README.md

metadata

title: RTI - Elastic Token Compression
emoji: 🧩
colorFrom: pink
colorTo: purple
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
short_description: Pixel-space t2i with a region-token budget dial
python_version: '3.12'
startup_duration_timeout: 30m
license: mit
models:
  - eduardzamfir/RTI
  - MiniT2I/MiniT2I

RTI — Elastic Token Compression for Pixel-Space Diffusion Transformers

Interactive demo for RTI (Region Token Interface) — arXiv:2608.29281. A frozen MiniT2I pixel-space diffusion backbone runs its middle blocks on R ≪ N content-shaped region tokens instead of all patches. One adapter checkpoint serves every budget: the region count was drawn at random during fine-tuning, so you can set R at inference to move along the quality/compute curve.

Budget dial (R at 512×512, N = 1024 patch tokens):

R	Meaning	Speedup (paper)
1024	dense backbone	1.0×
512	matches dense quality	~2.1×
256	small quality drop	~2.6×
64	extreme compression	—

Two released arms are preloaded — RTI-B/16 and RTI-L/16 (LoRA + the small interface on the frozen MiniT2I trunk) — sharing a single FLAN-T5-Large prompt encoder. Prompts are encoded at fp32, sampling runs bf16 Euler flow-matching with CFG 5.0 over 100 steps, exactly as in the authors' reference implementation (rti.sampling.sample).

The rti/ package in this Space is the inference subset of eduardzamfir/RTI (MIT license). Example prompts are the ones the authors showcase on the project page.