fix: eliminate scene canvas latency in remote Gradio event chain

enhanced/topbar.py

@@ -1664,6 +1664,11 @@ def wait_for_vlm_completion(check_interval=0.5):
         logger.error(f"VLM Error: {str(e)}")
         return None
 def avoid_empty_prompt_for_scene(prompt, state, canvas_image, input_image1, scene_theme, additional_prompt, additional_prompt_2):
+    if canvas_image is None and isinstance(state, dict):
+        _cache_key = state.get('user_did', '') or 'default'
+        _cached = util.get_scene_canvas_cache(_cache_key)
+        if _cached is not None:
+            canvas_image = _cached
     describe_prompt = None
     if not prompt and 'scene_frontend' in state:
         visible = _scene_disvisible_with_optional_inputs(state["scene_frontend"])
@@ -1879,6 +1884,11 @@ def _scene_director_audio_status_for_generation(scene_director_enabled=False, sc
 
 
 def process_before_generation(state_params, seed_random, image_seed, backend_params, scene_theme, scene_canvas_image, scene_input_image1, scene_input_image2, scene_input_image3, scene_input_image4, scene_additional_prompt, scene_additional_prompt_2, scene_var_number, scene_var_number2, scene_var_number3, scene_var_number4, scene_var_number5, scene_var_number6, scene_var_number7, scene_var_number8, scene_var_number9, scene_var_number10, scene_steps, scene_switch_option1, scene_switch_option2, scene_switch_option3, scene_switch_option4, scene_aspect_ratio, scene_image_number, scene_video, scene_audio, scene_original_video_path, active_video_source, sam3_input_video, sam3_original_video_path, sam3_mask_video, overwrite_width=None, overwrite_height=None, resolution_multiplier=1.0, resolution_quantize_step=None, resolution_edit_mode=None, resolution_original_input=False, sam3_trim_payload=None, overwrite_step=None, scene_director_enabled=False, scene_director_state=None, scene_video_duration=None, scene_reference_video=None, scene_reference_video_original_path=None):
+    if scene_canvas_image is None and isinstance(state_params, dict):
+        _cache_key = state_params.get('user_did', '') or 'default'
+        _cached = util.get_scene_canvas_cache(_cache_key)
+        if _cached is not None:
+            scene_canvas_image = _cached
     regen_scene_additional_prompt = scene_additional_prompt
     regen_scene_additional_prompt_2 = scene_additional_prompt_2
     user_did = _state_user_did(state_params)

modules/util.py

@@ -13,6 +13,8 @@
 
 import json
 import hashlib
+import threading
+import time as _time_module
 
 from PIL import Image
 
@@ -48,6 +50,36 @@ def log_ui_trace(target_logger, *args, **kwargs):
         pass
 
 
+_scene_canvas_cache: dict = {}
+_scene_canvas_cache_lock = threading.Lock()
+_SCENE_CANVAS_CACHE_TTL = 300
+
+
+def store_scene_canvas_cache(key, value):
+    with _scene_canvas_cache_lock:
+        now = _time_module.monotonic()
+        for k in list(_scene_canvas_cache.keys()):
+            if now - _scene_canvas_cache[k][1] > _SCENE_CANVAS_CACHE_TTL:
+                del _scene_canvas_cache[k]
+        _scene_canvas_cache[key] = (value, now)
+
+
+def get_scene_canvas_cache(key):
+    with _scene_canvas_cache_lock:
+        entry = _scene_canvas_cache.get(key)
+        if entry and _time_module.monotonic() - entry[1] <= _SCENE_CANVAS_CACHE_TTL:
+            return entry[0]
+        return None
+
+
+def clear_scene_canvas_cache(key=None):
+    with _scene_canvas_cache_lock:
+        if key is not None:
+            _scene_canvas_cache.pop(key, None)
+        else:
+            _scene_canvas_cache.clear()
+
+
 def erode_or_dilate(x, k):
     k = int(k)
     if k > 0:

