285 lines
14 KiB
Python
285 lines
14 KiB
Python
"""CPU regression tests; no ComfyUI server, checkpoint or GPU is required.
|
|
|
|
Run with the Python shipped with ComfyUI:
|
|
python -m unittest discover -s tests -v
|
|
"""
|
|
|
|
import base64
|
|
import importlib.util
|
|
import io
|
|
from pathlib import Path
|
|
import sys
|
|
import types
|
|
import unittest
|
|
from unittest.mock import patch
|
|
|
|
import numpy as np
|
|
from PIL import Image
|
|
import torch
|
|
import torch.nn.functional as F
|
|
|
|
SPEC = importlib.util.spec_from_file_location("adachi_backend_under_test", Path(__file__).resolve().parents[1] / "backend.py")
|
|
backend = importlib.util.module_from_spec(SPEC)
|
|
SPEC.loader.exec_module(backend)
|
|
|
|
|
|
def data_url(image):
|
|
buffer = io.BytesIO()
|
|
image.save(buffer, format="PNG")
|
|
return "data:image/png;base64," + base64.b64encode(buffer.getvalue()).decode("ascii")
|
|
|
|
|
|
def source(batch=1, height=96, width=160):
|
|
pixels = torch.linspace(0, 1, batch * height * width * 3).reshape(batch, height, width, 3)
|
|
mask = torch.zeros(batch, height, width)
|
|
mask[:, 30:50, 60:80] = 1
|
|
return pixels, mask
|
|
|
|
|
|
def prepare(image, mask, **options):
|
|
return backend._prepare(image, mask, width=64, height=64, padding=4, mask_blur=0, **options)
|
|
|
|
|
|
class FakeVAE:
|
|
downscale_ratio = 8
|
|
|
|
def __init__(self, channels=4):
|
|
self.channels = channels
|
|
self.encodes = 0
|
|
|
|
def encode(self, pixels):
|
|
self.encodes += 1
|
|
latent = F.interpolate(pixels.mean(-1)[:, None], scale_factor=1 / self.downscale_ratio, mode="bilinear", align_corners=False)
|
|
return latent.repeat(1, self.channels, 1, 1)
|
|
|
|
def decode(self, latent):
|
|
value = F.interpolate(latent[:, :1], scale_factor=self.downscale_ratio, mode="nearest")
|
|
return value.repeat(1, 3, 1, 1).permute(0, 2, 3, 1).clamp(0, 1)
|
|
|
|
|
|
class BackendTests(unittest.TestCase):
|
|
def test_comparison_previews_and_result_passthrough(self):
|
|
image, _ = source()
|
|
result = image.repeat_interleave(3, 0)
|
|
calls = []
|
|
class Preview:
|
|
def save_images(self, images, filename_prefix):
|
|
calls.append((images.clone(), filename_prefix))
|
|
return {"ui": {"images": [{"filename": f"{filename_prefix}_{i}.png", "subfolder": "", "type": "temp"} for i in range(len(images))]}}
|
|
with patch.dict(sys.modules, {"nodes": types.SimpleNamespace(PreviewImage=Preview)}):
|
|
output = backend.AdachiImageCompare().compare(image, result)
|
|
self.assertTrue(torch.equal(output["result"][0], result))
|
|
self.assertEqual(len(output["ui"]["adachi_originals"]), 1)
|
|
self.assertEqual(len(output["ui"]["adachi_results"]), 3)
|
|
self.assertTrue(torch.equal(calls[0][0], image))
|
|
self.assertTrue(torch.equal(calls[1][0], result))
|
|
|
|
def test_comparison_rejects_crop_and_incompatible_batch(self):
|
|
image, _ = source(batch=2)
|
|
with patch.dict(sys.modules, {"nodes": types.SimpleNamespace()}):
|
|
with self.assertRaisesRegex(ValueError, "crop"):
|
|
backend.AdachiImageCompare().compare(image, image[:, :32])
|
|
with self.assertRaisesRegex(ValueError, "batch"):
|
|
backend.AdachiImageCompare().compare(image, image[:1])
|
|
|
|
def test_original_protected_exactly_in_both_areas(self):
|
|
image, mask = source()
|
|
for area in ("whole_image", "only_masked"):
|
|
