LingBot-Map Training Checkpoints: Comparing lingbot-map, lingbot-map-long, and lingbot-map-stage1

The Robbyant/lingbot-map repository provides three distinct training checkpoints—lingbot-map (balanced default), lingbot-map-long (long-sequence optimized), and lingbot-map-stage1 (VGGT backbone stage-1)—each engineered for specific inference scenarios ranging from general pose estimation to extended video streams and bidirectional camera-to-world regression.

Selecting the correct LingBot-Map training checkpoint is critical for optimal performance in your specific deployment environment. Each variant targets different sequence lengths, memory constraints, and architectural backends as defined in the repository's model download specifications.

Overview of the Three LingBot-Map Checkpoints

According to the Model Download table in README.md (lines 137-140), the repository maintains three distinct weight files optimized for different computational requirements.

lingbot-map (Balanced Default)

The lingbot-map checkpoint serves as the general-purpose, balanced option suitable for most research and production use cases. It delivers optimal performance across both short and long video sequences and represents the exact checkpoint utilized in the paper, benchmark suite, and offline demo. As configured in benchmark/configs/methods/lingbot_map.yaml, this is the default weight file expected by evaluation scripts.

lingbot-map-long (Long-Sequence Optimized)

The lingbot-map-long checkpoint specifically targets very long sequences and large-scale scenes. It incorporates expanded capacity for the KV-cache and employs a longer training window, providing greater stability when streaming thousands of consecutive frames. This variant is essential for extended video processing where standard checkpoints might encounter memory pressure or temporal drift.

lingbot-map-stage1 (Stage-1 VGGT Backbone)

The lingbot-map-stage1 checkpoint captures the first stage of the two-stage training pipeline. Unlike the final checkpoints, this weight is specifically intended for bidirectional inference (camera-to-world or c2w) when loaded into the VGGT backbone. It enables coarse-to-fine pose regression within the VGGT architecture rather than the standard LingBot-Map inference path.

Technical Differences and Architecture

All three checkpoints instantiate the same core architecture defined in lingbot_map/models/gct_base.py, ensuring weight compatibility across variants. However, their training regimes and intended deployment contexts differ significantly:

• lingbot-map offers the best overall performance for mixed-length sequences, balancing accuracy and computational efficiency.
• lingbot-map-long sacrifices some generalization for extended context windows, making it robust for continuous streaming applications.
• lingbot-map-stage1 represents an intermediate training state compatible with the VGGT model, enabling specialized c2w bidirectional pose regression that the standard checkpoints do not support.

How to Load and Use Each Checkpoint

The CLI interface remains consistent across all variants. You select the desired behavior by specifying the path to the appropriate .pt file via the --model_path argument.

To use the balanced default checkpoint for standard inference:

python demo.py \
    --model_path /path/to/lingbot-map.pt \
    --image_folder example/courthouse --mask_sky

For extended video sequences requiring additional KV-cache capacity:

python demo.py \
    --model_path /path/to/lingbot-map-long.pt \
    --image_folder example/long_video \
    --keyframe_interval 2

To load the stage-1 checkpoint into the VGGT model for bidirectional c2w inference:

python demo.py \
    --model_path /path/to/lingbot-map-stage1.pt \
    --mode c2w --image_folder example/scene

Key Source Files and Configuration

Several critical source files govern how these checkpoints are referenced and loaded:

• README.md (lines 137-140): Contains the authoritative description of each checkpoint's intended use case and training characteristics.
• benchmark/configs/methods/lingbot_map.yaml: Defines the _checkpoint field that specifies which weight file the benchmark suite loads during evaluation.
• lingbot_map/models/gct_base.py: Implements the base model architecture that all three checkpoints instantiate.
• benchmark/methods/lingbot_map.py: Wraps the upstream package and consumes the checkpoint path defined in the YAML configuration, handling the actual model initialization for benchmark runs.

Summary

• lingbot-map: The balanced, default checkpoint optimized for general-purpose use across short and long sequences; used in the paper and standard benchmarks.
• lingbot-map-long: A high-capacity variant with expanded KV-cache and longer training windows for stable inference on very long video streams and large scenes.
• lingbot-map-stage1: An early-training checkpoint designed specifically for bidirectional c2w inference within the VGGT backbone architecture.
• All checkpoints utilize the same underlying architecture from gct_base.py and are interchangeable via the --model_path argument in demo.py.

Frequently Asked Questions

What is the default checkpoint for most LingBot-Map experiments?

The lingbot-map checkpoint serves as the default for most experiments. It provides balanced performance across sequence lengths and is the specific weight used in the paper, benchmark suite, and offline demo applications according to the repository documentation.

When should I use lingbot-map-long instead of the standard checkpoint?

Use lingbot-map-long when processing very long video sequences or large-scale scenes that span thousands of frames. Its expanded KV-cache capacity and optimized training window prevent instability and memory issues during extended streaming inference.

Can I use lingbot-map-stage1 with the standard inference pipeline?

No, lingbot-map-stage1 is specifically intended for the VGGT backbone and bidirectional camera-to-world (c2w) inference. It represents an intermediate training stage and requires the VGGT architecture to function correctly, rather than the standard LingBot-Map inference path defined in gct_base.py.

Where are checkpoint paths configured in the benchmark suite?

Checkpoint paths are configured in benchmark/configs/methods/lingbot_map.yaml via the _checkpoint field. The benchmark/methods/lingbot_map.py wrapper reads this configuration to load the specified weights during evaluation runs.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}