RTI

RTI — Elastic Token Compression for Pixel-Space Diffusion Transformers

Adapter checkpoints for the Region Token Interface (RTI), which lets a frozen MiniT2I backbone run its middle blocks on R ≪ N content-shaped region tokens.

• Paper: arXiv:2608.29281
• Code: https://github.com/eduardzamfir/RTI
• Project page: https://eduardzamfir.github.io/rti

One checkpoint serves every budget. The region count is drawn at random during fine-tuning, so the same weights run at R = 64 … 1024; pass --budget at inference to move along the quality/compute curve.

Files

These are adapters only — LoRA on the frozen trunk plus the small interface. The backbone is downloaded from the MiniT2I release at load time and is not redistributed here.

Usage

from rti import build_rti
from rti.minit2i_text import get_prompt_encoder
from rti.sampling import sample

model = build_rti("MiniT2I-B/16", ckpt="rti_b16_40k.pt", device="cuda").eval()
enc = get_prompt_encoder(device="cuda")
ctx, mask = enc.encode(["a red panda wearing a scarf, studio photo"])
imgs = sample(model, ctx, mask, budget=256, steps=100, cfg_scale=5.0)

The span and LoRA rank are read from the checkpoint metadata.

Results

GenEval / PartiPrompts (CLIP, PickScore, ImageReward), cfg 5.0, 100 steps:

Reproduced with the public code; see results/reproduction/ in the repository.

Citation

@article{zamfir2026rti,
  title   = {Elastic Token Compression for Pixel-Space Diffusion Transformers},
  author  = {Zamfir, Eduard and Reisswig, Christian and Wu, Zongwei and
             Xian, Yongqin and Timofte, Radu},
  journal = {arXiv preprint arXiv:2608.29281},
  year    = {2026}
}