Skip to content

Latest commit

 

History

History
121 lines (100 loc) · 6.5 KB

COOKBOOK.md

File metadata and controls

121 lines (100 loc) · 6.5 KB

Cosmos Cookbook

Start from what you want to achieve. Every recipe here leverages Cosmos to help you reach an objective.

Jump to: Inference · Simulation · Post-training · Distillation · Quantization · Evaluation

Legend: industry and GPU-budget codes
Code Industry Code GPU budget
ANY any industry S 1 GPU, up to 96 GB
AV autonomous vehicles M 1 GPU over 96 GB, or 2–8 GPUs
ROB robotics L 9–64 GPUs
INFRA smart infrastructure XL more than 64 GPUs
MFG manufacturing and industrial sites
AG agriculture
MED healthcare

Inference

Recipe Industry GPU budget
Generation
Text-to-image, text-to-video, image-to-video, video-to-video, with optional audio
ANY S M
Visual Reasoning
Captioning, VQA, grounding, temporal localization, physical plausibility
ANY INFRA S M
Action Policy
Forward dynamics, inverse dynamics, policy
AV ROB S
Transfer
Spatial controls: edge, blur, depth, segmentation, world-scenario map, multi-control
ANY AV S M
NIM deployment
Choose a run time (generation/reasoning), run accelerated inference from one container · support matrix
ANY S M

Simulation New

Recipe Industry GPU budget
Multi-view multi-sensor driving SDG
Fine-tune Cosmos on driving data for multi-view + LiDAR joint generation · Try example checkpoint
AV L
Autoregressive driving simulation
Build autoregressive multi-view driving simulator · Try example checkpoint
AV M
Autoregressive bimanual simulation
Build interactive neural simulator for robot policy· Try example checkpoint
ROB M
Autoregressive sim-to-real transfer
Build autoregressive sim-to-real transfer model · Try example checkpoint
ROB M
Surgical simulation
Coming soon
MED

Post-training

Recipe Industry GPU budget
Generation
Post-train with your captioned video: full SFT or LoRA
ANY M
Reasoning
Alignment SFT and physical-plausibility SFT
ANY M
Action policy
Post-train a policy, then evaluate it closed-loop: DROID, LIBERO, RoboCasa
ROB L XL
Forward dynamics
Post-train a forward-dynamics model on DROID
ROB M
Worker safety
Coming soon · Flag hazards in industrial video
MFG S
Action Policy (Light)
Coming soon · Post-train a LIBERO-10 manipulation policy on one GPU, evaluate it closed-loop
ROB M

Distillation

Recipe Industry GPU budget
4-step students
4-step distillation guide, text-to-image, image-to-video
ANY L

Quantization

Recipe Industry GPU budget
Quantization
Static FP8 export and serving: Nano/Super video, Super base and 4-step text-to-image/image-to-video
ANY S M

Evaluation

Benchmarks Industry GPU budget
Generation
PAI-Bench-G video · PAI-Bench-C controlled video · PhysicsIQ physics · UniGenBench images · RBench robotics
ANY ROB M
Reasoning
TAR & VANTAGE with VLMEvalKit, on any OpenAI-compatible endpoint
INFRA S