Training Free Guided Flow-Matching with Optimal Control
Authors: Luran Wang, Chaoran Cheng, Yizhen Liao, Yanru Qu, Ge Liu
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We show the effectiveness of our method with extensive empirical experiments, including text-guided image manipulation, controllable generation of small molecules on QM9, and energy optimization of flow-based all-atom peptide design. |
| Researcher Affiliation | Academia | 1University of Cambridge 2University of Illinois Urbana-Champaign 3Tsinghua University EMAIL, EMAIL EMAIL, EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 OC-Flow on Euclidean Space 1: Given: Pre-trained model: f p, initial state: x0 2: Initialize: Control terms θ0, learning rate η, weight decay β 3: for k = 0 to Max Iterations do 4: Solve for the state trajectory: Xθk t+ t = Xθk t + f p(t, Xθk t ) + θk t 5: Update control: θk+1 t = βθk t + η xtΦ(Xk 1 ) 6: end for |
| Open Source Code | No | The paper does not explicitly state that the source code for OC-Flow is provided, nor does it include a link to a code repository. The text mentions "due to the lack of code and implementation details of DFlow, we are unable to fully reproduce their implementation." but this refers to a different method (DFlow), not OC-Flow. |
| Open Datasets | Yes | We first validate our OC-Flow on the traditional text-to-image generation task... on the celebrity face dataset Celeb A-HQ (Karras, 2017)... controllable molecule generation on the QM9 dataset (Ruddigkeit et al., 2012; Ramakrishnan et al., 2014)... peptide backbone design using a test set derived from (Li et al., 2024) |
| Dataset Splits | No | For quantitative comparison, we used the Celeb A dataset, randomly sampling 1,000 images, which were manipulated based on text guidance... trained for each property on half of the QM9 data as the classifier... The test dataset split was also based on the one defined in the Pep Flow framework... we randomly selected 10 complexes from the dataset. After tuning, the full set of 162 complexes was used for guided sampling and evaluation. |
| Hardware Specification | Yes | In Table 8 we show memory usage and runtime on high-dimensional images, evaluated on a single A100 with 40G memory. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies such as libraries, frameworks, or programming languages used in the implementation. |
| Experiment Setup | Yes | For ours, we set time step of 100, step size η = 2.5, weight decay of 0.995, the weight of the extra constraint of 0.4, and the number of optimization steps of 15... For all properties, MAE was used as the optimization target and γ is the regularization coefficient such that γ R 1 0 θt 2 dt is the additional OC loss. For all optimization methods, we always used a fixed number of 50 Euler steps... Specifically, 200 time steps are used for ODE simulation, with 10 control terms, each controlling 20 time steps. As shown in Table 7, this approach reduces memory and runtime complexity without compromising the accuracy of the ODE simulation. Furthermore, for consistency and comparability across experiments, we strictly controlled the initial noise during reruns to ensure consistency and comparability across experiments. We used the pre-trained model as the initialization for our experiments, allowing us to build upon the pre-trained weights and achieve consistent performance improvements through hyperparameter adjustments. In OC-Flow(rot), we used α = 0.95 and β = 0.8; in OC-Flow(trans), α = 0.9 and β = 1.2; and in OC-Flow(both), α = 0.95 and β = 2.0. |