DIAMOND
A reinforcement learning method based on diffusion-model world models that generates the next-frame observation through a diffusion or denoising process; its visual fidelity exceeds that of RSSM-style methods, but its computational overhead is higher.
Timeline (last 90 days)
Simulusion has integrated the RSSM family (Dreamer, PlaNet) and the token Transformer family (IRIS), and diffusion models (DIAMOND) are also within its support scope.
DIAMOND uses a diffusion or denoising process to generate the next frame observation, and its latent state is closer to the intermediate representations along the denoising trajectory.
A single-step prediction by a diffusion world model requires multiple denoising iterations, so the computational overhead during planning is significantly higher than that of RSSM.
All Facts (3)
Simulusion has integrated the RSSM family (Dreamer, PlaNet) and the token Transformer family (IRIS), and diffusion models (DIAMOND) are also within its support scope.
50%UnverifiedDIAMOND uses a diffusion or denoising process to generate the next frame observation, and its latent state is closer to the intermediate representations along the denoising trajectory.
50%UnverifiedA single-step prediction by a diffusion world model requires multiple denoising iterations, so the computational overhead during planning is significantly higher than that of RSSM.
50%