---

PhotoQuilt is a training-free framework for generating photomosaics at arbitrary resolution. A low-resolution bootstrap pass fixes the global layout; the latent canvas is upscaled and re-noised; denoising then proceeds in fixed tiles so each region becomes its own convincing image while the overall scene stays coherent.

Installation

git clone https://github.com/KooroshRH/PhotoQuilt.git
cd PhotoQuilt

# Install PyTorch for your CUDA version first, then:
pip install -r requirements.txt

Model weights are downloaded from Hugging Face on first run.

Repository layout

PhotoQuilt/
├── flux1/     # FLUX.1-Krea-dev
├── flux2/     # FLUX.2-klein-9B (distilled)
├── sd21/      # Stable Diffusion 2.1
└── outputs/   # default output directory (created per run)

Run each backend from inside its folder:

cd flux1 && python run_flux1.py ...
cd flux2 && python run_flux2.py ...
cd sd21  && python run_sd21_mosaic.py ...

Method (four phases)

All backends follow the same high-level procedure:

1. Bootstrap — generate at --bootstrap-size or load --base-image.
2. Latent upscale — bicubic upscale to --height × --width.
3. Noise injection — inject noise at step steps − noise-steps.
4. Tiled denoise — complete the tail in fixed --tile-size windows.

In the paper, renoising strength (s) is noise-steps / steps (e.g. --steps 28 --noise-steps 17 → (s \approx 0.61), the FLUX.1 default in the paper).

Checkpoints

We evaluate PhotoQuilt with these Hugging Face checkpoints (used in the paper):

Backend	Checkpoint
FLUX.1	black-forest-labs/FLUX.1-Krea-dev
FLUX.2	black-forest-labs/FLUX.2-klein-9B (distilled)
SD 2.1	Manojb/stable-diffusion-2-1-base

Gallery mode on FLUX.1 additionally uses black-forest-labs/FLUX.1-Redux-dev.

FLUX.1 (flux1/)

Checkpoint: black-forest-labs/FLUX.1-Krea-dev (default --checkpoint Krea-dev).

Quick start

cd flux1
python run_flux1.py \
  --height 4096 --width 4096 \
  --steps 28 --noise-steps 17 \
  --tile-size 1024 \
  --prompt "a serene mountain lake at sunrise" \
  --tile-prompt "a bird"

Bootstrap from an image

cd flux1
python run_flux1.py \
  --base-image path/to/reference.jpg \
  --height 8192 --width 8192 \
  --steps 28 --noise-steps 17 \
  --tile-size 512 \
  --tile-prompt "a bird"

Gallery mosaic (per-tile reference images via FLUX.1 Redux)

cd flux1
python run_flux1.py \
  --base-image path/to/group_photo.png \
  --tile-image-dir path/to/gallery/ \
  --tile-prompt "a vivid photo collage tile" \
  --height 8192 --width 8192 \
  --steps 28 --noise-steps 17 \
  --tile-size 1024 \
  --seed 42

Multi-GPU (tile-parallel denoising + decode)

cd flux1
CUDA_VISIBLE_DEVICES=0,1 python run_flux1.py \
  --height 8192 --width 8192 \
  --steps 28 --noise-steps 17 \
  --tile-size 256 \
  --multi-gpu \
  --base-image path/to/reference.jpg \
  --tile-prompt "a bird"

Requires a tile-aligned canvas (height and width are exact multiples of --tile-size) and at least as many tile-rows as visible GPUs.

---

FLUX.2 Klein (flux2/)

Checkpoint: black-forest-labs/FLUX.2-klein-9B (distilled; pass --checkpoint klein). Default schedule: --steps 8 --noise-steps 5 ((s = 5/8)). Per-tile reference images use FLUX.2-native in-context conditioning (no Redux).

Quick start

cd flux2
python run_flux2.py \
  --checkpoint klein \
  --height 4096 --width 4096 \
  --steps 8 --noise-steps 5 \
  --tile-size 1024 \
  --prompt "a serene mountain lake at sunrise" \
  --tile-prompt "a bird"

Gallery mosaic

cd flux2
python run_flux2.py \
  --checkpoint klein \
  --base-image path/to/layout.png \
  --tile-image-dir path/to/gallery/ \
  --tile-prompt "a vivid photo collage tile" \
  --height 6144 --width 6144 \
  --steps 8 --noise-steps 5 \
  --tile-size 1024

---

Stable Diffusion 2.1 (sd21/)

Checkpoint: Manojb/stable-diffusion-2-1-base (default --model).

cd sd21
python run_sd21_mosaic.py \
  --height 2048 --width 2048 \
  --steps 50 --noise-steps 35 \
  --tile-size 512 \
  --prompt "a city skyline at dusk" \
  --tile-prompt "a bird"

---

Flags

Core options (all backends):

Flag	Description
--height / --width	Output resolution (pixels).
--bootstrap-size	Bootstrap long side (1024 FLUX, 768 SD 2.1).
--tile-size	Tile / attention window side.
--steps / --noise-steps	Denoising schedule; paper strength (s =) noise-steps / steps.
--prompt / --tile-prompt	Global layout prompt and per-tile prompt.
--base-image	Bootstrap from an image instead of generating.
--seed	Random seed.

Model-specific:

• FLUX.1 / FLUX.2: --tile-image-dir for gallery mosaics (FLUX.1 uses Redux; FLUX.2 uses native reference conditioning).
• FLUX.1: --multi-gpu for tile-parallel generation on multiple GPUs.
• FLUX.2: --checkpoint klein for the distilled model.
• SD 2.1: --model to pick the Hugging Face checkpoint.

Run python run_<backend>.py --help for all options (including advanced tuning flags).

Outputs

Images are saved under --output-dir (default outputs/). FLUX backends use descriptive timestamped filenames; SD 2.1 writes to a run subfolder with mosaic.png.

Requirements

• Python 3.10+
• CUDA GPU recommended (CPU offload is enabled automatically on FLUX backends when VRAM is below ~60 GB)
• See requirements.txt for package versions

Acknowledgments

This implementation adapts the inference layout and scripting conventions from SEGA (Rajabi et al., 2026). PhotoQuilt builds on diffusers, Hugging Face model hubs, and the FLUX / Stable Diffusion ecosystems.

Citation

If you find PhotoQuilt useful, please cite:

@misc{roohi2026photoquilt,
  title={PhotoQuilt: Training-Free Arbitrary-Resolution Photomosaics via Bootstrapped Tiled Denoising},
  author={Koorosh Roohi and Javad Rajabi and Andrew Fleet and Babak Taati},
  year={2026},
  eprint={2606.30968},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2606.30968},
}