Files

314 lines
11 KiB
Python

from __future__ import annotations
import importlib.util
from pathlib import Path
import sys
import numpy as np
import torch
TOOLS = Path(__file__).parents[1] / "tools"
def _load(name: str, filename: str):
path = TOOLS / filename
spec = importlib.util.spec_from_file_location(name, path)
assert spec is not None and spec.loader is not None
module = importlib.util.module_from_spec(spec)
sys.modules[name] = module
spec.loader.exec_module(module)
return module
runner = _load("h3_issue13_r25_gate_runner", "issue13_r25_gate_runner.py")
analyzer = _load("h3_issue13_r25_analyze", "issue13_r2_analyze.py")
soft_reanchor = _load(
"h3_issue13_soft_reanchor",
"gpu_diagnostic_node/soft_reanchor.py",
)
def _source_prompt() -> dict:
return {
"114": {"inputs": {"image": "old.png"}},
"119": {"inputs": {}},
"139": {"inputs": {"sampler_name": "res_multistep"}},
"150": {"inputs": {"enabled": True}},
"153": {"inputs": {"scheduler": "old", "steps": 20, "denoise": 0.5}},
"188": {"inputs": {"value": "old"}},
"191": {"inputs": {}},
"249": {"inputs": {}},
"286": {
"inputs": {
"lora_1": {"on": True, "lora": "old.safetensors", "strength": 1.0}
}
},
"297": {"inputs": {"value": 0.3}},
"305": {
"class_type": "H3ContinuumSamplerV37",
"inputs": {
"continuation_backend": "Standard",
"generation_mode": "Full Run",
"review_action": "Continue / Next",
"guide": ["900", 0],
"last_frame": ["901", 0],
},
},
"306": {"inputs": {}},
}
def test_screening_order_and_contract_are_fixed():
assert tuple(runner.CASE_SPECS)[:3] == (
"screen_3x5_turbo",
"screen_3x10_turbo",
"screen_6x5_turbo",
)
prompt, loras = runner.configure_prompt(
_source_prompt(),
case="screen_3x10_turbo",
output_prefix="video/test",
image_name="closeup.png",
seed=131325,
width=576,
height=864,
diagnostic_nonce=7,
)
inputs = prompt["305"]["inputs"]
assert inputs["chunks"] == 3
assert inputs["chunk_seconds"] == 10.0
assert inputs["width"] == 576
assert inputs["height"] == 864
assert prompt["139"]["inputs"]["sampler_name"] == "euler"
assert prompt["153"]["inputs"]["steps"] == 8
assert prompt["153"]["inputs"]["scheduler"] == "simple"
assert prompt["114"]["inputs"]["image"] == "closeup.png"
assert loras == [{"name": runner.TURBO_LORA, "strength": 1.0}]
assert prompt["188"]["inputs"]["value"].count("\n---\n") == 2
assert "dynv2" not in prompt["188"]["inputs"]["value"]
def test_six_by_five_uses_six_motion_steps():
prompt, _ = runner.configure_prompt(
_source_prompt(),
case="screen_6x5_turbo",
output_prefix="video/test",
image_name="closeup.png",
seed=1,
width=576,
height=864,
diagnostic_nonce=8,
)
assert prompt["305"]["inputs"]["chunks"] == 6
assert prompt["188"]["inputs"]["value"].count("\n---\n") == 5
def test_six_by_five_fixed_prompt_repeats_one_prompt():
prompt, loras = runner.configure_prompt(
_source_prompt(),
case="screen_6x5_turbo_fixed_prompt",
output_prefix="video/issue13_r26/fixed",
image_name="closeup.png",
seed=131325,
width=576,
height=864,
diagnostic_nonce=132600,
)
prompts = prompt["188"]["inputs"]["value"].split("\n---\n")
assert len(prompts) == 6
assert len(set(prompts)) == 1
assert prompt["305"]["inputs"]["chunks"] == 6
assert prompt["305"]["inputs"]["chunk_seconds"] == 5.0
assert prompt["153"]["inputs"]["steps"] == 8
assert loras == [{"name": runner.TURBO_LORA, "strength": 1.0}]
def test_six_by_five_fixed_prefix_changes_only_diagnostic_source_mode():
source = _source_prompt()
recursive, _ = runner.configure_prompt(
source,
case="screen_6x5_turbo_fixed_prompt",
output_prefix="video/issue13_r26/recursive",
image_name="closeup.png",
seed=131325,
width=576,
height=864,
diagnostic_nonce=132600,
)
fixed, _ = runner.configure_prompt(
source,
case="screen_6x5_turbo_fixed_prefix",
output_prefix="video/issue13_r26/fixed-prefix",
image_name="closeup.png",
seed=131325,
width=576,
height=864,
diagnostic_nonce=132601,
)
recursive_inputs = recursive["305"]["inputs"]
fixed_inputs = fixed["305"]["inputs"]
assert recursive_inputs["continuation_source_mode"] == "Recursive"
assert fixed_inputs["continuation_source_mode"] == "Fixed Group 1 Tail"
ignored = {"continuation_source_mode", "diagnostic_nonce"}
assert {
key: value for key, value in recursive_inputs.items() if key not in ignored
} == {key: value for key, value in fixed_inputs.items() if key not in ignored}
assert recursive["188"]["inputs"]["value"] == fixed["188"]["inputs"]["value"]
def test_six_by_five_soft_reanchor_changes_only_diagnostic_source_mode():
source = _source_prompt()
recursive, _ = runner.configure_prompt(
source,
case="screen_6x5_turbo_fixed_prompt",
output_prefix="video/issue13_r27/recursive",
