251 lines
9.6 KiB
Python
251 lines
9.6 KiB
Python
"""RR-R4 selected-Reference preparation and Qwen/DiT conditioning contracts."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
import torch
|
|
|
|
from ComfyUI_H3_Continuum_Join.reference import REFERENCE_SIZE_MATCH_OUTPUT
|
|
from ComfyUI_H3_Continuum_Join.v3.reference_routing import (
|
|
REFERENCE_SLOT_IDS,
|
|
compile_reference_routing_schedule,
|
|
)
|
|
from ComfyUI_H3_Continuum_Join.v3.reference_runtime import (
|
|
ReferenceInputSet,
|
|
ReferenceRoutingRuntime,
|
|
ReferenceRuntimeContractError,
|
|
rewrite_reference_tags,
|
|
)
|
|
from ComfyUI_H3_Continuum_Join.v3.reference_effective_plan import compile_group_picture_map
|
|
from ComfyUI_H3_Continuum_Join.v3.reference_storage_contract import (
|
|
ReferenceStorageContractError, build_reference_storage_plan,
|
|
)
|
|
|
|
|
|
def _image(value: float):
|
|
return torch.full((1, 64, 64, 3), value, dtype=torch.float32)
|
|
|
|
|
|
class _VAE:
|
|
def __init__(self):
|
|
self.values = []
|
|
|
|
def encode(self, image):
|
|
value = float(image.mean())
|
|
self.values.append(value)
|
|
return torch.full((1, 24, 1, 4, 4), value)
|
|
|
|
|
|
class _Clip:
|
|
def __init__(self):
|
|
self.calls = []
|
|
|
|
def tokenize(self, prompt, **kwargs):
|
|
self.calls.append((prompt, kwargs))
|
|
return prompt
|
|
|
|
def encode_from_tokens_scheduled(self, tokens):
|
|
return [[torch.zeros((1, 1, 2)), {"prompt": tokens}]]
|
|
|
|
|
|
def _selectors(**overrides):
|
|
selectors = {slot_id: "off" for slot_id in REFERENCE_SLOT_IDS}
|
|
selectors.update(overrides)
|
|
return selectors
|
|
|
|
|
|
def _runtime(*, slots, selectors, chunks=3, terminal=False):
|
|
images = [None] * 9
|
|
for slot_id, image in slots.items():
|
|
images[int(slot_id[1:]) - 1] = image
|
|
inputs = ReferenceInputSet(
|
|
tuple(images), 64, 64, REFERENCE_SIZE_MATCH_OUTPUT,
|
|
)
|
|
schedule = compile_reference_routing_schedule(
|
|
total_chunks=chunks,
|
|
terminal_merge_enabled=terminal,
|
|
mode="Custom",
|
|
selectors_by_slot=_selectors(**selectors),
|
|
)
|
|
return ReferenceRoutingRuntime(schedule=schedule, inputs=inputs)
|
|
|
|
|
|
def test_each_group_reaches_only_its_selected_image_in_vae_qwen_and_dit():
|
|
a, b = _image(0.25), _image(0.75)
