File size: 1,395 Bytes
93801e5 | 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 34 35 36 37 38 39 40 41 42 43 | """
ArtiGen Sampling — Flow Matching ODE Solver.
Rectified flow allows efficient 1-4 step generation via Euler solver.
"""
import torch
import torch.nn.functional as F
import numpy as np
try:
from artigen.model import ArtiGen
except ImportError:
from model import ArtiGen
def euler_solver(model, z_noise, text_embed, num_steps=4, device='cpu', cfg_scale=1.0):
model.eval()
z = z_noise.clone()
dt = -1.0 / num_steps
with torch.no_grad():
for i in range(num_steps):
t = torch.ones(z.shape[0], device=device) * (1.0 + i * dt)
if cfg_scale > 1.0:
v_cond, _ = model(z, t, text_embed, return_asdl=False)
v_uncond, _ = model(z, t, torch.zeros_like(text_embed), return_asdl=False)
v = v_uncond + cfg_scale * (v_cond - v_uncond)
else:
v, _ = model(z, t, text_embed, return_asdl=False)
z = z + dt * v
return z
def sample(model, text_embed, latent_shape=(4, 32, 32), num_steps=4, device='cpu', cfg_scale=1.5):
C, H, W = latent_shape
B = text_embed.shape[0]
z_noise = torch.randn(B, C, H, W, device=device)
z0 = euler_solver(model, z_noise, text_embed, num_steps=num_steps, device=device, cfg_scale=cfg_scale)
return z0
def decode_with_vae(vae, z, output_type='pil'):
img = torch.randn(z.shape[0], 3, 256, 256)
return img
|