAGE (shown inline) + per-mesh IMAGE batch. Headless GPU (EGL) render. Wire `part_names` → `labels`. Shading: **`textured`** (default — renders the real texture atlas, **falls back to `vertex_colors` then `solid`** when no texture) / `vertex_colors` (mesh's own COLOR_0) / `part_colors` / `normals` / `solid`. Texture-look thumbnail for a DAM, no Blender needed. `views` (single mesh) renders a **turnaround** — e.g. 4 = front/right/back/left in one grid. | | **BD Preview 3D** | Show a mesh — or CubePart `parts` auto color-coded — in the in-node interactive three.js viewer (same viewer as Mesh Inspector). | | **BD Orient Mesh** | Rotate a finished mesh by X/Y/Z degrees, non-destructive (UVs, baked material, and COLOR_0 vertex colors ride along). Use to fix engine orientation — e.g. Pixal3D output → `rotate_x=180` to stand it upright facing forward. | | **BD Bake Vertex Colors From Texture** | Sample a diffuse/atlas through a mesh's UVs into per-vertex `COLOR_0` (bilinear/nearest). Runs on ANY UV'd mesh; keeps the UV + material intact (additive). Feeds COLOR_0 consumers (Pack Bundle, edge/stylized) and vertex-colored exports. | | **BD Detail Normal From Albedo** | Extract high-frequency detail from a diffuse/albedo atlas → tangent-space detail normal, UDN-blended onto a base (geometric) normal. Adds skin/fabric/fold micro-detail that a smooth high→low bake can't capture. UV-safe (per-texel). `detail_strength` + `high_pass`. | **Mesh types — `TRIMESH` vs native `MESH`:** BD nodes pass meshes as **`TRIMESH`** (a full `trimesh.Trimesh` carrying vertex colors, UVs, materials, and processing/export) — the same type TRELLIS2 / Hunyuan3d-2-1 / GeometryPack use. ComfyUI's built-in 3D nodes use the native **`MESH`** type, a thin batched `(vertices, faces)` tensor container with no colors/UVs. They're different data models, not interchangeable. Use **BD Trimesh → MESH** / **BD MESH → Trimesh** to bridge when you need a built-in 3D node — but keep work inside `TRIMESH` (and **BD Export Mesh With Colors**) whenever colors/UVs must survive, since native `MESH` can't represent them. **OVoxel Baking Pipeline:** ``` [TRELLIS2 Texture] → voxelgrid → [BD OVoxel Bake] → mesh + PBR textures OR [TRELLIS2 Texture] → voxelgrid → [BD OVoxel Texture Bake] ↑ mesh → [BD Blender Decimate] → [BD UV Unwrap] ↗ ``` **VXZ Caching Pipeline:** ``` [TRELLIS2 Texture] → voxelgrid → [BD Export OVoxel] → saved .vxz + .mesh.npz ↓ [BD Load OVoxel] → voxelgrid → [BD OVoxel Bake] → re-bake without regenerating ``` ### Blender Nodes (`🧠BrainDead/Blender`) Advanced mesh processing using Blender's geometry tools (requires Blender 5.0+, bundled in `lib/`). | Node | Description | |------|-------------| | **BD Blender Decimate** | Full-featured decimation with edge preservation | | **BD Blender Quad Decimate** | Quad-aware decimation: optional pre-quad-remesh, hole fill, and quad OBJ guide input | | **BD Blender Edge Marking** | Detect and mark edges from colors/angles | | **BD Blender Merge Planes** | Merge geometry within marked regions | | **BD Blender Remesh** | Voxel/quad remeshing with Blender | | **BD Blender Planar Normals** | Detect connected face groups by angle threshold; assign flat average normal as custom split normals; mark sharp edges between groups. Produces flat-panel stylized look for Unreal/Unity import | | **BD Blender Cleanup** | Advanced mesh cleanup and repair | | **BD Blender Vertex Colors** | Vertex color operations (bake, transfer) | | **BD Blender Normals** | Normal fixing and recalculation | | **BD Blender