|
|
# R9 is deliberately not a Tensor. An unselected image must not be validated.
|
|
runtime = _runtime(
|
|
slots={"R1": a, "R4": b, "R9": object()},
|
|
selectors={"R1": "1", "R4": "2"},
|
|
)
|
|
vae, clip = _VAE(), _Clip()
|
|
first = _image(0.1)
|
|
|
|
first_group = runtime.prepare_group(
|
|
physical_group=1, prompt="@R1 runs; <Subject 1> watches",
|
|
clip=clip, video_vae=vae, first_image=first, last_image=None,
|
|
)
|
|
assert first_group.selected_assets.source_slot_ids == ("R1",)
|
|
assert first_group.effective_prompt == "<Picture 2> runs; <Subject 1> watches"
|
|
assert first_group.picture_map.references[0].source_slot_id == "R1"
|
|
assert first_group.picture_map.references[0].picture_number == 2
|
|
assert len(vae.values) == 1 and vae.values[0] == pytest.approx(0.25)
|
|
first_items = [item["data"] for item in clip.calls[0][1]["minimax_ref_items"]]
|
|
assert first_items[0] is first and torch.equal(first_items[1], a)
|
|
assert len(first_group.conditioning[0][1]["minimax_refs"]) == 1
|
|
assert float(first_group.conditioning[0][1]["minimax_refs"][0]["latent"].mean()) == pytest.approx(0.25)
|
|
|
|
second_group = runtime.prepare_group(
|
|
physical_group=2, prompt="@R4 runs; <Subject 1> watches",
|
|
clip=clip, video_vae=vae, first_image=first, last_image=None,
|
|
)
|
|
assert second_group.selected_assets.source_slot_ids == ("R4",)
|
|
assert second_group.effective_prompt == "<Picture 2> runs; <Subject 1> watches"
|
|
assert vae.values == pytest.approx([0.25, 0.75])
|
|
second_items = [item["data"] for item in clip.calls[1][1]["minimax_ref_items"]]
|
|
assert second_items[0] is first and torch.equal(second_items[1], b)
|
|
assert float(second_group.conditioning[0][1]["minimax_refs"][0]["latent"].mean()) == pytest.approx(0.75)
|
|
|
|
|
|
def test_empty_custom_group_never_falls_back_to_all_references():
|
|
runtime = _runtime(slots={"R1": _image(0.25)}, selectors={"R1": "1"})
|
|
vae, clip = _VAE(), _Clip()
|
|
result = runtime.prepare_group(
|
|
physical_group=2, prompt="no reference here", clip=clip, video_vae=vae,
|
|
first_image=None, last_image=None,
|
|
)
|
|
assert result.selected_assets is None
|
|
assert result.picture_map.references == ()
|
|
assert result.conditioning[0][1].get("minimax_refs") is None
|
|
assert clip.calls == [("no reference here", {"images": []})]
|
|
assert vae.values == []
|
|
|
|
|
|
def test_empty_nonfinal_group_keeps_legacy_last_image_timing():
|
|
runtime = _runtime(slots={}, selectors={}, chunks=2)
|
|
first, last = _image(0.1), _image(0.9)
|
|
clip = _Clip()
|
|
initial = runtime.prepare_group(
|
|
physical_group=1, prompt="running", clip=clip, video_vae=_VAE(),
|
|
first_image=first, last_image=last, include_last_image=False,
|
|
)
|
|
final = runtime.prepare_group(
|
|
physical_group=2, prompt="running", clip=clip, video_vae=_VAE(),
|
|
first_image=first, last_image=last, include_last_image=True,
|
|
)
|
|
assert initial.picture_map.last_picture_number is None
|
|
assert final.picture_map.last_picture_number == 2
|
|
assert clip.calls[0][1]["images"] == [first]
|
|
assert clip.calls[1][1]["images"] == [first, last]
|
|
|
|
|
|
def test_same_prompt_different_group_reference_is_encoded_twice():
|
|
runtime = _runtime(
|
|
slots={"R1": _image(0.25), "R4": _image(0.75)},
|
|
selectors={"R1": "1", "R4": "2"},
|
|
)
|
|
vae, clip = _VAE(), _Clip()
|
|
results = [
|
|
runtime.prepare_group(
|
|
physical_group=group, prompt="running", clip=clip, video_vae=vae,
|
|
first_image=None, last_image=None,
|
|
)
|
|
for group in (1, 2)
|
|
]
|
|
assert len(clip.calls) == 2
|
|
assert results[0].picture_map.references[0].image_sha256 != results[1].picture_map.references[0].image_sha256
|
|
assert results[0].conditioning is not results[1].conditioning
