JoLT

Official inference code for JoLT: Joint Latent Trajectories for Context-Guided High-Resolution Tiled Generation.

Project page · arXiv paper

JoLT jointly denoises a low-resolution global trajectory and a tiled high-resolution trajectory. The global branch maintains composition, while the tiled branch adds local detail. Both branches exchange information at every denoising step.

Setup

The released configuration uses FLUX.2 Klein-9B at 4096 × 4096 resolution and is intended for an NVIDIA CUDA GPU with 80 GB of memory. Request access to FLUX.2 Klein-9B, then install the environment and authenticate with Hugging Face:

uv sync
uv run hf auth login

Inference

PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True uv run jolt \
  --prompt "A flooded hotel atrium contains a functioning print shop with reed boats, mechanical clocks, stained-glass fish, and citrus trees; boats carry paper, clocks drive the presses, and fish-shaped light moves across the leaves." \
  --seed 195 \
  --output output.png

The command uses the paper configuration: four denoising steps, a 4096 × 4096 canvas, 1024 × 1024 windows, 512-pixel strides, and α = 0.25. --alpha controls high-to-low conditioning strength, --detail-prompt controls local refinement, and --seed controls sampling.

Evaluation prompts

prompts.txt contains the 200 semantically dense prompts used in the paper, one prompt per line. Evaluation used each prompt's zero-based line index as its seed.

Citation

@article{koroglu2026jolt,
  title={JoLT: Joint Latent Trajectories for Context-Guided High-Resolution Tiled Generation},
  author={Koroglu, Mathis and Jeanneret, Guillaume and Caselles-Dupré, Hugo and Cord, Matthieu and Dapogny, Arnaud},
  journal={arXiv preprint arXiv:2608.15395},
  year={2026}
}

License

Apache 2.0.