Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation
Python agentic-video-generationmutli-agentmutli-agent-systemspatial-consistencyvideo-generationvideo-generator
import json
from types import SimpleNamespace
import numpy as np
import pytest
from atas.shot.camera_planner import plan_camera, observation_indices
def test_recapture_does_not_change_event_time():
assert observation_indices({'frame_index':10,'recon_frame_count':150},{'render_mode':'recapture'})==[10]
assert observation_indices({'frame_index':149,'recon_frame_count':150},{})==[149,74,0]
def test_selected_historical_frame_is_bound_to_gpu_context_and_cached(tmp_path,monkeypatch):
from atas.shot import camera_planner as c
from atas.providers import vlm
source=tmp_path/'story_story/source';source.mkdir(parents=True)
(source/'result.mp4').write_bytes(b'film');out=tmp_path/'target';out.mkdir()
base=dict(context_id='last',source_shot_id='source',frame_index=9,recon_frame_count=10)
calls=[]
def ctx(intent):
n=intent.get('camera_grounding_frame_index',9)
return {**base,'frame_index':n,'context_id':str(n)}
monkeypatch.setattr(c,'reference_frame',lambda v,d:(np.full((8,8,3),d['frame_index'],dtype=np.uint8),d['frame_index']))
monkeypatch.setattr('atas.world.identity.appearance_pairs',lambda s:[])
monkeypatch.setattr('atas.request.planning_context',lambda s:{})
response=dict(schema_version=1,status='ok',reference_observation_index=1,
subjects=[dict(label='Alice',kind='dynamic',part='face',visible=True,confidence=.95,bbox=[.1,.1,.8,.8],anchor_uv=[.4,.4])],
framing=dict(side_degrees=0,elevation_degrees=0,screen_position=[.5,.5],vertical_fraction=.6,horizontal_fraction=.6),
motion=dict(name='locked',degrees=0,travel_ratio=0))
def chat(*args,images,**kwargs):
calls.append([int(im[0,0,0]) for im in images]);return json.dumps(response)
monkeypatch.setattr(vlm,'chat_with_retry',chat)
shot=SimpleNamespace(story_id='story',shot_id='target',entity_ids=['Alice'],prompt='Alice',raw={},
shot_perspective='',shot_size='close_up',camera_angle='',camera_move_name='locked',camera_side_degrees=0,camera_movement_degrees=0)
g=SimpleNamespace(camera_context=ctx)
a=plan_camera(g,{'camera_relation':'continuous'},shot,tmp_path,out)
b=plan_camera(g,{'camera_relation':'continuous'},shot,tmp_path,out)
assert a==b and calls==[[9,4,0]]
assert a['camera_grounding_frame_index']==4
assert a['camera_plan']['context']['context_id']=='4'
assert a['camera_relation']=='continuous'
def test_missing_evidence_never_retries_same_images_even_on_resume(tmp_path):
from atas.runner.ai_recovery import corrected_json, AIRecoveryExhausted
from atas.shot.camera_planner import _parse
calls=[]
def generate(history):
calls.append(1)
return json.dumps(dict(schema_version=1,status='unresolved',reason='actor absent in all three frames'))
for _ in range(2):
with pytest.raises(AIRecoveryExhausted):corrected_json(tmp_path,'key',generate,_parse)
assert calls==[1]
def test_planned_prop_is_not_observed_or_allowed_in_recapture():
from atas.shot.camera_planner import validate_for_shot
shot=SimpleNamespace(raw={},entity_ids=['Alice'],camera_move_name='locked',prompt='Alice holds a wooden palette')
plan=dict(motion={'name':'locked'},subjects=[dict(label='Alice',kind='dynamic',framing_role='focal')],
planned_props=[dict(label='wooden palette',anchor_label='Alice',screenplay_evidence='holds a wooden palette',
placement='in her visible hand',status='planned_unobserved')])
assert validate_for_shot(plan,shot,{} ) is plan
with pytest.raises(ValueError,match='recapture'):validate_for_shot(plan,shot,{'render_mode':'recapture'})
plan['planned_props'][0]['label']='Alice'
with pytest.raises(ValueError,match='impersonate'):validate_for_shot(plan,shot,{})