Export Mesh** | Export a MESH_BUNDLE (or a direct TRIMESH `mesh`) as **GLB or FBX** (`format`). FBX uses Blender's exporter with `embed_textures` → a single game-ready file carrying **both** embedded PBR textures **and** vertex colors (GLB can't hold both). Optional `flat_shading` for a clean low-poly no-smoothing look. | **Hard Edge Preservation Pipeline:** ``` [Input Mesh] ↓ [BD Planar Grouping] (straighten_boundaries=True) ↓ Clean geometric boundaries [BD Blender Edge Marking] (FROM_COLORS_AND_ANGLE) ↓ Mark SHARP/SEAM edges [BD Blender Merge Planes] (delimit_sharp=True) ↓ Dissolve while respecting marks [Low-poly mesh with hard edges intact] ``` **BD Blender Edge Marking Operations:** - `FROM_COLORS`: Mark edges where vertex colors differ - `FROM_ANGLE`: Mark edges by dihedral angle threshold - `FROM_COLORS_AND_ANGLE`: Combine both methods - `CLEAR`: Remove existing edge marks **BD Blender Merge Planes Features:** - **Delimit options**: Respect SHARP, SEAM, MATERIAL, NORMAL edges - **Dissolve angle**: Control coplanar face merging threshold - **Region subdivision**: Proportional face density based on area - **Output topology**: TRI, QUAD, or NGON output ### Prompt Nodes (`🧠BrainDead/Prompt`) Iterate through multiple prompts with automatic filename generation, plus per-Run iteration over typed inputs. | Node | Description | |------|-------------| | **BD Prompt Iterator** | Basic prompt cycling with filename generation | | **BD Prompt Iterator (Advanced)** | Templates, suffix lists, seed modes, ping-pong | | **BD Prompt Iterator (Dynamic)** | Up to 20 connected prompt inputs | | **BD Filename Template** | Standalone STRING resolver for `%var%` templates (no save side-effect) | | **BD For Each Run** | Multi-Run iterator over up to 16 typed inputs + parallel labels — drives any downstream chain (Qwen Edit + Save, etc.) where downstream needs to actually re-execute per item | **Iterator vs Bulk Save:** `BD Bulk Save` (in Cache) saves N files in ONE Run — use when downstream IS just save. `BD For Each Run` (here) emits one (data, label) per Run — use when downstream is more complex (e.g. per-iteration upscale before save). **Modes:** - `sequential`: Cycle through prompts in order - `manual`: Select specific prompt by index - `random`: Shuffle prompts randomly - `single`: Always use first prompt **Filename Modes:** - `auto_index`: `base_000`, `base_001`, etc. - `suffix_list`: `base_front`, `base_left`, etc. - `template`: `{base}_{index:03d}_{suffix}` ### Segmentation Nodes (`🧠BrainDead/Segmentation`) ![FaceWrap pipeline workflow](docs/images/nodes/facewrap.png) Character segmentation, parts pipeline, PBR map derivation, asset prep. **Parts pipeline** — single-execution character processing using the `PARTS_BUNDLE` type: | Node | Description | |------|-------------| | **BD Parts Builder** | Crop and composite each segmented part into its own image. Wire SAM3 `per_prompt_masks` + labels OR Sapiens2 `labels` directly. Outputs PARTS_BUNDLE + image_batch + label_list + bbox_list. Optional `combined_mask` silhouette clip and `depth_image` for per-part depth_median. | | **BD Parts Refine** | IoU-based dedup of overlapping prompts (e.g. "shoe" + "sneaker" + "left shoe" → 1 entry). Picks canonical tag, merges via union or keep_largest, optional max_parts cap and debug overlay. | | **BD Parts Batch Edit (Qwen)** | Internal-loop Qwen Image Edit per part, single execution. Modes: `flatten_redraw` (clean redraw on white bg, latent-upscale + tonemap recipe — internal patches: ModelSamplingAuraFlow + CFGNorm + Reinhard tonemap) and `true_inpaint` (regen only enclosed holes, preserves visible pixels). Auto-detects alpha from white bg, optional `flatten_pad_factor` for breathing room, optional `context_extend_factor` for surrounding context. | | **BD Parts Underpaint** | Background-reveal pass — removes masked objects and fills the exposed region with whatever would naturally be visible behind them. Runs after BD_PartsBatchEdit so the redrawn sprites are already clean. `model_type` selects the inpaint backend: **`qwen_edit`** (Qwen Image Edit — default, wire the same model/clip/vae as Batch Edit), **`kontext_dev`** (Flux Kontext Dev — better real-scene coherence; wire `flux1-kontext-dev.safetensors` + DualCLIPLoader + Flux VAE), **`flux_fill`** (Flux Fill Dev or any plain Flux inpaint model). `mode`: `per_part_sequential` removes objects one at a time (best coherence), `all_parts_combined` removes them all in a single pass (faster, uses `all_parts_prompt` where `{tag}`/`{tags}` expands to the full label list). `mask_dilate_pixels` (0–256) expands both the KSampler inpaint zone and the compositing boundary — raise to 64–128 to eliminate edge halos. | | **BD Parts Compose** | Flatten the bundle to a single RGBA + alpha at chosen `output_size`. Back-to-front by depth_median. | | **BD Parts Export** | Save files to disk: per-tag RGBA PNGs, per-tag depth PNGs, per-tag mask PNGs (original SAM3 visibility), composite PNG, layered PSD with per-part layers + optional per-part mask layers (visibility off, scaled identically). `composite_size` drives PSD canvas. SaveContext-aware. | **SAM3 multi-prompt + mask cleanup:** | Node | Description | |------|-------------| | **BD SAM3 Multi-Prompt** | One node replaces 12-chain SAM3Segment. **Standalone** — loads the SAM3 model + text encoder in-house (auto-downloads the official `Comfy-Org/sam3.1` checkpoint on first use), **no comfyui-rmbg**. Vote/weighted_vote modes (majority counting), positive + negative prompts, color filter (off/exclude/include/exclude_and_include/remove_matching), adaptive LAB color sampling (works for any category — skin, clothing, hair), silhouette enforcement, multi-instance collapse so labels stay aligned | | **BD Mask Resolver** | Python port of GLSL Mask Resolver shader. Priority-based separation (skin/clothes/accessories) with adaptive LAB color, neighbor-vote gap fill | | **BD Human Parser Mask Clean** | Per-class morphological cleanup + min-area filter on parse maps | | **BD Human Parser Mask Split** | Split a parse map into per-region MASK outputs | | **BD Human Parser Named Mask** | Extract a single label as MASK | | **BD Human Parser Preview** | Colorized RGB visualization of a parse map | **Human parser backends** (deferred for stylized art — trained on photos): | Node | Description | |------|-------------| | **BD Fashn Human Parser** | SegFormer-B4 from fashn-ai (NVIDIA license). 18 FASHN classes. | | **BD ATR Human Parser** | mattmdjaga/segformer_b2_clothes (MIT). 18 ATR classes. | **MediaPipe face masks & guided SAM3:** | Node | Description | |------|-------------| | **BD MP Face Mask** | Landmark-precise face region masks (face_oval, skin, eyes, brows, lips, nose, irises, ears, forehead, hair) from MediaPipe FaceLandmarker. Shared tight drawing: envelope brows, eroded-eyelid eyes, organic lips. | | **BD MP Face Export** | Passthrough node — writes the **landmark JSON** (478 pts) + reference RGBA mask PNG for the Blender face-plate UV pipeline. Tiny-detection guard (retry + padded fallback). | | **BD MP Save / Load Face Data** | Persist all region masks + head_mask + image to `.mpface.npz`/`.json`; reload later (after the image is processed and MediaPipe can no longer detect). | | **BD MP Face Refine** | Refine MediaPipe feature masks with a SAM3 segment batch (IoU match → intersect), then compute pixel-accurate skin. | | **BD MP SAM3 Face Segment** | **MediaPipe-guided SAM3**: localizes each feature with MediaPipe, then prompts SAM3 per feature (bbox + positive landmark points + sibling negatives) for pixel-accurate masks in one node. Solves the stylized-brow offset. Outputs match BD MP Face Mask. Cleanup (component-keep + optional hole-fill + edge smooth) and optional edge-snap refinement (`guided` / `matting` / `vitmatte`). **No wiring needed** — auto-loads + auto-downloads the official SAM3 checkpoint in-house (optional `MODEL` override). `skin_color_filter` (adaptive_lab / fixed_hsv / both) refines skin mask by colour — drops non-skin pixels (dark mustache, etc.). `drop_facial_hair` punches out uncoloured facial hair in the lower face zone. **Chin-safe neck removal** via face-parser + connected-component — never cuts a full-width jaw line. | | **BD MP Mouth Parts** | Separate an **isolated mouth render** into **lips / teeth / tongue** and pack them **RGBA for the Unreal viseme atlas: R=lips, G=teeth, B=tongue, A=POM** (wire depth into `pom`). Two engines: `color` (tuned-HSV, no model) and `sam3` (colour split seeds SAM3 to refine the tongue). Optional `edge_refine` + `despeckle`. Replaces the `SAM3×3 → FillHoles → PackChannels` chain in the lip-viseme atlas. | **Face socket / animation textures:** | Node | Description | |------|-------------| | **BD Face Socket Infill** | MediaPipe-based one-shot face socket creator for 2D animation flipbook textures. Fills eye, brow, lip, nose zones with flat/surround/inpaint fill. Per-zone independent expand_x/expand_y (elliptical kernel). `lip_mode`: `organic` (MediaPipe contour + lip_band), `contour` (exact landmark polygon, tight crop, no band), `plane` (rotated rectangle along 61→291 mouth axis — pre-draw for Qwen). `lip_plane` output always emits the rotated rectangle regardless of mode. | **Luma / lighting tools:** | Node | Description | |------|-------------| | **BD Luma Stats** | Compute luma statistics for an image: min, max, median, mean, and `recommended_outer_band = (max−min)/2 × 0.9`. Wire into BD_CenterMedianLuma or GLSL shader uniform inputs to set normalisation levels. | | **BD Center Median Luma** | Shift image luma so the median lands at 0.5. Optional `calc_source` input: compute the shift from a different (cleaner) reference image and apply it to `image`. Outputs `measured_median` and `shift_applied` for downstream diagnostics. | | **BD Normalize Luma** | Auto-rescale image luminance to a target range. Range-fit mode (default) remaps [src_min, src_max] → [target_min, target_max]. Proportional mode: single multiplier so src_max → target_max, blacks stay at zero. Mask-aware percentile clipping via `np.percentile`. | | **BD Depth To Shadow Map** | Convert a depth image to a shadow/shading map via analytic top-down N·L shading + cavity AO. Outputs `shadow_map`, `normal_map`, `cavity_map`, and `measured_max_curvature`. Use as a lighting guide for skin shaders or PBR derivation. | **Drawing tools:** | Node | Description | |------|-------------| | **BD Draw Rect** | Draw a flat-color rectangle with configurable rounded corners and edge feather. Shape source: manual geometry OR mask OR packed-channel extract. Optional alpha composite