Case Study / AI Pipeline

Blender → SDXL Stylization Pipeline

Single-command script that renders Blender passes, generates depth maps, and produces stylized SDXL output via ControlNet. Multiple style presets — no re-rendering needed. Designed for artists who don't want to touch code.

Rolle
Pipeline Architect
Dauer
Ongoing (internal tool)
Tools
Blender 5.1, Python 3.14, ComfyUI, SDXL, ControlNet
Problemstellung

Why Build This?

The gap: 3D artists and product designers need stylized renders for different contexts — studio product shot, material study, spatial concept, artistic variant. Usually, this means re-rendering or manual post-processing for every variant.

The insight: Blender already has the 3D data. SDXL can generate stunning stylized outputs. The missing piece: a reliable bridge that requires zero manual steps.

The result: One command. One Blender file. Unlimited style variants.

Vorgehen

The Pipeline Architecture

01

Render Passes

Blender Python API renders alpha, object-ID, normals, and depth in a single script run. Scene validation + aspect ratio handling (square, 16:9, portrait, scene-native).

02

Depth Conditioning

16-bit depth map → ComfyUI ControlNet Depth (fp16, strength 0.85). SDXL Base 1.0 as the generative backbone. Dual CLIP encoders for full prompt understanding. dpmpp_2m+karras scheduler, CFG 7.0, 25 steps.

03

Style Presets

Multiple prompt presets: studio still life, material study, spatial concept, floating onyx, material alchemy. Switch presets without touching Blender geometry or re-rendering.

04

Post-Processing

RealESRGAN 4× upscale → lanczos resize to 2048² → WebP q=90 output. All in one automated chain. SSD-1B variant for rapid prototyping (8 steps).

Ergebnis

What It Produces

From a single Blender scene, the pipeline generates variants that range from photorealistic product shots to abstract material studies. The same geometry, different visual languages.

Open Source Reference

Customizable for specific style profiles or client workflows. Single Python entry point (depth_pipeline.py) handles ComfyUI API calls, upload, queue, poll, and download. A second pipeline (canny_icon_pipeline.py) covers CSS → Canny → SDXL icon generation via Chrome Headless.

5
Style Presets
2
Model Modes (SDXL + SSD-1B)
4
Render Pass Types
Erkenntnisse

What This Enables

  • Artists iterate on style, not on technical setup. Change a prompt, not a render setting.
  • Depth maps are the universal bridge between 3D and generative AI. Once this works, every Blender scene becomes an AI input.
  • The pipeline design pattern (extract → transform → generate → post-process) applies to any creative tool chain.
  • Production quality (SDXL, 25 steps, 2048²) and prototype speed (SSD-1B, 8 steps) can coexist in one pipeline.