Camera Overhaul, Local Edit Mode & Blender 5.1 Support

[sakalond]

StableGen v0.2.0: Camera Overhaul, Local Edit Mode & Blender 5.1 Support

This is the biggest StableGen update yet — a ground-up rework of the camera system, a brand new Local Edit mode, Blender 5.1 support with GPU-accelerated projection, Apple Silicon support, and a suite of quality-of-life improvements across the board.

⚠️ Blender Compatibility: StableGen v0.2.0 supports Blender 4.2 – 4.5 (using OSL shaders) and Blender 5.1+ (using native Raycast nodes with GPU acceleration). Blender 5.0 is not supported — OSL is broken in 5.0 and the native Raycast node was not introduced until 5.1. Blender 5.1 is currently in beta but is expected to release soon.

✨ What's New in v0.2.0:

• 📷 Camera System Overhaul:

  • The camera placement system has been completely rewritten with 7 placement strategies including fully automatic modes:
    • Orbit Ring — the original circular arrangement.
    • Fan Arc — cameras spread in an arc facing the subject.
    • Hemisphere — even distribution across a hemisphere.
    • PCA-Axis — cameras aligned to the mesh's principal axes.
    • Normal-Weighted K-means — clusters camera directions based on surface normals, biasing toward faces that need coverage.
    • Greedy Occlusion — iteratively picks directions that maximise visible, uncovered surface.
    • Interactive Visibility — real-time scroll-to-adjust preview that lets you balance occlusion filtering live, with HUD and camera count preview.
  • Per-Camera Optimal Aspect Ratios: Each camera now gets its own resolution computed from the mesh's silhouette extent in that viewing direction. No more wasted pixels on letterboxing — portraits get tall frames, landscapes get wide ones.
  • No More 8-Camera Limit: The hardcoded limit has been removed — use as many cameras as you need. (Thanks to hickVieira for the suggestion!)
  • Camera Generation Order: New reorder list in Viewpoint Blending Settings lets you control the exact order cameras are processed in Sequential mode. Includes 6 preset strategies: Alphabetical, Front→Back→Sides, Back→Front→Sides, Alternating Opposite, Top→Bottom, and Reverse.
  • New camera operators:
    • Clone Camera — duplicate a camera and immediately enter fly mode to reposition.
    • Mirror Camera — mirror a camera across X/Y/Z axis through the mesh center.
    • Toggle Camera Labels — show floating per-camera prompt text in the viewport.

  Here are some of the automatic camera adding. As you can see, the process is now much simpler:

  • restaurant model by Jellepostma from Sketchfab
  • woman model by Patrix from Sketchfab

• 🖥️ Blender 5.1 Support (GPU-Accelerated Projection):

  • StableGen now fully supports Blender 5.1, including all API changes to the compositor, UV selection, animation channels, and Eevee engine naming.
  • On Blender 5.1+, projection uses native Raycast shader nodes instead of OSL scripts. This means projection and visibility rendering now run on GPU (CUDA/OptiX/Metal) instead of being locked to CPU — a major speed boost on complex scenes.
  • Blender 4.2 – 4.5 remain fully supported using the original OSL pipeline.

• 🎯 Local Edit Mode:

  • A brand new generation mode that replaces the old "Preserve Original Texture" toggle. Point cameras at specific areas you want to modify — the new generation blends seamlessly over the original using angle-based and vignette-based feathering, leaving untouched areas pristine. (Based on the refine/mirror PR by ManglerFTW — thank you!)
  • Works with all architectures (SDXL, Flux, and Qwen Image Edit) — Qwen gets its own dedicated local_edit generation method.
  • Silhouette Edge Feathering: New controls for softly blending projection boundaries based on screen-space frustum distance — prevents hard seams at projection edges. Powered by a new dedicated feather.osl shader (and its native Blender 5.1 equivalent).
  • Separate Angle & Vignette Controls: Angle ramp (black/white point) and vignette (width, softness) are now independently tunable for precise control over where new texture fades into old.
  • Original Render IPAdapter mode, which enables SDXL / FLUX to match the styling of the previous generation for usecases such as improving detail / quality in specific areas.

• 🎨 Qwen Refine Mode:

  • New Refine and Local Edit generation methods for Qwen Image Edit — restyle or locally modify existing textures using the image editing workflow.
  • Supports optional previous-view reference and depth map as additional context images.

