Upload app.py with huggingface_hub

app.py

@@ -67,7 +67,7 @@ stream = StreamDiffusionV2Pipeline(
     device=device,
     height=HEIGHT,
     width=WIDTH,
-    step=1,            # fewer denoising stages -> shallower pipeline -> lower lag
+    step=2,            # 2 denoising steps (quality); TAEHV keeps it fast
     noise_scale=NOISE_SCALE,
     model_type="T2V-1.3B",
     use_taehv=True,    # tiny-VAE decode -> much faster per-chunk -> lower lag
@@ -127,13 +127,24 @@ def run_session() -> str:
         pass
     yield READY_SENTINEL
 
-    prompt = _read_prompt() or DEFAULT_PROMPT
+    cur_prompt = _read_prompt() or DEFAULT_PROMPT
     buffer = []
     session = None
     deadline = time.time() + SESSION_DURATION
     last = None
 
     while time.time() < deadline:
+        # Live prompt change: re-encode text and reset the cross-attn cache,
+        # WITHOUT touching the rolling self-attn KV (keeps temporal continuity).
+        new_prompt = _read_prompt() or DEFAULT_PROMPT
+        if new_prompt != cur_prompt and session is not None:
+            cur_prompt = new_prompt
+            cond = PM.pipeline.text_encoder(text_prompts=[cur_prompt])
+            cond["prompt_embeds"] = cond["prompt_embeds"].repeat(PM.pipeline.batch_size, 1, 1)
+            PM.pipeline.conditional_dict = cond
+            for blk in PM.pipeline.crossattn_cache:
+                blk["is_init"] = False
+
         drained = 0
         while drained < 256:
             try:
@@ -156,7 +167,7 @@ def run_session() -> str:
 
         t0 = time.time()
         if session is None:
-            session, init_video = PM.start_stream_session(prompt, vid, NOISE_SCALE)
+            session, init_video = PM.start_stream_session(cur_prompt, vid, NOISE_SCALE)
             outs = [init_video]
         else:
             outs = PM.run_stream_batch(session, vid)