The diffusion sampling environment itself is general enough. Defining a new target distribution would be enough for most cases. A few interesting variants of the environment are: - Conditional sampling (e.g., [VAE](https://github.com/GFNOrg/gfn-diffusion/tree/main/vae)) - relatively easy to implement - Fine-tuning (e.g., [RTB](https://github.com/GFNOrg/diffusion-finetuning), [outsourced diffusion](https://arxiv.org/abs/2502.06999), and [outsourced Schrodinger bridge](https://arxiv.org/abs/2509.26364)) - requiring much more work _Originally posted by @hyeok9855 in https://github.com/GFNOrg/torchgfn/pull/419#discussion_r2454597468_