3.38.0

Prompt-to-3D: generate or edit the source image with text

The image-to-3D pipeline no longer requires importing an image:

• Generate from a prompt — type a description in AI: Image → 3D (or qtmesh generate3d --prompt "a goblin warrior" -o out.glb, or MCP generate_mesh_from_image {prompt}) and the app creates the source image with FLUX.2-klein-4B (Apache-2.0, ~5.2 GB, downloadable in AI Model Settings; any Stable Diffusion checkpoint works as a fallback), then reconstructs it in 3D as usual.
• Edit with a prompt — with an image loaded, the prompt edits it (FLUX.2 reference conditioning): "give him golden armor" transforms the image while preserving identity, pose and composition. A 🗑 button clears the image to start fresh. Works in the GUI, CLI (image.png --prompt "...") and MCP (image_path + prompt).
• Progress bar during generation, click-to-enlarge preview, and text inputs now regain focus with a single click after switching apps.

Windows: TRELLIS.2 runtime now bundled

The bundled trellis-cli.exe ships in the Windows package (static MinGW build, portable CPU baseline) — Windows users no longer see "TRELLIS.2 runtime not installed": download the models in AI Model Settings and generate. (CPU backend for now; GPU via Vulkan is a tracked follow-up.)

Full Changelog: 3.37.8...3.38.0