Open-source ArtSmoker pipeline turns text prompts into textured, Blender-ready 3D models on your own AWS
niravdd · reddit · 2026-09-03
A developer released ArtSmoker, an MIT-licensed open-source studio app that chains text→2D asset→editing→textured 3D model→engine export behind one self-hosted UI, running entirely inside your own AWS account — aimed at small teams without local GPUs.
- 2D generation: Bedrock models (SD3.5 Large, Stable Image Ultra) or one-click self-deploys of FLUX.2 [dev], HunyuanImage 3.0, and Qwen-Image onto SageMaker GPU endpoints with packaging, NF4/BF16 quantization, scale-to-zero, and job tracking. Includes prompt enhancement, multi-model comparison grids, and exact seed reproducibility.
- In-place editing: inpaint/outpaint, recolor, search-and-replace, strength-ladder img2img remix, and instruction-based editing via self-hosted Qwen-Image-Edit.
- 2D→3D: TripoSG geometry plus TRELLIS.2 texturing produce real PBR GLBs, then headless-Blender exports FBX/USDZ, LOD chains, collision meshes, and per-engine texture packing for Unreal/Unity/Godot.
- Cost & provenance: per-generation cost estimates shown upfront (warm Qwen-Image BF16 on 4×L40S ≈ $0.44), with full prompt/seed/model provenance per asset.
The author stresses it's a cloud tool, not a ComfyUI replacement, and invites discussion of deployment learnings (NF4 quantization ceilings on L40S, FlashInfer on Blackwell, SageMaker cold-start traps). Repo: github.com/niravdd/ArtSmoker
More from Infra
- On-device LLM inference on Android with QNN, ExecuTorch, LiteRT and Gemma showcased at EdgeAI event — carrycooldude · 2026-09-03
- llama.cpp adds support for NVIDIA's Nemotron-3-Puzzle-75B-A9B MoE model — pmttyji · 2026-09-03
- NVIDIA's 75B hybrid MoE Nemotron-3-Puzzle is now runnable locally in llama.cpp — jacek2023 · 2026-09-03
- 20VC: Nvidia's $96.2B quarter, near-$12.9B Hugging Face deal, Cognition at $46BN — 20VC · 2026-09-03
- One int8 op makes exact rescoring nearly free in PLAID-style late-interaction search — antoine_chaffin · 2026-09-03
- KV cache, not parameter count, may be the real bottleneck for long-context local models — jonejy · 2026-09-03