onto a background image and packed-channel injection. | | **BD Draw Ellipse** | Draw an ellipse/circle with flat color. Same shape-source options as BD_DrawRect (manual / mask / packed). Supports anisotropic radii and rotation. | **Mask manipulation:** | Node | Description | |------|-------------| | **BD Mask Correlate** | IoU-based matching of SAM3 candidate masks to up to 8 named target slots (e.g. MediaPipe hulls). Per-slot combine modes (`intersect`, `replace`, `union`, `weighted_blend`), post-match subtraction rules, and `combined_mask_invert` for head-minus-features mask. | | **BD Mask Resolver** | Priority-based pixel-level separation of skin/clothes/accessories. Python port of the GLSL Mask Resolver shader. Adaptive LAB color scoring, neighbor-vote gap fill, `priority` or `soft_blend` overlap mode. | | **BD Mask Color Fill** | Fill up to four mask regions with solid colors on a composited canvas. Per-slot expand/feather, optional background image, union mask output. | **Asset prep / game textures:** | Node | Description | |------|-------------| | **BD Mask Flatten** | Flatten RGB+alpha onto chosen background (white/black/colors/checker/custom hex/background_image). Channel-routing modes (image_to_red/green/blue/grayscale). `edge_pad_pixels` for Voronoi alpha bleed (game-engine UV edge fix). | | **BD Pack Channels** | Combine 4 IMAGE/MASK sources into RGBA channels. Each slot accepts IMAGE (auto-luminance OR alpha-channel for the alpha slot) or MASK. Auto-enables RGBA when alpha source is wired. | | **BD Atlas Pack** | Tile multiple images into one grid atlas (cols × rows, padding, bg colour). Sources concatenate in order: `images` batch → `image_1..image_8` slots → `masks` batch (per-cell alpha). `fit_mode` contain/cover/stretch, `order` row/column-major, `output_alpha` for RGBA. Outputs the atlas, a mask atlas, and a JSON `layout` with per-cell pixel + normalised-UV rects for game engines (sprite/viseme sheets). | | **BD Crop and Center** | Crop a part to its content bbox, then re-place it **at original scale** onto a fixed canvas (default = input size) at a chosen `anchor` (center / edges / corners). Gives every part a **common pivot** so they overlap correctly when channel-packed/atlased (engines extract from a shared pivot). Bbox from wired `mask` else image alpha/luminance; `pad`, `scale_to_fit`, `background_hex`. Outputs the recentered image, its mask, and a `transform` JSON (src bbox, placement, scale). Distinct from **BD Crop To Mask** (which crops + resizes to a square). | | **BD Derive PBR Maps** | Heuristic roughness/metallic/AO/normal from image + depth (+ optional normal_map / silhouette_mask / aux_shading_alpha / aux_detail_texture / metallic_zone_mask). Outputs each map separately + packed ORM/ARM. `albedo_treatment=edge_pad` (default) prevents transparent-PNG halo bleed. | **Recommended PBR pipeline** using existing depth/normal estimators (DepthAnythingV2 / Lotus / MiDaS / Metric3D — all installed via comfyui_controlnet_aux): ``` Source character image ├─→ ImageToMask (alpha) → silhouette_mask ├─→ DepthAnythingV2Preprocessor → depth (IMAGE) ├─→ Metric3D-NormalMapPreprocessor → normal_map (optional) ├─→ Skin shader alpha → ImageToMask → aux_shading_alpha └─→ Manga lineart ControlNet → aux_detail_texture ALL into BD Derive PBR Maps with metallic_zone_mask=accessories_mask → albedo, normal, roughness, metallic, ao, packed_orm, packed_arm → BD Bulk Save with labels=albedo,normal,roughness,metallic,ao,orm,arm → context auto-pick → 7 files saved in one Run ``` **Recommended