with self.subTest(area=area):
|
|
pixels, prepared_mask, context = prepare(image, mask, area=area)
|
|
result = backend._composite(torch.ones_like(pixels), context)
|
|
self.assertEqual(result.shape, image.shape)
|
|
self.assertTrue(torch.equal(result[mask == 0], image[mask == 0]))
|
|
self.assertTrue(torch.allclose(result[mask == 1], torch.ones_like(result[mask == 1]), atol=1e-6))
|
|
self.assertEqual(tuple(prepared_mask.shape), (1, 64, 64))
|
|
|
|
def test_empty_mask_and_denoise_zero_skip_vae_and_sampling(self):
|
|
image, mask = source()
|
|
vae = FakeVAE()
|
|
sampler = types.SimpleNamespace(common_ksampler=lambda *a, **k: self.fail("Sampling should be skipped"))
|
|
with patch.dict(sys.modules, {"nodes": sampler}):
|
|
for denoise, tested_mask in ((0, mask), (0.75, torch.zeros_like(mask))):
|
|
output, _, _ = backend.AdachiWebUIInpaint().sample(None, [], [], vae, image, tested_mask,
|
|
denoise=denoise, batch_size=3, width=64, height=64, padding=4, mask_blur=0)
|
|
self.assertTrue(torch.equal(output, image.repeat_interleave(3, 0)))
|
|
self.assertEqual(vae.encodes, 0)
|
|
|
|
def test_invert_repaints_outside(self):
|
|
image, mask = source()
|
|
pixels, _, context = prepare(image, mask, mask_mode="outside", area="whole_image")
|
|
result = backend._composite(torch.zeros_like(pixels), context)
|
|
self.assertTrue(torch.equal(result[mask == 1], image[mask == 1]))
|
|
self.assertTrue(torch.equal(result[mask == 0], torch.zeros_like(result[mask == 0])))
|
|
|
|
def test_corner_masks_are_never_cropped_in_mask_only_mode(self):
|
|
image, _ = source()
|
|
for y, x in ((0, 0), (0, 159), (95, 0), (95, 159)):
|
|
for resize_mode in ("stretch", "crop", "fill"):
|
|
mask = torch.zeros(1, 96, 160)
|
|
mask[0, y, x] = 1
|
|
pixels, _, context = prepare(image, mask, resize_mode=resize_mode)
|
|
result = backend._composite(torch.ones_like(pixels), context)
|
|
self.assertTrue(torch.equal(result[0, y, x], torch.ones(3)))
|
|
self.assertTrue(torch.equal(result[mask == 0], image[mask == 0]))
|
|
|
|
def test_bbox_padding_and_target_ratio(self):
|
|
image, mask = source(height=200, width=300)
|
|
_, _, context = prepare(image, mask)
|
|
x0, y0, x1, y1 = context["records"][0]["box"]
|
|
self.assertEqual(x1 - x0, y1 - y0)
|
|
self.assertLessEqual(x0, 56)
|
|
self.assertGreaterEqual(x1, 84)
|
|
self.assertLessEqual(y0, 26)
|
|
self.assertGreaterEqual(y1, 54)
|
|
|
|
def test_batch_grouping_and_single_mask_broadcast(self):
|
|
image, mask = source(batch=2)
|
|
pixels, _, context = prepare(image, mask[:1])
|
|
variants = torch.stack([torch.full_like(pixels[0], value) for value in (0.1, 0.2, 0.3, 0.4)])
|
|
output = backend._composite(variants, context)
|
|
self.assertEqual(output.shape[0], 4)
|
|
for index in range(4):
|
|
self.assertTrue(torch.equal(output[index][mask[0] == 0], image[index // 2][mask[0] == 0]))
|
|
self.assertTrue(torch.allclose(output[index, 40, 70], torch.full((3,), (index + 1) / 10), atol=1e-6))
|
|
|
|
def test_mask_growth_and_feather_have_effective_support(self):
|
|
image, mask = source()
|
|
_, _, context = prepare(image, mask, mask_grow=3)
|
|
self.assertEqual(float(context["mask"][0, 27, 57]), 1)
|
|
_, _, eroded = prepare(image, mask, mask_grow=-3)
|
|
self.assertEqual(float(eroded["mask"][0, 30, 60]), 0)
|
|
_, _, softened = backend._prepare(image, mask, width=64, height=64, mask_blur=2)