|
|
|
|
|
|
def test_disconnected_selected_slot_warns_without_changing_prompt_or_sampling_contract():
|
|
runtime = _runtime(slots={}, selectors={"R7": "1"}, chunks=1)
|
|
result = runtime.prepare_group(
|
|
physical_group=1, prompt="@R7 moves", clip=_Clip(), video_vae=_VAE(),
|
|
first_image=None, last_image=None,
|
|
)
|
|
assert result.effective_prompt == "@R7 moves"
|
|
assert result.selected_assets is None
|
|
assert any("R7 is routed" in warning for warning in result.warnings)
|
|
assert any("@R tag was left unchanged" in warning for warning in result.warnings)
|
|
|
|
|
|
def test_duplicate_content_keeps_two_fixed_slot_identities_and_picture_numbers():
|
|
image = _image(0.4)
|
|
runtime = _runtime(
|
|
slots={"R1": image, "R4": image},
|
|
selectors={"R1": "all", "R4": "all"}, chunks=1,
|
|
)
|
|
result = runtime.prepare_group(
|
|
physical_group=1, prompt="@R1 and @R4", clip=_Clip(), video_vae=_VAE(),
|
|
first_image=None, last_image=None,
|
|
)
|
|
refs = result.picture_map.references
|
|
assert [(item.source_slot_id, item.picture_number) for item in refs] == [
|
|
("R1", 1), ("R4", 2)
|
|
]
|
|
assert refs[0].image_sha256 == refs[1].image_sha256
|
|
assert result.effective_prompt == "<Picture 1> and <Picture 2>"
|
|
|
|
|
|
def test_terminal_conflict_is_not_silently_merged_or_split():
|
|
runtime = _runtime(
|
|
slots={"R1": _image(0.25), "R4": _image(0.75)},
|
|
selectors={"R1": "2", "R4": "3"}, terminal=True,
|
|
)
|
|
with pytest.raises(ReferenceRuntimeContractError, match="conflicting terminal"):
|
|
runtime.check_physical_contract(chunks=3, terminal_merge_enabled=True)
|
|
with pytest.raises(ReferenceRuntimeContractError, match="no single Reference route"):
|
|
runtime.prepare_group(
|
|
physical_group=2, prompt="running", clip=_Clip(), video_vae=_VAE(),
|
|
first_image=None, last_image=None,
|
|
)
|
|
|
|
|
|
def test_raw_identity_excludes_never_selected_image_and_changes_with_route():
|
|
bad_unselected = object()
|
|
left = _runtime(
|
|
slots={"R1": _image(0.25), "R9": bad_unselected},
|
|
selectors={"R1": "1"}, chunks=1,
|
|
)
|
|
right = _runtime(
|
|
slots={"R1": _image(0.25), "R9": bad_unselected},
|
|
selectors={"R9": "1"}, chunks=1,
|
|
)
|
|
# Neither selected nor unselected raw image content is touched merely to
|
|
# construct an RR-R4 queue identity; RR-R5 owns persistent reuse identity.
|
|
assert len(left.global_identity("base")) == 64
|
|
right._run_nonce = left._run_nonce
|
|
assert left.global_identity("base") != right.global_identity("base")
|
|
|
|
|
|
def test_reference_rewrite_is_exact_and_never_rewrites_subject_or_partial_tokens():
|
|
picture_map = compile_group_picture_map(
|
|
source_slot_ids=("R4",),
|
|
reference_image_hashes=("a" * 64,),
|
|
has_first_image=True,
|
|
has_last_image=True,
|
|
)
|
|
effective, warnings = rewrite_reference_tags(
|
|
"@R4 <Subject 4> @R5 @R10 X@R4 @R4suffix", picture_map
|
|
)
|
|
assert effective == "<Picture 3> <Subject 4> @R5 @R10 X@R4 @R4suffix"
|
|
assert len(warnings) == 1 and "R5" in warnings[0]
|
|
|
|
|
|
def test_rr_r5_changed_selected_image_is_rejected_before_vae_or_prompt_encode():
|
|
image = _image(0.25)
|
|
runtime = _runtime(slots={"R1": image}, selectors={"R1": "1"}, chunks=1)
|
|
plan = build_reference_storage_plan(
|
|
runtime=runtime, prompts=["@R1 runs"],
|
|
first_frame_hash="none", last_frame_hash="none",
|
|
)
|
|
image.fill_(0.75)
|
|
vae, clip = _VAE(), _Clip()
|
|
with pytest.raises(ReferenceStorageContractError, match="changed after"):
|
|
runtime.prepare_group(
|
|
physical_group=1, prompt="@R1 runs", clip=clip, video_vae=vae,
|
|
first_image=None, last_image=None,
|
|
expected_group_contract=plan.group_for(1),
|
|
)
|
|
assert vae.values == []
|
|
assert clip.calls == []
|