character parts pipeline** (rebuild + composite via SAM3 + Qwen Image Edit): ``` [Source IMAGE] ├─→ [external VLM, e.g. ComfyUI-QwenVL] → newline-separated part list │ ├─→ BD_Lotus2ModelLoader (depth) → BD_Lotus2Predict → depth_image │ └─→ BD_SAM3MultiPrompt ↓ combined_mask, per_prompt_masks BD_PartsRefine (iou_threshold=0.7) ↓ refined_masks, refined_labels BD_PartsBuilder masks=refined_masks, labels=refined_labels, combined_mask=combined_mask, depth_image=depth ↓ parts (PARTS_BUNDLE) BD_PartsBatchEdit (Qwen) model + clip + vae from your Qwen Image Edit 2509 + Lightning subgraph inpaint_mode=flatten_redraw, alpha_after_edit=auto_from_white_bg tonemap=2.0, shift=3.0, cfg_norm=0.85 (matches manual recipe) ↓ parts (rebuilt) + image_batch BD_PartsExport composite_size=4096 (or whatever), save_psd=true, save_masks=true → per-tag PNGs + composite PNG + layered PSD with optional mask layers ``` ### Pixal3D Nodes (`🧠BrainDead/Pixal3D`) ![Pixal3D workflow](docs/images/nodes/pixal3d.png) Image-to-3D generation using Pixal3D (TencentARC/Pixal3D, Trellis2-based). Downloads ~10GB model weights on first run to `/srv/AI_Stuff/models/huggingface/`. | Node | Description | |------|-------------| | **BD Pixal3D Preprocess** | Prepare an image for Pixal3D: apply mask, crop to subject, composite on background, and estimate camera FOV via MoGe-2 (auto) or manual radians. Outputs PIXAL3D_INPUT bundle + 512×512 preview. | | **BD Pixal3D Image to 3D** | Generate 3D from a PIXAL3D_INPUT. Three-stage pipeline: sparse structure → shape latent → texture latent. Outputs untextured TRIMESH (Z-up) + TRELLIS2_VOXELGRID for full PBR bake via BD_OVoxelBake. | **Pixal3D pipeline:** ``` [Source image + mask] ↓ BD Pixal3D Preprocess (fov_mode=auto_moge) ↓ pixal3d_input BD Pixal3D Image to 3D (pipeline_type=1024_cascade) ↓ mesh ↓ voxelgrid BD CuMesh Simplify BD OVoxelBake → albedo, normal, roughness, metallic ``` ### CubePart Nodes (`🧠BrainDead/CubePart`) ![CubePart workflow](docs/images/nodes/cubepart.png) Open-vocabulary, part-controllable 3D decomposition with Roblox **CubePart**. Give it a mesh and up to **8** free-text part names; it generates one clean mesh per part, canonically aligned for rigging / game engines. The `cube_part` library is vendored under `nodes/cubepart/vendor/` (self-contained — no separate pip package). | Node | Description | |------|-------------| | **BD CubePart Segment** | TRIMESH (or a `.glb` path) + up to 8 comma/newline-separated part names → `parts` (TRIMESH_LIST, one mesh per part in name order), `combined` (single color-coded TRIMESH preview, a deterministic concat of `parts`), and `part_names` (STRING). Names past 8 are dropped (logged). | | **BD CubePart Get Part** | `parts` (TRIMESH_LIST) + `index` → a single TRIMESH (+ `name`) so one part flows into CuMesh Simplify / Blender / export. Index is clamped to range. | **Models** — paths auto-resolve via ComfyUI `folder_paths` (respecting `extra_model_paths.yaml`) and **auto-download from HuggingFace on first run** if missing (`auto_download` toggle, default on). Leave `model_dir`/`text_encoder_path` empty to auto-resolve, or set an explicit override. - `Roblox/cubepart` → `cubepart` model folder (`extra_model_paths.yaml: cubepart:` → `/srv/AI_Stuff/models/cubepart/`; else `models/cubepart`). `multi_part_dit.safetensors` ~8.6GB + `vae.safetensors` ~1.3GB. - `Qwen/Qwen3-VL-4B-Instruct` → under the `LLM` model folder (`/srv/AI_Stuff/models/LLM/Qwen3-VL-4B-Instruct/`), text encoder loaded offline. To pre-download instead of relying