|
|
edge = float(softened["mask"][0, 30, 60])
|
|
self.assertGreater(edge, 0)
|
|
self.assertLess(edge, 1)
|
|
|
|
def test_fill_uses_surrounding_color_and_preserves_unmasked_pixels(self):
|
|
image = torch.ones(1, 64, 64, 3) * 0.8
|
|
image[:, 20:44, 20:44] = 0.1
|
|
mask = torch.zeros(1, 64, 64)
|
|
mask[:, 20:44, 20:44] = 1
|
|
original, _, _ = prepare(image, mask, area="whole_image", fill="original")
|
|
filled, prepared_mask, _ = prepare(image, mask, area="whole_image", fill="fill")
|
|
self.assertTrue(torch.equal(filled[prepared_mask == 0], original[prepared_mask == 0]))
|
|
self.assertGreater(float(filled[0, 32, 32].mean()), 0.7)
|
|
|
|
def test_resize_fill_content_is_removed_during_composite(self):
|
|
image, mask = source()
|
|
mask.fill_(1)
|
|
pixels, _, context = prepare(image, mask, area="whole_image", resize_mode="fill")
|
|
left, top, right, bottom = context["records"][0]["content"]
|
|
self.assertGreater(top, 0)
|
|
generated = torch.zeros_like(pixels)
|
|
generated[:, top:bottom, left:right] = 0.7
|
|
output = backend._composite(generated, context)
|
|
self.assertTrue(torch.allclose(output, torch.full_like(output, 0.7), atol=1e-6))
|
|
|
|
def test_embedded_source_and_mask_convention(self):
|
|
rgba = Image.new("RGBA", (8, 4), (10, 20, 30, 255))
|
|
rgba.putpixel((0, 0), (10, 20, 30, 0))
|
|
pixels, alpha_mask = backend.AdachiInlineMask().load(source_data=data_url(rgba))
|
|
self.assertEqual(tuple(pixels.shape), (1, 4, 8, 3))
|
|
self.assertEqual(float(alpha_mask[0, 0, 0]), 1)
|
|
self.assertEqual(float(alpha_mask[0, 1, 1]), 0)
|
|
blackwhite = Image.new("L", (8, 4), 0)
|
|
blackwhite.putpixel((2, 1), 255)
|
|
_, explicit_mask = backend.AdachiInlineMask().load(source_data=data_url(rgba), mask_data=data_url(blackwhite))
|
|
self.assertEqual(float(explicit_mask[0, 1, 2]), 1)
|
|
self.assertEqual(float(explicit_mask[0, 0, 0]), 0)
|
|
|
|
def test_reject_invalid_data_dimensions_and_batch(self):
|
|
for data in ("data:image/svg+xml;base64,PHN2Zz4=", "not an image", "data:image/png;base64,!!"):
|
|
with self.assertRaises(ValueError):
|
|
backend._data_image(data)
|
|
image, mask = source()
|
|
for size in (63, 65, 8200):
|
|
with self.assertRaises(ValueError):
|
|
backend._prepare(image, mask, width=size, height=64)
|
|
with self.assertRaises(ValueError):
|
|
prepare(image, mask.repeat(2, 1, 1))
|
|
|
|
def test_latent_noise_zero_modes_and_nonfour_channel_vae(self):
|
|
image, mask = source(height=64, width=64)
|
|
mask.fill_(1)
|
|
captured = []
|
|
def sample(*args, **kwargs):
|
|
captured.append({key: value.clone() if isinstance(value, torch.Tensor) else value for key, value in args[8].items()})
|
|
return (args[8],)
|
|
with patch.dict(sys.modules, {"nodes": types.SimpleNamespace(common_ksampler=sample)}):
|
|
node = backend.AdachiWebUIInpaint()
|
|
for fill in ("latent_noise", "latent_noise", "latent_nothing", "original"):
|
|
node.sample(None, [], [], FakeVAE(channels=16), image, mask, fill=fill, seed=42,
|
|
width=64, height=64, area="whole_image", mask_blur=0, batch_size=2)
|
|
self.assertEqual(tuple(captured[0]["samples"].shape), (2, 16, 8, 8))
|
|
self.assertEqual(tuple(captured[0]["noise_mask"].shape), (2, 1, 8, 8))
|
|
self.assertTrue(torch.equal(captured[0]["samples"], captured[1]["samples"]))
|
|
