Amagine3D is the open-source 3D capability layer from Amagine: describe a hardware product, add reference images and key dimensions, and an agent writes editable build123d CAD, checks assembly interference and motion in a real geometry runtime, then exports STEP, STL, or color-aware 3MF. We review the 3D-native agent loop, the web refs control, and where it stands vs text-to-3D mesh generators.
camera-to-blender is an MIT-licensed open-source pipeline (created 2026-09-03, 440+ stars and 47 forks in three days) that turns a single phone photo of a real object into a 3D model sitting inside your Blender scene in under a minute. Point your phone at an object, tap the shutter, and a relay server removes the background (optionally with Gemini), generates a mesh via the Tripo3D API, and pushes it straight into a connected Blender session over WebSocket — no saving files, no manual import. This review covers the four-part architecture (Python relay server, phone camera web app, Blender add-on, ngrok tunnel), the exact setup steps, the 30–60 second generation flow, the troubleshooting guide, and how it differs from prompt-based 3D tools like Meshy and from multi-shot photogrammetry: it is a capture pipeline for real objects aimed at Blender users, with honest limits around model fidelity and the paid Tripo3D API dependency.
dream-loop is an MIT-licensed Agent Skill (created 2026-09-07, 130+ stars in its first day) by Anshu Chimala that turns a single prompt into a game, app or 3D scene with genuinely impressive visuals — by closing a loop most agents never close. Step 1: the agent 'dreams' a high-quality target screenshot with an image-generation model, styled as an in-engine screenshot of the ideal result. Step 2: it builds toward that target with real assets — Blender modeling preferred for 3D, image-gen textures, normal maps and skyboxes. Step 3: a separate subagent 'judge' with a clean context compares a live screenshot of the build against the concept and scores it on a gated five-tier ladder (shape 0–3, light and color 3–5, materials and surfaces 5–7, fine detail 7–9, indistinguishable 9–10), returning blocking directives with concrete magnitudes. Step 4: the builder loops until the judge scores 8+, or recognizes a stall and makes one big structural change instead of tweaking. This review covers the full loop mechanics, the judge prompt and its anti-nagging rules, the exit criteria, how dream-loop upgrades an existing product by re-rendering a live screenshot, and the honest prerequisites: a strong multimodal agent with image generation, vision and subagents — currently tested only with GPT-6 Astra in Codex.