Spaces:
Running on Zero
Running on Zero
File size: 1,668 Bytes
6eff13b ccf7a99 6eff13b ccf7a99 6eff13b ccf7a99 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 | ---
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](https://arxiv.org/abs/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](https://github.com/eduardzamfir/RTI) (MIT license). Example
prompts are the ones the authors showcase on the
[project page](https://eduardzamfir.github.io/rti).
|