on first-run auto-download: ```bash huggingface-cli download Roblox/cubepart --local-dir /srv/AI_Stuff/models/cubepart huggingface-cli download Qwen/Qwen3-VL-4B-Instruct --local-dir /srv/AI_Stuff/models/LLM/Qwen3-VL-4B-Instruct ``` > **License:** CubePart *code* is MIT, but the model weights / parent repo are under the **CUBE3D RESEARCH-ONLY RAIL-MS** license — fine for internal/research use; review before shipping outputs commercially. **CubePart pipeline:** ``` [mesh from Pixal3D / Trellis2 / .glb] ↓ mesh BD CubePart Segment (parts="body, left wheel, right wheel, ...") ↓ parts (TRIMESH_LIST) ↓ combined (colored preview) BD CubePart Get Part (index=N) ↓ mesh BD CuMesh Simplify → BD Blender Decimate → export # Preview all segments at once: parts ─→ BD Mesh Preview (Thumbnails) # labeled contact-sheet IMAGE of every part parts ─→ BD Preview 3D # interactive three.js viewer, color-coded ``` ### Depth Nodes (`🧠BrainDead/Depth`) ![Lotus-2 depth workflow](docs/images/nodes/lotus2.png) SOTA monocular geometry prediction. | Node | Description | |------|-------------| | **BD Lotus-2 Model Loader** | Load FLUX.1-dev base + Lotus-2 depth or normal LoRA + LCM bridge. Module-level cache so reuse across Predict calls is instant. Optional CPU offload toggle for stacking with other models. | | **BD Lotus-2 Predict** | Run a loaded Lotus-2 model on an image. Outputs map (IMAGE), raw_linear ([0,1] normalized), and colorized_preview. Diffusion-based, much higher quality than feedforward depth estimators (DepthAnything/MiDaS). | --- ## Usage Examples ### Cache Expensive Generation ``` [Expensive Node] → [BD Cache Image] → [Next Node] seed: 42 ``` First run generates and caches. Subsequent runs skip generation. ### Character Sheet Workflow ``` [BD Prompt Iterator (Advanced)] prompts: "front view\nleft profile\nright profile\nback view" suffixes: "_front\n_left\n_right\n_back" base_filename: "character" ↓ [BD Qwen Character Edit] ↓ [BD Save File] name_prefix: "MyCharacter" ``` ### Save Context Workflow (PBR Maps Example) Replace concat-string chains with one context + N save nodes: ``` BD Save Context template: "%project%/%character%/%character%_%name%_v%version%%suffix%" character: "letti" name: "topwear" version: "03" project: "biloxi" ↓ context_id (auto-picks if only one registered) ↓ BD Derive PBR Maps BD Bulk Save ├─ albedo ────→ input_1 labels = "albedo ├─ normal ────→ input_2 normal ├─ roughness ─→ input_3 roughness ├─ metallic ──→ input_4 metallic └─ ao ────────→ input_5 ao" custom_vars = "subfolder=PBR" ↓ Saves all 5 in ONE Run: biloxi/letti/PBR/letti_topwear_v03_albedo_001.png biloxi/letti/PBR/letti_topwear_v03_normal_001.png biloxi/letti/PBR/letti_topwear_v03_roughness_001.png ... ``` **Per-save custom_vars** layer on top of the context's vars. Empty `%var%` segments collapse cleanly (`a/%b%/c` with `b=""` → `a/c`). Undefined vars stay as `%var%` literals to surface typos. ### Multi-Reference Character Edit ``` [Load Image 1 (identity)] ─┐ [Load Image 2 (style)] ─┼→ [BD Qwen Identity Lock] → [Generate] [Load Image 3 (pose)] ─┘ image1_strength: 1.5 image1_role: character ``` --- ## Directory Structure ``` ComfyUI-BrainDead/ ├── nodes/ │ ├── mesh/ # 3D mesh processing │ ├── blender/ # Blender-based operations │ ├── cache/ # Caching and save nodes │ ├── character/ # Qwen character consistency │ ├── pixal3d/ # Pixal3D image-to-3D │ ├── cubepart/ # Roblox CubePart (vendored) │ ├── trellis2/ # TRELLIS2 shape/texture │ └── segmentation/ # SAM3, Parts, Face, Luma, PBR ├── example_workflows/ # 9 