Idol: a single image to 3D Gaussians. Inference SDK and demo.
Python
import numpy as np
import torch
from tqdm import tqdm
@torch.no_grad()
def sample_flow(model, noise, cond, neg_cond, steps: int = 50, rescale_t: float = 3.0, cfg_strength: float = 3.0,
verbose: bool = True, batch_cfg: bool = False):
"""Integrate from noise (t = 1) to data (t = 0) in ``steps`` Euler steps.
rescale_t shifts the time grid towards t = 1 (more steps at high noise).
cfg_strength: guided velocity = (1 + w) * v(cond) - w * v(neg_cond).
batch_cfg: run cond and neg_cond as one batch-2 forward instead of two passes (same formula, faster on GPU)."""
def velocity(sample, t, condition):
return model(sample, torch.tensor([1000 * t] * sample.shape[0], device=sample.device, dtype=torch.float32),
condition)
sample = noise
t_seq = np.linspace(1, 0, steps + 1)
t_seq = rescale_t * t_seq / (1 + (rescale_t - 1) * t_seq)
for i in tqdm(range(steps), desc='Sampling', disable=not verbose):
t, t_prev = t_seq[i], t_seq[i + 1]
if cfg_strength > 0 and batch_cfg:
n = sample.shape[0]
both = velocity(torch.cat([sample, sample]), t, torch.cat([cond, neg_cond]))
v = (1 + cfg_strength) * both[:n] - cfg_strength * both[n:]
else:
v = velocity(sample, t, cond)
if cfg_strength > 0:
v = (1 + cfg_strength) * v - cfg_strength * velocity(sample, t, neg_cond)
sample = sample - v * (t - t_prev)
return sample