INRFlow: Flow Matching for INRs in Ambient Space
Authors: Yuyang Wang, Anurag Ranjan, Joshua M. Susskind, Miguel Ángel Bautista
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical results demonstrate that INRFlow effectively handles different data modalities such as images, 3D point clouds and protein structure data, achieving strong performance in different domains and outperforming comparable approaches. [...] 4. Experiments We evaluate INRFlow on two challenging problems: image generation (FFHQ-256 (Karras et al., 2019), LSUN-Church256 (Yu et al., 2015), Image Net-128/256 (Russakovsky et al., 2015)), image-to-3D point cloud generation (Objaverse (Deitke et al., 2023)) and protein folding (Swiss Prot (Boeckmann et al., 2003)). |
| Researcher Affiliation | Industry | 1Apple, Machine Learning Research 2Work done while at Apple. Correspondence to: Yuyang Wang <EMAIL>, Miguel Angel Bautista <EMAIL>. |
| Pseudocode | No | The paper describes the model architecture in Section 3.4 and visually in Figure 2, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code for the methodology described, nor does it provide a link to a code repository. |
| Open Datasets | Yes | We evaluate INRFlow on two challenging problems: image generation (FFHQ-256 (Karras et al., 2019), LSUN-Church256 (Yu et al., 2015), Image Net-128/256 (Russakovsky et al., 2015)), image-to-3D point cloud generation (Objaverse (Deitke et al., 2023)) and protein folding (Swiss Prot (Boeckmann et al., 2003)). From an ML perspective, this problem is a conditional 3D generation problem where we are given the amino-acid sequence (e.g. a sequence of discrete symbols from a vocabulary of 20 possible amino-acids) and we need to generate a 3D coordinate for each atom in the protein, where different amino-acids can have different numbers of atoms. In our experiments we use Swiss Prot set (Boeckmann et al., 2003) taking the ground truth structures from the Alpha Fold Database (Varadi et al., 2022). [...] For completeness we also tackle unconditional 3D point cloud generation on Shape Net (Chang et al., 2015). |
| Dataset Splits | No | The paper mentions using specific datasets (e.g., FFHQ-256, Image Net, Objaverse, Swiss Prot, Shape Net) and sometimes specifies training data size or sampling rates (e.g., 'We select a random set of 10k protein structures to train INRFlow.', 'For each object in Objaverse, we sample point cloud with 16k points.'). However, it does not provide explicit training, validation, and test splits (e.g., percentages or exact counts for all splits) for all experiments, nor does it consistently reference standard splits for all benchmarks in a way that ensures reproducibility of data partitioning. |
| Hardware Specification | No | The paper discusses training cost in terms of "total training Gflops" and compares models based on "# params", "bs it." (batch size * iterations), and "NFE" (number of function evaluation). However, it does not specify any particular GPU models (e.g., NVIDIA A100), CPU models, TPUs, or detailed specifications of the computing cluster used for the experiments. |
| Software Dependencies | No | The paper specifies training configuration details such as the optimizer used ('optimizer= Adam W') and its hyperparameters ('adam_beta1=0.9 adam_beta2=0.999 adam_eps=1e-8 learning_rate=1e-4 weight_decay=0.0 gradient_clip_norm=2.0 ema_decay=0.999 mixed_precision_training=bf16'). While these are important for reproducibility, it does not list specific software libraries or frameworks with their version numbers (e.g., PyTorch 1.9, CUDA 11.1), which would be necessary for a reproducible software environment. |
| Experiment Setup | Yes | default training config: optimizer= Adam W adam_beta1=0.9 adam_beta2=0.999 adam_eps=1e-8 learning_rate=1e-4 weight_decay=0.0 gradient_clip_norm=2.0 ema_decay=0.999 mixed_precision_training=bf16. [...] On image generation, all models are trained with batch size 256, except for INRFlow-XL reported in Tab. 2 and Tab. 3, which are trained for 1.7M steps with batch size 512. [...] We train an image-to-point-cloud INRFlow model on Objaverse (Deitke et al., 2023)... We train INRFlow with batch size 384 for 500k iterations. During sampling, we use an Euler-Maruyama sampler (Ma et al., 2024) with 500 steps to generate point clouds. [...] For this task we train a XL size model for 100k iterations with batch size 256. |