Flow Matching

1.0.10 verified Sat May 09 auth: no python

Flow Matching for Generative Modeling, a library by Meta Research for training continuous normalizing flows via flow matching objectives. Current version 1.0.10, supports Python >=3.9, actively maintained.

pip install flow-matching

Common errors

error AttributeError: module 'flow_matching' has no attribute 'FlowMatching' ↓

cause Importing the high-level wrapper incorrectly after v1.0.0, or using outdated documentation.

fix

Use from flow_matching import CondOTFlowMatching (or another specific flow class) instead of from flow_matching import FlowMatching.

error TypeError: __init__() got an unexpected keyword argument 'dim' ↓

cause Passing the old `dim` parameter to a class that expects different arguments (e.g., `CondOTFlowMatching` does not take `dim` directly; it takes `input_dim` in some versions, or defaults to 2).

fix

Check the constructor signature: for CondOTFlowMatching(input_dim=2), or simply omit dim and set it later.

error RuntimeError: Expected all tensors to be on the same device, but found at least two devices ↓

cause Model parameters and input tensors are on different devices (CPU vs GPU).

fix

Move model and tensors to same device: model = CondOTFlowMatching(dim=2).to(device); x0 = x0.to(device); x1 = x1.to(device).

Warnings

breaking As of v1.0.0, the API was completely rewritten. Imports like `from flow_matching import FlowMatching` now return a high-level wrapper, not the old `CNF` class. If upgrading from <1.0, review all imports. ↓

fix Update imports: use `from flow_matching import CondOTFlowMatching` instead of old `from flow_matching import FlowMatching`.

deprecated The `models.CNF` class is deprecated since v1.0.0. Use `CondOTFlowMatching` or other specific flow matching classes. ↓

fix Replace `from flow_matching.models import CNF` with `from flow_matching import CondOTFlowMatching`.

gotcha The library expects input tensors (x0, x1) to have dtype `torch.float32`. Using float64 may cause silent errors or performance degradation. ↓

fix Ensure tensors are float32: `x0 = x0.float()`.

Imports

FlowMatching
```
from flow_matching import FlowMatching
```

CNF

wrong

from flow_matching.models import CNF

correct

from flow_matching import CNF

Common mistake: trying to import from submodule `models`

CondOTFlowMatching

from flow_matching import CondOTFlowMatching

FMObjective

from flow_matching.objectives import FMObjective

Quickstart

Minimal example: create a flow matching model, compute the flow matching objective between random source and target samples.

import torch
from flow_matching import CondOTFlowMatching
from flow_matching.objectives import FMObjective

dim = 2
flow = CondOTFlowMatching(dim=dim)
objective = FMObjective(flow)

x0 = torch.randn(100, dim)
x1 = torch.randn(100, dim)
t = torch.rand(100)
loss = objective(x0, x1, t)
print(f"Loss: {loss.item()}")