MMnode/tests/test_sam3_refs_composite.py
turtle89431 4ea8431cd0
Some checks are pending
Publish to Comfy registry / Publish Custom Node to registry (push) Waiting to run
init
2026-09-26 11:11:23 -07:00

218 lines
6.7 KiB
Python

import importlib.util
import json
import sys
import types
from pathlib import Path
import torch
from test_multitrack_info_output import _load_basic_module
def _load_image_module():
_load_basic_module()
path = Path(__file__).parents[1] / "nodes" / "image.py"
spec = importlib.util.spec_from_file_location("easy_media.nodes.image", path)
module = importlib.util.module_from_spec(spec)
sys.modules[spec.name] = module
spec.loader.exec_module(module)
return module
def test_make_refs_composite_schema_accepts_lists_and_exposes_expected_widgets():
module = _load_image_module()
schema = module.MakeRefsCompositeBySam3.define_schema()
assert schema.node_id == "easy makeRefsCompositeBySam3"
assert schema.is_input_list is True
assert [input_.name for input_ in schema.inputs] == [
"model",
"clip",
"images",
"prompt",
"width",
"height",
"detection_threshold",
"background",
"composite_mode",
]
assert schema.inputs[-1].kwargs["options"] == ["original", "sam3_masked"]
assert schema.inputs[-1].kwargs["default"] == "original"
assert [output.name for output in schema.outputs] == ["composite", "mask"]
def test_make_refs_composite_has_complete_chinese_localization():
locale_path = Path(__file__).parents[1] / "locales" / "zh" / "nodeDefs.json"
node_defs = json.loads(locale_path.read_text(encoding="utf-8"))
translation = node_defs["easy makeRefsCompositeBySam3"]
assert translation["display_name"] == "SAM3 制作参考图拼贴"
assert set(translation["inputs"]) == {
"model",
"clip",
"images",
"prompt",
"width",
"height",
"detection_threshold",
"background",
"composite_mode",
}
assert translation["inputs"]["background"]["options"] == {
"white": "白色",
"black": "黑色",
}
assert translation["inputs"]["composite_mode"]["options"] == {
"original": "原图",
"sam3_masked": "SAM3 遮罩分割",
}
assert translation["outputs"] == {
"0": {"name": "拼贴图"},
"1": {"name": "遮罩"},
}
def test_expand_ref_images_splits_a_single_batched_input():
module = _load_image_module()
batch = torch.stack((torch.zeros(2, 3, 3), torch.ones(2, 3, 3)))
images = module._expand_ref_images([batch])
assert len(images) == 2
assert all(image.shape == (1, 2, 3, 3) for image in images)
assert torch.count_nonzero(images[0]) == 0
assert torch.all(images[1] == 1)
def test_compose_refs_ignores_empty_masks_and_centers_the_remaining_cutout():
module = _load_image_module()
ignored = torch.ones(1, 2, 2, 3)
included = torch.zeros(1, 2, 2, 3)
included[..., 0] = 1
empty_mask = torch.zeros(1, 2, 2)
full_mask = torch.ones(1, 2, 2)
composite, mask = module._compose_refs(
[(ignored, empty_mask), (included, full_mask)],
width=6,
height=4,
background="white",
)
assert composite.shape == (1, 4, 6, 3)
assert mask.shape == (1, 4, 6)
assert torch.all(composite[:, :, :1] == 1)
assert torch.all(composite[:, :, -1:] == 1)
assert torch.all(composite[0, :, 1:5, 0] == 1)
assert torch.all(composite[0, :, 1:5, 1:] == 0)
assert torch.all(mask[0, :, 1:5] == 1)
def test_compose_refs_uses_grid_when_full_height_items_do_not_fit():
module = _load_image_module()
image = torch.ones(1, 2, 4, 3)
mask = torch.ones(1, 2, 4)
composite, composite_mask = module._compose_refs(
[(image, mask), (image, mask), (image, mask)],
width=8,
height=4,
background="black",
)
assert composite.shape == (1, 4, 8, 3)
assert composite_mask.shape == (1, 4, 8)
assert torch.count_nonzero(composite) > 0
assert torch.count_nonzero(composite_mask) > 0
assert torch.all(composite[composite_mask.unsqueeze(-1).expand_as(composite) == 0] == 0)
def test_make_refs_composite_runs_sam3_for_each_image_in_a_batch(monkeypatch):
module = _load_image_module()
calls = []
masks = [torch.ones(1, 2, 2), torch.ones(1, 2, 2)]
class FakeVideoTrack:
@classmethod
def execute(cls, **kwargs):
calls.append(kwargs)
return types.SimpleNamespace(result=({"index": len(calls) - 1},))
class FakeTrackToMask:
@classmethod
def execute(cls, track_data):
return types.SimpleNamespace(result=(masks[track_data["index"]],))
sam3_module = types.ModuleType("comfy_extras.nodes_sam3")
sam3_module.SAM3_VideoTrack = FakeVideoTrack
sam3_module.SAM3_TrackToMask = FakeTrackToMask
monkeypatch.setitem(sys.modules, "comfy_extras", types.ModuleType("comfy_extras"))
monkeypatch.setitem(sys.modules, "comfy_extras.nodes_sam3", sam3_module)
class FakeClip:
def __init__(self):
self.prompts = []
def tokenize(self, prompt):
self.prompts.append(prompt)
return {"tokens": prompt}
def encode_from_tokens_scheduled(self, tokens):
return [(tokens, {})]
clip = FakeClip()
batch = torch.stack((torch.zeros(2, 2, 3), torch.ones(2, 2, 3)))
output = module.MakeRefsCompositeBySam3.execute(
model=[object()],
clip=[clip],
images=[batch],
width=[4],
height=[4],
prompt=["person"],
detection_threshold=[0.7],
background=["black"],
composite_mode=["sam3_masked"],
)
composite, composite_mask = output.values
assert clip.prompts == ["person"]
assert len(calls) == 2
assert all(call["detection_threshold"] == 0.7 for call in calls)
assert all(call["conditioning"] == [({"tokens": "person"}, {})] for call in calls)
assert composite.shape == (1, 4, 4, 3)
assert torch.count_nonzero(composite) > 0
assert torch.count_nonzero(composite_mask) > 0
def test_original_mode_composites_all_images_without_running_sam3():
module = _load_image_module()
class UnusedClip:
def tokenize(self, prompt):
raise AssertionError("Original mode must not encode a SAM3 prompt.")
foreground = torch.zeros(2, 2, 3)
foreground[..., 0] = 1
background = torch.zeros(2, 2, 3)
background[..., 1] = 1
batch = torch.stack((foreground, background))
output = module.MakeRefsCompositeBySam3.execute(
model=[object()],
clip=[UnusedClip()],
images=[batch],
width=[4],
height=[4],
prompt=[""],
detection_threshold=[0.5],
background=["black"],
composite_mode=["original"],
)
composite, composite_mask = output.values
assert torch.count_nonzero(composite[..., 0]) > 0
assert torch.count_nonzero(composite[..., 1]) > 0
assert torch.count_nonzero(composite_mask) > 0