File size: 1,116 Bytes
c5e4c7b | 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 | """
IRIS: Iterative Refinement Image Synthesizer
=============================================
A mobile-first image generation architecture that combines:
- PDE-SSM spatial mixing (O(N log N), native 2D, no scanning)
- Weight-shared iterative refinement (6 blocks × R iterations)
- Structured latent canvas (DC-AE with coarse-to-fine channels)
- Tiny PixelShuffle decoder (0.1M params)
- Multi-Query Attention + 2D RoPE
- Flow matching training with logit-normal timestep sampling
Usage:
from iris.model import IRIS
from iris.configs import get_model_config
from iris.flow_matching import flow_matching_loss, euler_sample
model = IRIS(**get_model_config("iris-small"))
# Training
losses = flow_matching_loss(model, latents, text_embeddings)
# Sampling
images = euler_sample(model, noise, text_embeddings, num_steps=20)
"""
from .model import IRIS
from .configs import get_model_config, CONFIGS
from .flow_matching import flow_matching_loss, euler_sample
__version__ = "0.1.0"
__all__ = ["IRIS", "get_model_config", "CONFIGS", "flow_matching_loss", "euler_sample"]
|