A curated collection of 195+ state-of-the-art open source AI capabilities, packaged in self-contained Docker environments. Find a recipe, test it, copy what you need.

CVlization stands on the shoulders of giants - we package and test amazing open source projects so you can use them with confidence.

git clone --depth 1 https://github.com/kungfuai/CVlization
cd CVlization

# Install CLI (optional - or just use bash scripts)
pip install .

# Optional: install with remote execution support
pip install .[remote]    # SSH runner
pip install .[aws]       # SageMaker runner
pip install .[deploy]    # Serverless deployment (Cerebrium)

# Browse examples
cvl list                  # compact grouped overview
cvl list -k gpt           # search by keyword
cvl list --format list    # detailed view
# or browse examples/ on GitHub

# Run any example
cvl run nanogpt build
cvl run nanogpt train

# Copy into your project (bundled with cvlization)
cvl export perception/image_classification/torch -o your-project/

That's it! Each example is self-contained with its own Dockerfile and dependencies. (We battled CUDA versions and dependency conflicts so you don't have to.)

• Examples
• Running an Example
• Benefits
• Requirements
• For Contributors
• License

Our examples/ directory is organized by capability (what the model does) rather than modality (what data it processes).

examples/
  analytical/          # Prediction & forecasting (time series, tabular ML)
  perception/          # Understand signals (vision, speech, multimodal)
  generative/          # Create content (text, images, video, audio, avatars)
  physical/            # Robotics & embodied AI (vision-language-action models)
  agentic/             # AI agents (RAG, tool use, optimization, workflows)

Legend: ✅ = Tested and maintained | 🧪 = Experimental

Capability	Example Directory	Implementations	Status
Image Classification	examples/perception/image_classification	torch, cifar10-speedrun	✅
Object Detection	examples/perception/object_detection	mmdet, torchvision, rt-detr, yolov13	✅
Segmentation	examples/perception/segmentation	instance (mmdet, torchvision), semantic (mmseg, torchvision), panoptic (detectron2, mmdet, torchvision), sam, sam_lora_finetuning, sam3, sam3_finetuning, flowrvs	✅
Pose Estimation	examples/perception/pose_estimation	dwpose, mmpose	✅
Tracking	examples/perception/tracking	global_tracking_transformer, soccer_visual_tracking	✅
Line Detection	examples/perception/line_detection	torch	✅
Document AI	examples/perception/doc_ai	OCR (chandra_ocr, deepseek_ocr, docling, doctr, dots_ocr, nanonets_ocr, olmocr_2, paddleocr_vl, surya), VLMs (donut_doc_classification, donut_doc_parse, granite_docling, granite_docling_finetune), Layout (doclayout_yolo), Parsing (dolphin_v2, churro_3b, extract0, nvidia_nemotron_parse), Leaderboard (leaderboard)	✅
Vision-Language Models	examples/perception/vision_language	florence_2 (+ finetune), gemma3_vision (+ grpo, sft), internvl3, joycaption_llava, kosmos2_grounded_ocr, lighton_ocr, llama3_vision, llava_next_video, minicpm_v_2_6, molmoe_1b, moondream2 (+ finetune), moondream3, owl_vit, paligemma2 (detection, segmentation), phi_3_5_vision_instruct, phi_4_multimodal_instruct, pixtral_12b, qwen3_vl	✅
3D Reconstruction	examples/perception/3d_reconstruction	dust3r, mast3r, monst3r, hunyuanworld_mirror, map_anything, nerf_tf (experimental)	✅

Capability	Example Directory	Implementations	Status
LLMs (text generation)	examples/generative/llm	Pretraining (nanogpt, modded_nanogpt, nanomamba), Full pipeline (nanochat: pretrain, sft, rl), Fine-tuning (peft_mistral7b_sft, trl_sft, miles_grpo, unsloth: gpt_oss_grpo, gpt_oss_sft, llama_3b_sft, qwen_7b_sft, gemma3_4b_sft), Inference (mixtral8x7b, sglang, vllm, dllm, nanbeige4_3b_thinking, nomos_1, rnj_1_instruct), Diffusion LLM (semicat), Interpretability (gemma_scope_2_270m_it)	✅
Image Generation	examples/generative/image_generation	cfm, ddpm, diffuser_unconditional, dit, dreambooth, edm2, flux, next_scene_qwen, qwen_image_layered, rae, repa, stable_diffusion, uva_energy (experimental), vqgan	✅
Video Generation	examples/generative/video_generation	animate_diff, animate_diff_cog, animate_x, cogvideox, ctrl_world, deforum, dove, flashvsr, framepack, hunyuan_video_1_5, kandinsky_5, krea_realtime_scope (experimental), longcat_video, ltx2, mimic_motion, minisora, phantom (experimental), propainter, real_video, reward_forcing, skyreals, svd_cog, svd_comfy, turbodiffusion, vace, vace_comfy (experimental), video2x, video_enhancement, video_in_between, wan2gp, wan2gp_wan, wan_animate, wan_comfy, worldcanvas	✅
Text-to-Speech (TTS)	examples/generative/audio	cosyvoice3, vibevoice_realtime, voxcpm	✅
Avatar & Talking Head	examples/generative/video_generation/avatar	anytalker, egstalker, fastavatar, flashportrait, hunyuanvideo_avatar, imtalker, lite_avatar, live_avatar, livetalk, longcat_video_avatar, personalive, wan_s2v	✅

Note: These examples are regularly updated and tested to ensure compatibility with the latest dependencies. We recommend starting with the nanogpt example.

