image2garment/GarmentParticles-Realistic

Garment Particles (Realistic Image Fine-Tuned Checkpoints)

Official fine-tuned checkpoints for Garment Particles, adapted for realistic and rendered image conditioning.

• Paper: Garment Particles: A 2D--3D Symmetric Garment Representation for Generation and Editing
• Project Page: https://garment-particles.github.io
• Code: https://github.com/garment-particles/GarmentParticles
• Base Models: georgeNakayama/GarmentParticles

---

Overview

This repository hosts the realistic image-conditioned Stage 1 (PGF) fine-tuned checkpoints, the shared Stage 2 Edge Model, and 200 evaluation test images.

The Stage 1 model was fine-tuned for 35,000 steps across 16 × NVIDIA H100 GPUs using PyTorch FSDP2 on 46,119 realistic garment outfits generated with GPT Image 2 from GarmentCodeData-v2.

Validation Loss Trajectory

Checkpoint	Steps	Validation Loss	Notes
Step 5,000	5k	0.8781	Initial adaptation
Step 10,000	10k	0.7472	Rapid feature alignment
Step 15,000	15k	0.6867	Stable generation
Step 20,000	20k	0.6466	High geometric detail
Step 25,000	25k	0.6255	Fine-grain pattern refinement
Step 30,000	30k	0.6150	Near-convergence milestone
Step 35,000	35k	0.6131	Final fully converged checkpoint (-30.2% loss reduction)

---

Repository Structure

Path	Description	Contents
pgf_image_realistic_step35000/	Recommended: Fully converged Stage 1 PGF flow matching model	15 GB (16 shards)
pgf_image_realistic_step30000/	Step 30,000 Stage 1 checkpoint	15 GB (16 shards)
edge/	Pretrained Stage 2 Edge & Sewing Pattern Reconstruction Model	8.8 GB (8 shards)
test_images/	200 evaluation garment test images (eval_set_40_prompt5)	200 PNG images (206 MB)

---

Quickstart & Inference

Clone the GarmentParticles repository:

git clone https://github.com/garment-particles/GarmentParticles.git
cd GarmentParticles

# Install dependencies
pip install torch torchvision
pip install -r src/requirements.txt
pip install flash-attn --no-build-isolation
export PYTHONPATH=$PWD/src:$PYTHONPATH

Download the checkpoints and test images:

hf download image2garment/GarmentParticles-Realistic --local-dir checkpoints_hub/realistic

Run two-stage image-conditioned inference:

torchrun --standalone --nproc_per_node=1 inference/infer_twostage.py \
  eval.sample_per_batch=1 eval.n_samples=0 eval.evaluate=False \
  train.exp_name=realistic_img_samples sample.num_sampling_steps=100 \
  gpf_ckpt=null \
  dataset.front_only=True dataset.use_all_captions=True \
  dataset.img_drop_prob=0 dataset.text_drop_prob=1 \
  model.use_qknorm=True model.use_rope=False model.in_channels=6 model.freeze_everything=False \
  edge_model.use_qknorm=True \
  edge_model_ckpt=checkpoints_hub/realistic/edge \
  model=sparse_lightningdit_v3_xl1_w_img_text_v2 \
  pgf_weight_init=checkpoints_hub/realistic/pgf_image_realistic_step35000 \
  --config-name sparselightningdit_xl_garment_particle_inference

---

Architecture Specifications

• Stage 1 (PGF): SparseLightningDiTV3CrossAttnVarlenImgTextV2 (28 DiT layers, 1152 hidden dimension, 16 attention heads, QK-Norm, DINOv2-Large image encoder + CLIP text conditioning, velocity flow matching).
• Stage 2 (Edge Model): LightningCrossAttnDiTV3EdgeModelVarlen (24 DiT layers, 1024 hidden dimension, 16 attention heads, RoPE, RMSNorm, SwiGLU, FlashAttention-2, 37 panels, 38 curves).

---

Citation

@inproceedings{garmentparticles2026,
  title={Garment Particles: A 2D--3D Symmetric Garment Representation for Generation and Editing},
  author={George Nakayama and others},
  booktitle={SIGGRAPH Conference Papers},
  year={2026}
}