fix: eliminate scene canvas latency in remote Gradio event chain

[lonrencn]

Problem

scene_canvas_image is a gr.Textbox that stores a full base64 PNG JSON string (multiple MB). Gradio's walk_and_store_blobs() returns [] for string inputs, so every event-chain step re-uploads the entire base64 payload in the POST body.

For remote deployments (browser → reverse proxy → server), this adds ~15-20s of dead time between clicking Generate and actual generation start.

Root Cause

The Generate button triggers a 13-step Gradio event chain. Steps 1-12 run with queue=False, each as a separate HTTP POST. Three of these steps receive scene_canvas_image as an input:

Step	Handler	Purpose
2	cache_input_image_func	Caches input image
4	process_before_generation	Pre-generation setup
7	avoid_empty_prompt_for_scene	Prompt fallback

Each step sends the multi-MB base64 string over the wire.

Solution

Server-side cache bypass: Store the canvas data once during step 2 (cache_input_image_func), then pass a lightweight gr.State(None) in steps 4 and 7. The Python handlers transparently recover the cached value via user_did as the cache key.

Changes (3 files, +47/-2)

• modules/util.py: Thread-safe in-memory cache with 300s TTL auto-expiry
• webui.py: _scene_canvas_pass = gr.State(None) dummy component; cache store in step 2; steps 4 & 7 inputs changed
• enhanced/topbar.py: Cache fallback in process_before_generation() and avoid_empty_prompt_for_scene()

Measured Results (remote deployment, China → server)

Metric	Before	After
Click → Generation start	~27s	~10s
Payload per chain step	multi-MB base64	~4 bytes (null)

Local deployments are unaffected (sub-ms transfers). The cache auto-expires after 300s and uses user_did as the key, so concurrent users are isolated.

[lonrencn]

Tested on a remote deployment (browser in China → reverse proxy → server). After applying this fix, the click-to-generate delay dropped from ~27s to ~10s. The improvement is very noticeable in actual use.

[Windecay]

Manually merged and update.Thanks!