Browse on GitHub: perception/ • generative/ • analytical/ • physical/ • agentic/

Browse via CLI: cvl list for a compact overview, or cvl list -k <keyword> to search

You can run examples using the cvl CLI or directly with bash scripts.

Option 1: Using cvl CLI (recommended)

cvl run nanogpt build
cvl run nanogpt train

Option 2: Using bash scripts directly

cd examples/generative/llm/nanogpt
bash build.sh
bash train.sh

More examples:

# RL post-training with GRPO (cutting-edge reasoning model training)
cvl run miles_grpo train

# Generate video from text prompt (Tencent HunyuanVideo)
cvl run hunyuan-video-1-5 predict --prompt "A cat playing piano"

# Document AI extraction (IBM Granite-Docling)
cvl run granite-docling predict -i input_pdf=@document.pdf

# RAG agent with LangGraph
cvl run agentic-rag-langgraph-helpdesk predict --question "How do I list examples?"

For detailed instructions and available options, see the README.md in each example directory.

License Note: Each example may reference projects with different licenses. Check the license file in each example directory.

Run examples on cloud infrastructure instead of locally:

# AWS SageMaker (managed training)
cvl run nanogpt train --runner sagemaker --spot --output-path s3://bucket/outputs

# SkyPilot (any cloud: AWS, GCP, Azure, Lambda Labs)
cvl run nanogpt train --runner skypilot --gpu A100:1 --cloud aws

# SSH (existing GPU server)
cvl run nanogpt train --runner ssh user@gpu-host

Manage remote jobs:

cvl jobs list --runner sagemaker
cvl jobs logs <job-id>
cvl jobs kill <job-id>

See cvl/runners/README.md for setup instructions.

Deploy inference endpoints to serverless GPU platforms:

# Deploy to Cerebrium
cvl deploy ltx2 --gpu L40

# Manage deployed services
cvl services list
cvl services status ltx2
cvl services logs ltx2
cvl services delete ltx2

See cvl/deployers/README.md for setup and supported platforms.

Centralized Caching: All examples use ~/.cache/cvlization/ for models and datasets, avoiding re-downloads across examples:

• Automatic caching for HuggingFace Hub, PyTorch, and custom downloads
• Managed by build scripts - no manual setup required
• Shared across all examples to save disk space and bandwidth

Self-Contained Docker Environments: Each example is isolated with pinned dependencies:

• CUDA and dependency conflicts already resolved - saves hours or days of setup
• Source code is mounted at runtime (not baked into images) for easy iteration
• Dependency versions are pinned where possible for reproducibility

Production-Ready Patterns: Copy what works into your projects:

• Consistent build/train/predict script structure across all examples
• Battle-tested configurations for 190+ AI capabilities
• Examples regularly updated and tested for compatibility

• Docker (Install Docker)
• NVIDIA GPU (most examples require 16GB+ VRAM; A10 or better recommended)
• nvidia-docker for GPU access

  # Ubuntu
  sudo apt-get install -y nvidia-container-toolkit

No installation needed - run examples directly in Google Colab: Colab notebook: CIFAR-10 classification

CVlization includes Claude Code skills for AI-assisted development and automated verification:

• verify-training-pipeline - Validates training examples are properly structured, build successfully, train without errors, and log appropriate metrics
• verify-inference-example - Validates inference examples build correctly and run inference successfully

These skills enable Claude to automatically verify examples end-to-end, helping maintain code quality across the repository.

Contributing Guidelines:

• See CONTRIBUTING.md for standardization patterns and best practices
• All examples should follow the build/train/predict script pattern
• Use verification metadata in example.yaml to track testing status

• examples/: Dockerized AI examples (perception, generative, analytical, physical, agentic)
• cvl/: CLI tool source code
• cvlization/: Optional reusable library components
• doc/: Project documentation, including doc/runners/ for cloud setup guides
• tests/: Unit and integration tests

The cvlization/ directory provides optional reusable components:

Available:

• Training pipeline abstractions (image classification, object detection, LLMs, diffusion)
• Dataset builders with caching (PyTorch, TensorFlow, HuggingFace)
• Model factories (Torchvision, MMDetection, MMSegmentation)
• Utilities (metrics, logging, download helpers)

Installation: pip install -e .

Note: Examples are self-contained and don't require the library. For production use, copying example code directly is often simpler than depending on the library package.

Detailed documentation can be found in the doc/ directory:

• Computer vision model training workflow and quality checks
• Multi-task multi-input models: a common pattern
• Reusable model components

CVlization Library & CLI: MIT License

• The cvlization package and cvl CLI tool are released under the MIT License
• Safe for commercial use

Examples Directory: Mixed Licenses

• Examples may reference projects with various licenses (copyleft, non-commercial, etc.)
• Examples are NOT included when you pip install cvlization
• Always check the license file in each example directory before using in production

Note: Each example packages different open-source projects with their own licenses. Review licenses carefully for your use case.