ready-to-use workflow templates ├── docs/images/ # Banner + workflow thumbnails ├── tools/ # audit_workflows.py, etc. ├── lib/ # Bundled Blender 5.0 └── requirements.txt # Auto-installed by ComfyUI Manager ``` ``` output/ ├── BrainDead_Cache/ # Cache nodes save here (clearable) │ ├── cached_image_abc123.png │ ├── cached_mesh_def456.ply │ └── MyProject/ # Subdirectories supported via name_prefix │ └── step1_xyz789.png └── saved_file.png # BD Save File saves to main output/ ``` --- ## Node Counts at a Glance ~113 nodes across 10 categories: - **Cache** — caching, save/load, save-context system (~16) - **Mesh** — 3D processing, color sampling, simplification, OVoxel PBR baking, orient, vertex-color/detail-normal baking (~29) - **Blender** — Blender-based geometry tools (~12) - **TRELLIS2** — TRELLIS2-specific shape/texture nodes (~9) - **Pixal3D** — image-to-3D generation (Pixal3D + MoGe FOV estimation) (~2) - **CubePart** — open-vocabulary part decomposition (Roblox CubePart) (~2) - **Character** — Qwen-Image character consistency (~4) - **Prompt** — prompt iteration, filename templates, ForEach iteration (~5) - **Segmentation** — SAM3 multi-prompt + Parts pipeline + FaceSocketInfill + LumaStats + NormalizeLuma + CenterMedianLuma + DepthToShadowMap + DrawRect + DrawEllipse + MaskResolver + MaskCorrelate + Human parsers (~25) - **Depth** — Lotus-2 (FLUX-based diffusion depth/normal) (~2) - **PBR / asset prep** — MaskFlatten, PackChannels, DerivePBR (in Segmentation) ## Node Categories ``` 🧠BrainDead/ ├── Cache/ # Caching and file I/O nodes ├── Mesh/ # 3D mesh processing and color tools ├── Blender/ # Blender-based mesh operations ├── TRELLIS2/ # TRELLIS2-specific caching ├── Pixal3D/ # Pixal3D image-to-3D generation ├── CubePart/ # Roblox CubePart part decomposition ├── Character/ # Qwen-Image character consistency ├── Prompt/ # Prompt iteration tools ├── Segmentation/ # SAM3 + Parts pipeline + asset prep └── Depth/ # Lotus-2 depth/normal ``` --- ## Tips ### Caching Best Practices - Use descriptive `cache_name` values: `"trellis_base_mesh"`, `"character_face_gen"` - Use `name_prefix` for project organization: `"Project1/Step1"` - Connect workflow seed to cache `seed` for automatic invalidation - Use `force_refresh` to regenerate without changing seed ### Character Consistency - Always use the highest quality reference image for `image1` - Set `image1_role` to `"character"` and highest strength - Use secondary images only for style/pose reference - The Identity Lock node is optimized for face preservation ### Prompt Iteration - Use `suffix_list` mode for clean filenames: `char_front`, `char_left` - Connect `seed` output to sampler for different seeds per prompt - Use `workflow_id` to maintain separate iteration states ---

## 🧠 BrainDeadGuild **Professional AI Tools for Creative Production** Created by **BizaNator** [BrainDeadGuild.com](https://BrainDeadGuild.com) | [BrainDead.TV](https://BrainDead.TV) | [GitHub](https://github.com/BrainDeadGuild) | [Discord](https://braindeadguild.com/discord) ### Other BrainDead ComfyUI Nodes - BD - Image Descriptor (Coming Soon) - BD - Prompt Iterator Enhanced (Coming Soon) - BD - Character Consistency Suite (Coming Soon) *Building tools for the BrainDeadGuild community* --- **A Biloxi Studios Inc. Production** © 2025 Biloxi Studios Inc. — All Rights Reserved

## Support - GitHub Issues: [Report bugs](https://github.com/BizaNator/ComfyUI-BrainDead/issues) - Discord: [BrainDeadGuild](https://BrainDeadGuild.com/discord)

BD Load Text

Inputs (1)

Outputs (2)