Blender

Higgsfield for Blender is an AI add-on that brings Higgsfield's generation tools into Blender, generating and importing 3D scenes, meshes, rigs, textures, images, and video. It adds a floating viewport bar with tabs for scene building, 3D models, character animation, images, video, cameras, and assets; results can be inserted into the open scene as editable geometry, planes, materials, rigs, keyframes, or video outputs. Scene Builder creates editable objects, layouts, and lighting from a prompt. The character-animation workflow produces a fitted weighted rig with timeline keyframes. An optional Higgsfield Bridge MCP endpoint at https://bridge.higgsfield.ai/mcp connects an external AI agent such as Claude to the add-on for operations including building blockouts, generating meshes at the 3D cursor, and requesting Seedance renders in the open Blender scene. Generation runs on Higgsfield's servers, requires an internet connection and a signed-in Higgsfield account, uses the account's existing Higgsfield credits, and supports Blender 4.2 through 5.1.

Spirula Studio is a cross-vendor 3D Gaussian Splatting trainer that processes photos or video into Gaussian splats and textured meshes in a self-contained binary. It combines video frame extraction, AI masking, structure-from-motion, Gaussian Splatting training, meshing, depth and normal generation, and skybox processing, with native support for fisheye and 360-degree equirectangular datasets. Its Vulkan backend runs on NVIDIA, AMD, Intel, and Apple GPUs, while a CUDA backend supports CUDA-capable NVIDIA GPUs; the application provides GUI and CLI workflows for Windows, Linux, and macOS. Quantized training is designed to support up to 10 million SH3 Gaussians in 8 GB of VRAM, and masking requires a separately downloaded SAM checkpoint.

ABot-Recon is a streaming 3D-reconstruction model and inference tool for reconstructing long video streams from video frames alone. At each step, it caches key-value features from the preceding 11 frames, predicts a point map in the current camera coordinate system, estimates the adjacent relative camera pose, and composes those poses sequentially into a global trajectory and point cloud without persistent learned long-range memory. It can output camera poses, relative poses, local and world point maps, colors, confidence maps, and run metadata. An optional loop-closure backend retrieves candidate revisited-frame pairs using DINOv2-SALAD descriptors, predicts relative-pose constraints with ABot-Recon, and refines the trajectory through sparse pose-graph optimization. The repository provides Python and command-line interfaces, a public model checkpoint, a demo, evaluation code, and visualization utilities. The released configuration targets Linux, Python 3.10 or later, PyTorch 2.5.1, and CUDA 12.1; source code is licensed under Apache License 2.0, while the model weights have separate terms in MODEL_LICENSE.md.

OpenMontage is an open-source, agent-driven video production system that turns plain-language instructions into video projects through an AI coding assistant. Its workflow covers research, scripting, production planning, asset generation, editing, and final composition, with production pipelines, approval gates, cost estimates, and post-render checks. The system can produce image-based videos as well as videos assembled from real motion footage: its agent can build a corpus from free stock footage and open archives, retrieve clips, edit them into a timeline, and render the result. It runs locally with FFmpeg and Remotion and includes 12 production pipelines, more than 100 tools, and hundreds of agent-skill and production-knowledge files.

Soup is an open-source Python CLI for fine-tuning and post-training large language models from a YAML configuration and a single training command. It supports supervised fine-tuning, LoRA, QLoRA, preference optimization, quantization, automatic batch-size and GPU detection, and local training without a cloud service. Its optional layer-streaming mode keeps the frozen base model out of VRAM and feeds it to the GPU one decoder layer at a time, reducing memory use for low-VRAM systems. The project describes training an 8B model on a 4 GB laptop GPU with this mode; layer streaming is marked beta.