Bi-Anchor Interpolation Solver for Accelerating
Generative Modeling

 

Hongxu Chen¹   ·  Hongxiang Li¹   ·  Wang Zhen¹   ·  Long Chen^1*

¹Hong Kong University of Science and Technology   ^*Corresponding Author  

Summary: We presented the Bi-Anchor Interpolation Solver (BA-solver), a novel framework that effectively bridges the gap between computationally expensive training-free solvers and resource-intensive training-based acceleration methods. By introducing a lightweight SideNet, we endow frozen flow matching backbones with bidirectional temporal perception, enabling high-order efficient numerical integration. BA-solver achieves state-of-the-art generation quality with as few as 5 to 10 NFEs, matching the performance of standard Euler solvers requiring 100+ NFEs.

1. Environment setup

conda create -n BA_solver python=3.11
conda activate BA_solver
pip install -r requirements.txt

2. Dataset

Dataset download

The experiments for ImageNet are provided. You can place the data that you want and can specifiy it via --data-dir arguments (e.g., "../train") in training scripts. We didn't apply preprocessing process for extracting latent in our experiments.

3. Training

export WANDB_API_KEY="YOUR_API_KEY"
export WANDB_ENTITY="WANDB_ENTITY"
export WANDB_PROJECT="WANDB_PROJECT"

accelerate launch --num_processes 1 train_BA_side.py \
  --data-dir <DATA_PATH> \
  --pretrained-ckpt <PRETRAINED_MODEL_PATH> \
  --resolution 256 \
  --model SiT-XL/2 \
  --exp-name <EXP_NAME> \
  --batch-size 168 \
  --learning-rate 0.0001 \
  --SideNet-depth 4 \
  --cfg-scale 4.0 \
  --sampling-steps 1000 \
  --checkpointing-steps 1000

All experiments are conducted using the SiT-XL/2 backbone. You can adjust the configuration using the following arguments:

• --SideNet-depth: Depth of the SideNet module.
• --SideNet-in-channels: Number of input channels for SideNet.
• --SideNet-base-channels: Base channel width for SideNet.
• --SideNet-h-emb-dim: Embedding dimension for the offset parameter $h$.
• --data-dir: Path to the ImageNet dataset directory.
• --pretrained-ckpt: Path to the pre-trained SiT-REPA checkpoint. You can download it here.
• --resolution: Input image resolution (e.g., 256 or 512).

For ImageNet 512x512, please use the following script:

export WANDB_API_KEY="YOUR_API_KEY"
export WANDB_ENTITY="WANDB_ENTITY"
export WANDB_PROJECT="WANDB_PROJECT"

accelerate launch --num_processes 1 train_BA_side.py \
  --data-dir <DATA_PATH> \
  --pretrained-ckpt <PRETRAINED_MODEL_PATH> \
  --resolution 512 \
  --model SiT-XL/2 \
  --exp-name <EXP_NAME> \
  --batch-size 96 \
  --learning-rate 0.0001 \
  --SideNet-depth 8 \
  --cfg-scale 4.0 \
  --sampling-steps 1000 \
  --checkpointing-steps 1000

4. Evaluation

Utilizing trained SideNet, you can generate ImageNet-256 images (and the .npz file can be used for ADM evaluation suite) through the following script:

torchrun --nnodes=1 --nproc_per_node=8 generate.py \
    --model SiT-XL/2 \
    --resolution 256 \
    --num-fid-samples 50000 \
    --per-proc-batch-size 32 \
    --base-ckpt <PRETRAINED_MODEL_PATH> \
    --side-ckpt <SIDENET_MODEL_PATH> \
    --SideNet-depth 4 \
    --num-steps <STEPS> \
    --sample-dir <OUTPUT_DIR> \
    --cfg-scale <CFG> \
    --cfg-interval-start <CFG_START> \

You can also generate ImageNet-512 images through the following script:

torchrun --nnodes=1 --nproc_per_node=8 generate.py \
    --model SiT-XL/2 \
    --resolution 512 \
    --num-fid-samples 50000 \
    --per-proc-batch-size 10 \
    --base-ckpt <PRETRAINED_MODEL_PATH> \
    --side-ckpt <SIDENET_MODEL_PATH> \
    --SideNet-depth 8 \
    --num-steps <STEPS> \
    --sample-dir <OUTPUT_DIR> \
    --cfg-scale <CFG> \
    --cfg-interval-start <CFG_START> \

We provided SideNet checkpoints here for ImageNet-256 and -512 generation.

5. Results

Here are some visual samples on ImageNet-512 with only 7 NFEs.

Acknowledgement

This code is mainly built upon DiT, SiT, edm2, RCG, and REPA repositories.

BibTeX

@article{chen2026bi,
  title={Bi-Anchor Interpolation Solver for Accelerating Generative Modeling},
  author={Chen, Hongxu and Li, Hongxiang and Wang, Zhen and Chen, Long},
  journal={arXiv preprint arXiv:2601.21542},
  year={2026}
}