self.assertFalse(torch.equal(captured[0]["samples"][0], captured[0]["samples"][1]))
|
|
self.assertTrue(torch.equal(captured[2]["samples"], torch.zeros_like(captured[2]["samples"])))
|
|
self.assertGreater(float(captured[3]["samples"].abs().sum()), 0)
|
|
|
|
def test_refiner_switch_uses_two_contiguous_sampler_stages(self):
|
|
image, mask = source(height=64, width=64)
|
|
calls = []
|
|
def sample(*args, **kwargs):
|
|
calls.append((args, kwargs))
|
|
return (args[8],)
|
|
base_model, refiner = object(), object()
|
|
with patch.dict(sys.modules, {"nodes": types.SimpleNamespace(common_ksampler=sample)}):
|
|
backend.AdachiWebUIInpaint().sample(base_model, ["base+"], ["base-"], FakeVAE(), image, mask,
|
|
steps=20, refiner_model=refiner, refiner_positive=["ref+"], refiner_negative=["ref-"],
|
|
refiner_switch=0.8, width=64, height=64, mask_blur=0)
|
|
self.assertEqual(len(calls), 2)
|
|
self.assertIs(calls[0][0][0], base_model)
|
|
self.assertIs(calls[1][0][0], refiner)
|
|
self.assertEqual(calls[0][1]["last_step"], 16)
|
|
self.assertEqual(calls[1][1]["start_step"], 16)
|
|
self.assertTrue(calls[1][1]["disable_noise"])
|
|
self.assertEqual(calls[1][0][6], ["ref+"])
|
|
|
|
def test_dedicated_inpaint_uses_native_conditioning_and_noise_mask(self):
|
|
image, mask = source(height=64, width=64)
|
|
calls = []
|
|
class NativeConditioning:
|
|
def encode(self, **kwargs):
|
|
calls.append(kwargs)
|
|
return ["native+"], ["native-"], {"samples": kwargs["vae"].encode(kwargs["pixels"]), "native_field": "preserved"}
|
|
def sample(*args, **kwargs):
|
|
self.assertEqual(args[6], ["native+"])
|
|
self.assertEqual(args[7], ["native-"])
|
|
self.assertEqual(args[8]["native_field"], "preserved")
|
|
self.assertEqual(tuple(args[8]["noise_mask"].shape), (1, 1, 8, 8))
|
|
return (args[8],)
|
|
fake_nodes = types.SimpleNamespace(common_ksampler=sample, InpaintModelConditioning=NativeConditioning)
|
|
with patch.dict(sys.modules, {"nodes": fake_nodes}):
|
|
backend.AdachiWebUIInpaint().sample(None, [], [], FakeVAE(), image, mask,
|
|
inpaint_model=True, width=64, height=64, mask_blur=0)
|
|
self.assertEqual(len(calls), 1)
|
|
self.assertTrue(calls[0]["noise_mask"])
|
|
|
|
def test_vae_compression_alignment_restores_original_dimensions(self):
|
|
image, mask = source()
|
|
vae = FakeVAE()
|
|
vae.downscale_ratio = 16
|
|
captured = []
|
|
def sample(*args, **kwargs):
|
|
captured.append(tuple(args[8]["samples"].shape))
|
|
return (args[8],)
|
|
with patch.dict(sys.modules, {"nodes": types.SimpleNamespace(common_ksampler=sample)}):
|
|
output, crop, prepared_mask = backend.AdachiWebUIInpaint().sample(None, [], [], vae, image, mask,
|
|
width=72, height=88, mask_blur=0)
|
|
self.assertEqual(captured, [(1, 4, 6, 5)])
|
|
self.assertEqual(tuple(crop.shape), (1, 96, 80, 3))
|
|
self.assertEqual(tuple(prepared_mask.shape), (1, 96, 80))
|
|
self.assertEqual(output.shape, image.shape)
|
|
self.assertTrue(torch.equal(output[mask == 0], image[mask == 0]))
|
|
|
|
def test_incompatible_refiner_latent_format_is_rejected(self):
|
|
def model(channels):
|
|
return types.SimpleNamespace(model=types.SimpleNamespace(latent_format=types.SimpleNamespace(latent_channels=channels)))
|
|
with self.assertRaisesRegex(ValueError, "incompatible"):
|
|
backend._check_refiner(model(4), model(16))
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|