Best-Quality Configuration Recommendations for Trellis.2 Inference

I am currently tuning inference parameters for Trellis.2 with the goal of achieving the highest possible output quality, prioritizing geometric fidelity and texture detail over speed or memory efficiency.

Below is the configuration I am using as a baseline:

{
  "resolution": "1536",
  "decimation_target": 800000,
  "texture_size": 4096,

  "ss_guidance_strength": 8.0,
  "ss_guidance_rescale": 0.7,
  "ss_sampling_steps": 50,
  "ss_rescale_t": 6.0,

  "shape_slat_guidance_strength": 8.5,
  "shape_slat_guidance_rescale": 0.5,
  "shape_slat_sampling_steps": 50,
  "shape_slat_rescale_t": 6.0,

  "tex_slat_guidance_strength": 2.5,
  "tex_slat_guidance_rescale": 0.2,
  "tex_slat_sampling_steps": 50,
  "tex_slat_rescale_t": 6.0
}

Questions:

• Are these values close to the recommended best-quality configuration for Trellis.2?

• Which parameters most strongly affect:

  • fine geometric detail?
  • structural stability?
  • texture sharpness and consistency?

• Are there known upper bounds or diminishing returns for:

  • guidance strength
  • sampling steps
  • resolution or texture size?

• Are different presets recommended for organic vs. hard-surface objects?

Any insights, ablation results, or officially recommended presets for quality-first inference would be greatly appreciated.

No files, tests, or entry points are named. Start by examining the Trellis.2 inference configuration in the issue and gather official presets, ablation results, and guidance on quality limits for the listed parameters and object types. Done means documenting evidence-based quality-first recommendations and any known diminishing returns.