tf.image.rgb_to_hsv and tf.image.hsv_to_rgb are not numerically reversible for large float32 inputs

Issue type

Bug

Have you reproduced the bug with TensorFlow Nightly?

Yes

Source

source

TensorFlow version

tf 2.13.0

Custom code

Yes

OS platform and distribution

Ubuntu 20.04

Mobile device

No response

Python version

3.9

Bazel version

No response

GCC/compiler version

No response

CUDA/cuDNN version

No response

GPU model and memory

No response

Current behavior?

For very large float32 RGB values (on the order of 1e38), the pair of functions tf.image.rgb_to_hsv → tf.image.hsv_to_rgb fails to recover the original input values.

tf.image.rgb_to_hsv() produces finite outputs.

tf.image.hsv_to_rgb() does not return the same RGB tensor — values differ significantly even though both functions should be approximately reversible.

Standalone code to reproduce the issue

import numpy as np
import tensorflow as tf

x = np.array([[[1.0870518e+38, 2.0447458e+38, 6.2133610e+37]]],
dtype=np. float32)
tf_result1 = tf.image.rgb_to_hsv(x)
#[[[3.3333334e-01, 6.9613039e-01, 2.0447458e+38]]]
tf_result2 = tf.image.hsv_to_rgb(tf_result1)
#[[[6.2133610e+37, 2.0447458e+38,6.2133610e+37]]]

print(tf_result1)
print(tf_result2)

Relevant log output

[[[3.3333334e-01, 6.9613039e-01, 2.0447458e+38]]]
[[[6.2133610e+37, 2.0447458e+38,6.2133610e+37]]]