• 📦 New Presets:

  • LOCAL REFINE — SDXL local edit mode with original_render IPAdapter, depth ControlNet, and Lightning 8-step LoRA. Preconfigured with angle-based blending, vignette feathering (width 0.3), and edge feather projection (30 px). Great for touching up detail on an existing texture.
  • LOCAL EDIT (QWEN) — Qwen local edit mode with Lightning LoRA. Vignette feathering (width 0.1) and edge feather projection (15 px) enabled; angle blending disabled for a softer blend. Ideal for rewriting text, adding details, or changing colours in a targeted area.
  • REFINE (QWEN) — Qwen refine mode for global restyling. Applies changes uniformly across all camera views — change the overall colour scheme, art style, or surface appearance in one pass.

Examples of Local Refine (before/after):

Before	After

Examples of Refine (Qwen):

Before	"change the color scheme to red and black"	"remove all text"

• 🟣 Qwen Voronoi Projection Mode:

  • New option under Qwen Image Edit guidance (sequential generate mode). Instead of zeroing weights of non-generated cameras, keeps natural angle-based weights and projects the guidance fallback colour from cameras not yet generated.
  • With a high weight exponent, each surface point is dominated by its closest camera — creating Voronoi-like segmentation where the discard-over-angle setting becomes irrelevant.
  • Dedicated Voronoi presets: New QWEN EDIT VORONOI and QWEN EDIT VORONOI (NUNCHAKU) presets — preconfigured with Voronoi mode, exponent 1000, and post-generation exponent reset to 15.

• 🍎 Apple Silicon Support:

  • StableGen now runs natively on Apple silicon Macs. The extension manifest includes macos-arm64 platform wheels for Pillow, OpenCV, and imageio_ffmpeg.

• 🔬 Weight Normalization for High Exponents:

  • Projection weights are now max-relative normalized before applying the weight exponent. This prevents numerical underflow at any exponent value — you can now push the Weight Exponent up to 1000 without black edge artifacts.
  • Mathematically: pow(cos/max_cos, exp) — the blending ratios are identical to the original formula, but computed entirely in [0, 1] range.
  • This enables Voronoi-like hard segmentation at high exponents without any artifacts.

• 🔄 Reset Exponent After Generation:

  • New toggle to automatically reset the Weight Exponent to a different value after generation completes — useful when generating with a high exponent (e.g., 1000 for Voronoi segmentation) but wanting a lower exponent (e.g., 15) for the final blended result.

• 🔍 Debug & Diagnostic Tools:

  • New debug toolbox (gated behind addon preferences toggle) with operators that visualize the projection pipeline without running AI:
    • Project Solid Colors — one colour per camera to verify coverage and boundaries.
    • Project Grid Patterns — UV alignment checkerboard per camera.
    • Visualize Weight Material — see the raw blended weights as rendered.
  • All operators call the real project_image() and export_visibility() functions, so debug output faithfully matches what generation produces.

• 📁 Portable .blend Files:

  • OSL scripts are now embedded as internal text datablocks in the .blend file. Saved files no longer require the addon folder to be present for the projection shaders to work.

• 🛠️ Improvements:

  • Advanced Resolution Rescaling: Configurable target megapixels (default 1.0 MP) and Qwen-specific alignment rounding (multiples of 112 for the Qwen2.5-VL vision encoder window).
  • More Qwen Guidance Map Options: Added Workbench Render and Viewport Render as guidance map types, beyond depth and normal maps.
  • Color Matching for View Blending: New color matching module with multiple algorithms (MKL, Reinhard, Histogram, MVGD, hybrid). Match each generated view's colors to the current texture before blending.
  • Mirror Axis Selection moved from panel to operator dialog for cleaner UX.
  • Refine Output Crop & Alpha Vignette — generated images are post-processed to remove hard borders and softly vignette edges.

• 🐛 Bug Fixes:

  • Active UV map is now properly restored after projection completes.
  • Prevented image datablock collisions — get_or_load_image() now verifies file paths, not just names, avoiding incorrect texture assignments when multiple images share a filename.
  • Improved error handling in generation callbacks — exceptions in sequential/refine callbacks are now caught and properly reported instead of causing silent failures.
  • Fixed the issue caused by update to ComfyUI Qwen related nodes causing errors and failed generations.

---

Contributors: Thanks to ManglerFTW (PR #63 — refine/mirror support that evolved into Local Edit mode) and hickVieira (suggestion to remove the 8-camera limit).

Full Changelog: v0.1.1...v0.2.0

---

This discussion was created from the release Camera Overhaul, Local Edit Mode & Blender 5.1 Support.