Generating an identical sequence of random tensorflow operations for parallel instances

I'm trying to write my own data generator for an image segmentation task in order to have more control over the transformations that I can apply. To start, I'm writing a toy generator that outputs a randomly cropped image using tf.image.random_crop:

import tensorflow.keras.preprocessing.image as k_img
import os
import numpy as np
import matplotlib
import matplotlib.pyplot as plt

class DataGen:
    seed=1
    def __init__(self,seed=1):
        self.seed = seed

    def flow(self):
        #tf.random.set_seed(self.seed)
        img = np.asarray(k_img.load_img('data/train/images/all/image.png',color_mode='grayscale'))
        while(True):
            img_crop = tf.image.random_crop(img,(32,32),seed=self.seed)
                        yield img_crop

The use case looks like this:

gen2 = DataGen(1234).flow()
plt.imshow(next(gen))
plt.imshow(next(gen))
plt.imshow(next(gen2))
plt.imshow(next(gen2))

The desired behavior is for image 1 & 3 to be the same and for image 2 & 4 to be the same. However, setting the seeds within tf.image.random_crop in this context does not do this--nothing is repeated. Setting tf.random.set_seed (commented out) does enforce a repeatable sequence, but it is global rather than local.

How do I achieve the desired behavior?