Dear authors, thank you for providing another possibility for image generation with the Tarflow model.
Based on this, we have accelerated Tarflow's sampling module by applying GS-Jacobi Iteration to accelerate the non-linear RNN in Tarflow's sampling process, achieving speed-ups of 4.53x in Img128cond, 5.32x in AFHQ, 2.96x in Img64uncond, and 2.51x in Img64cond without degrading FID scores or sample quality.
We have also open-sourced checkpoints that reproduce the training results from the Tarflow paper, including Img128cond, AFHQ, Img64uncond, and Img64cond. Our paper is available at https://arxiv.org/abs/2505.12849.
Finally, we would like to thank the author team for their assistance during the reproduction process.
Dear authors, thank you for providing another possibility for image generation with the Tarflow model.
Based on this, we have accelerated Tarflow's sampling module by applying GS-Jacobi Iteration to accelerate the non-linear RNN in Tarflow's sampling process, achieving speed-ups of 4.53x in Img128cond, 5.32x in AFHQ, 2.96x in Img64uncond, and 2.51x in Img64cond without degrading FID scores or sample quality.
We have also open-sourced checkpoints that reproduce the training results from the Tarflow paper, including Img128cond, AFHQ, Img64uncond, and Img64cond. Our paper is available at https://arxiv.org/abs/2505.12849.
Finally, we would like to thank the author team for their assistance during the reproduction process.