image_name="closeup.png",
seed=131325,
width=576,
height=864,
diagnostic_nonce=132700,
)
reanchored, _ = runner.configure_prompt(
source,
case="screen_6x5_turbo_soft_reanchor",
output_prefix="video/issue13_r27/soft-reanchor",
image_name="closeup.png",
seed=131325,
width=576,
height=864,
diagnostic_nonce=132701,
)
recursive_inputs = recursive["305"]["inputs"]
reanchored_inputs = reanchored["305"]["inputs"]
assert recursive_inputs["continuation_source_mode"] == "Recursive"
assert reanchored_inputs["continuation_source_mode"] == "Soft Re-anchor Group 4"
ignored = {"continuation_source_mode", "diagnostic_nonce"}
assert {
key: value for key, value in recursive_inputs.items() if key not in ignored
} == {key: value for key, value in reanchored_inputs.items() if key not in ignored}
assert recursive["188"]["inputs"]["value"] == reanchored["188"]["inputs"]["value"]
def test_soft_reanchor_moves_global_appearance_halfway_without_geometry_change():
current = torch.zeros((1, 4, 8, 8, 3), dtype=torch.float32)
current[..., 0] = 0.8
current[..., 1] = 0.3
current[..., 2] = 0.2
reference = torch.zeros_like(current)
reference[..., 0] = 0.4
reference[..., 1] = 0.5
reference[..., 2] = 0.6
adjusted, report = soft_reanchor.soft_match_rgb_appearance(
current,
reference,
strength=0.5,
)
assert adjusted.shape == current.shape
assert torch.equal(current[..., 0], torch.full_like(current[..., 0], 0.8))
before_gap = abs(
report["before"]["luma_mean"] - report["reference"]["luma_mean"]
)
after_gap = abs(
report["after"]["luma_mean"] - report["reference"]["luma_mean"]
)
assert after_gap < before_gap
assert report["strength"] == 0.5
assert report["geometry_unchanged"] is True
def test_soft_reanchor_vae_roundtrip_preserves_shape_and_inputs():
class FakeVideoVAE:
def decode(self, latent):
value = float(latent.mean().item())
pixels = torch.full((1, 22, 32, 32, 3), value, dtype=torch.float32)
pixels[..., 0] = value + 0.1
return pixels.clamp(0.0, 1.0)
def encode(self, pixels):
value = float(pixels.mean().item())
return torch.full((1, 24, 7, 2, 2), value, dtype=torch.float32)
current = torch.full((1, 24, 7, 2, 2), 0.7, dtype=torch.float32)
reference = torch.full((1, 24, 7, 2, 2), 0.3, dtype=torch.float32)
current_before = current.clone()
reference_before = reference.clone()
encoded, report = soft_reanchor.soft_reanchor_context(
current,
reference,
video_vae=FakeVideoVAE(),
strength=0.5,
)
assert encoded.shape == current.shape
assert encoded.dtype == current.dtype
assert torch.isfinite(encoded).all()
assert torch.equal(current, current_before)
assert torch.equal(reference, reference_before)
assert report["current_context_unchanged"] is True
assert report["reference_context_unchanged"] is True
assert report["latent_shape_unchanged"] is True
def test_boundary_metrics_report_flow_and_frame_jump():
frames = []
for index in range(12):
frame = np.zeros((24, 24, 3), dtype=np.uint8)
frame[:, max(0, min(20, index)) : max(1, min(24, index + 4))] = 80
if index >= 6:
frame = np.clip(frame + 80, 0, 255).astype(np.uint8)
frames.append(frame)
rows = analyzer._boundary_metrics(frames, chunks=2)
assert len(rows) == 1
assert rows[0]["target_group"] == 2
assert rows[0]["frame_mae"] > 0.0
assert rows[0]["trajectory_flow_discontinuity"] >= 0.0
def test_motion_pair_ab_uses_normal_twenty_step_and_no_turbo():
off, off_loras = runner.configure_prompt(
_source_prompt(),
case="motion_off_3x5_normal",
output_prefix="video/test",
image_name="closeup.png",
seed=1,
width=576,
height=864,
diagnostic_nonce=9,
)
both, both_loras = runner.configure_prompt(
_source_prompt(),
case="motion_both_3x5_normal",
output_prefix="video/test",
image_name="closeup.png",
seed=1,
width=576,
height=864,
diagnostic_nonce=10,
)
assert off_loras == []
assert off["153"]["inputs"]["steps"] == 20
assert "dynv2" in off["188"]["inputs"]["value"]
assert [item["name"] for item in both_loras] == [
runner.MOTION_BOOSTER_LORA,
runner.MOTION_REPAIR_LORA,
]
assert both["153"]["inputs"]["steps"] == 20
assert all(runner.TURBO_LORA != item["name"] for item in both_loras)
def test_continuation_rgb_proxy_uses_pre_boundary_context_window():
frames = [np.full((8, 8, 3), index, dtype=np.uint8) for index in range(30)]
result = analyzer._continuation_input_rgb_proxy(
frames, chunks=3, context_frames=4
)
assert result["context_frames"] == 4
assert len(result["boundaries"]) == 2
first = result["boundaries"][0]
assert first["boundary_frame"] == 10
assert first["window_start"] == 6
assert first["window_end_exclusive"] == 10
assert first["window_frames"] == 4
assert abs(sum(first["luma_histogram"]) - 1.0) < 1e-9