Atlas Viewport 🧊

Atlas Viewport 🧊 (class AtlasBlockoutViewport) is the node you'll spend the most time staring at. It opens a real-time 3D viewport in the ComfyUI browser, pre-aligned to your recovered camera, so you can actually look through the scene before you export it. It's the pack's review room, blockout tool, camera-path author, and angle-extractor all in one node.

What it does

The pattern mirrors the Yedp Blockout approach: the browser renders, sends base64 JSON to Python, Python decodes it, and out come proxy image outputs. Concretely:

1. Connect an ATLAS_SOLVE and source_image, queue once.
2. The Three.js viewport opens, camera locked to the photo.
3. Place primitive geometry - box, plane, person card - to block out the scene.
4. Click Render Proxy Passes; it fills client_data and re-queues.
5. Four proxy/LDR outputs come alive: shaded, depth, normal, mask - wire them into ControlNet or compositing.

resolution sets the long-edge render size (128–8192, default 1024); the short side auto-follows your source image's aspect. You can also drag the node's resize handle. A shot_cam from AtlasDefineShotCam conforms it to a project format instead.

The two features that make it the hub

Camera Path mode. Author a keyframed move in the viewport. "⏺ Bake Full Path" produces path_frames (a proxy image batch - feed it to repair preview or Video Combine) plus camera_path (the raw keyframes, feed it to AtlasExportCameraPathUSD for a real animated camera). "📷 Bake Repair Frame" stores only the final preview frame for path-guided repair. Frames outside the recovered camera's cone show the documented black/undefined regions - expected, not a bug.

Extract Angle 📐. This is the multi-angle patch workflow's best friend. Orbit to the view you want a patch generated at, click 📐, and the node measures the orbit delta from the recovered camera, snaps it to the Qwen Multiple-Angles LoRA's nearest named views, and outputs four strings: patch_azimuth_view, patch_elevation_view, patch_distance (wire into AtlasAddPatchView's widgets) and patch_prompt (the ready-to-use <sks> [azimuth] [elevation] [distance] LoRA prompt). The delta uses the same pivot the backend uses, so the snapped views round-trip exactly. Until you extract an angle these outputs return an ExecutionBlocker - downstream branches stay paused rather than wasting a generation on a zero-orbit patch.

What else is under the hood

preview_expand dilates geometry for wider orbit coverage - leave it at 1.0 if you use Project mode, or the dilated fringe has no real photo data and renders black. clean_plate projects a background plate through the same registered projector so tears reveal a clean backdrop. drawn_fill_px (default 96) fills drawn shapes by smearing surrounding plate colors - deterministic edge-extend, not an inpaint; set 0 and route drawn_mask to a real inpaint if you prefer. primary_depth powers the ray-traced occlusion-culling (✂ Occlude) toggle. And if the node feels cramped, connect an AtlasViewportControls node to controls to move every button and panel off this one - the link carries no data, it just lets the two frontends find each other.

The node outputs a lot: the four proxy passes, path_frames, camera_path, the four patch strings, patch_render_mask, the pass-through solve, drawn_mask, and draw_report. It's an output node, so it renders even with nothing downstream.

Install is the pack's clone-and-go. The depth-dependent culling wants the [neural] extra, and heavy orbits want a GPU, but the core viewport runs on CPU fine.