webui.py

@@ -2625,6 +2625,7 @@ def _launch_root_app_with_frontend_port_retry(**kwargs):
     scene_original_video_path = gr.State(None)
     scene_original_video_backup = gr.State(None)
     scene_reference_video_original_path = gr.State(None)
+    _scene_canvas_pass = gr.State(None)
     active_video_source = gr.State(None)
     resolution_source_meta = gr.Textbox(value="{}", visible="hidden", elem_id="resolution_source_meta", elem_classes=["resolution-hidden-control"])
     resolution_quantize_step = gr.Number(value=flags.default_resolution_quantize_step, visible="hidden", elem_id="resolution_quantize_step", elem_classes=["resolution-hidden-control"])
@@ -8559,6 +8560,8 @@ def cache_input_image_func(state_params, enhance_enabled, tab, uov, inpaint, enh
             enhance_enabled_bool = bool(enhance_enabled)
 
             if is_scene:
+                _cache_key = state_params.get('user_did', '') or 'default'
+                util.store_scene_canvas_cache(_cache_key, scene_canvas)
                 processed, label, scene_hidden, scene_candidates = resolve_scene_comparison_input(state_params, scene1, scene_canvas)
                 if processed is not None:
                     util.log_ui_trace(
@@ -8999,7 +9002,7 @@ def bind_generation_failure_cleanup(event):
             topbar.process_before_generation,
             inputs=[
                 state_topbar, seed_random, image_seed, params_backend, scene_theme,
-                scene_canvas_image, scene_input_image1, scene_input_image2, scene_input_image3, scene_input_image4,
+                _scene_canvas_pass, scene_input_image1, scene_input_image2, scene_input_image3, scene_input_image4,
                 scene_additional_prompt, scene_additional_prompt_2,
                 scene_var_number, scene_var_number2, scene_var_number3, scene_var_number4,
                 scene_var_number5, scene_var_number6, scene_var_number7, scene_var_number8,
@@ -9017,7 +9020,7 @@ def bind_generation_failure_cleanup(event):
         ))
         generate_event = bind_generation_failure_cleanup(generate_event.success(_sync_model_params_state_from_ui, inputs=model_state_ui_inputs, outputs=model_params_state, show_progress=False, queue=False))
         generate_event = bind_generation_failure_cleanup(generate_event.success(topbar.wait_for_vlm_completion, outputs=[], show_progress=False, queue=False))
-        generate_event = bind_generation_failure_cleanup(generate_event.success(topbar.avoid_empty_prompt_for_scene, inputs=[prompt, state_topbar, scene_canvas_image, scene_input_image1, scene_theme, scene_additional_prompt, scene_additional_prompt_2], outputs=prompt, show_progress=False, queue=False))
+        generate_event = bind_generation_failure_cleanup(generate_event.success(topbar.avoid_empty_prompt_for_scene, inputs=[prompt, state_topbar, _scene_canvas_pass, scene_input_image1, scene_theme, scene_additional_prompt, scene_additional_prompt_2], outputs=prompt, show_progress=False, queue=False))
         generate_event = bind_generation_failure_cleanup(generate_event.success(apply_scene_director_prompt_for_generation, inputs=[prompt, params_backend, scene_director_enabled, scene_director_state, state_topbar, scene_theme], outputs=[prompt, params_backend], show_progress=False, queue=False))
         generate_event = bind_generation_failure_cleanup(generate_event.success(select_random_aspect_ratio, inputs=[random_aspect_ratio_checkbox, random_aspect_ratio_state, aspect_ratios_selection], outputs=[overwrite_width, overwrite_height, aspect_ratios_selection, random_aspect_ratio_state], show_progress=False, queue=False))
         generate_event = bind_generation_failure_cleanup(generate_event.success(sync_quick_enhance_for_generation, inputs=[quick_enhance, quick_enhance_uov_strength, enhance_checkbox, enhance_uov_method, enhance_uov_strength], outputs=[enhance_checkbox, enhance_uov_method, enhance_uov_strength], show_progress=False, queue=False, js=sync_enhance_submit_state_js))