TRIM: Trajectory Data for Latent Selector Training

Denoising trajectory latents and quality scores used to train the latent selector of TRIM (NeurIPS 2025). Code: github.com/zeyuanyin/TRIM.

Synthesized with DiffSplat (SD3.5-Medium) on 100 ChatGPT-4o prompts x 64 random seeds, 20 denoising steps per trajectory.

Files

• latent_data_4o_data100_seed64_step{k}in20.pt for k = 0...19: one tensor per denoising step, each of size (100, 64, 4, 16, 32, 32), covering 100 prompts x 64 seed trajectories, each entry being the Gaussian latent at that step (4 views x 16 channels x 32x32).
• score_results_4o/sd3_num_inference_steps_20_seed_{s}/: CLIP similarity and ImageReward scores for seed s, computed on the final render of each trajectory, one score per prompt. The same scores apply to every step, since they measure the quality of the whole trajectory.

Usage

import torch

latents = torch.load("latent_data_4o_data100_seed64_step9in20.pt")  # step 9 of 20, ~50% progress
print(latents.shape)  # torch.Size([100, 64, 4, 16, 32, 32])

Citation

@inproceedings{
    Yin2025TRIM,
    title={{TRIM}: Scalable 3D Gaussian Diffusion Inference with Temporal and Spatial Trimming},
    author={Yin, Zeyuan and Liu, Xiaoming},
    booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS)},
    year={2025}
}