Speed up fast mode: lower resolution cap, skip needless VAE tiling, fewer steps

Fast mode's 768px (was 1024px) output now stays under the VAE's tiling
threshold, so it decodes in a single pass instead of a tiled decode with
blend overhead that was only ever needed for high-detail's 2048px case.
Also gives fast mode a lower default step count (3 vs 4), following the
existing mode-scoped-default pattern already used for GPU duration.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

app.py

@@ -29,7 +29,6 @@ _log_uploader = LogUploader(
 
 MAX_SEED = np.iinfo(np.int32).max
 LANCZOS = getattr(Image, "Resampling", Image).LANCZOS
-MAX_OUTPUT_DIM = 2048
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
@@ -71,7 +70,7 @@ torch.backends.cudnn.allow_tf32 = True
 print("[startup] TF32 enabled", flush=True)
 
 print("[startup] importing dimensions...", flush=True)
-from dimensions import compute_output_dimensions, max_dim_for_mode
+from dimensions import compute_output_dimensions, max_dim_for_mode, MAX_OUTPUT_DIM_FAST
 print("[startup] importing diffusers...", flush=True)
 from diffusers import FlowMatchEulerDiscreteScheduler
 from diffusers.models.normalization import RMSNorm
@@ -190,7 +189,7 @@ pipe = QwenImageEditPlusPipeline.from_pretrained(
     torch_dtype=dtype,
 )
 _hb.set()
-pipe.vae.enable_tiling()
+pipe.vae.enable_tiling(tile_sample_min_height=MAX_OUTPUT_DIM_FAST, tile_sample_min_width=MAX_OUTPUT_DIM_FAST)
 print(f"[startup] pipeline loaded in {time.perf_counter()-_t1_load:.1f}s", flush=True)
 
 print("[startup] setting cuDNN SDPA attention processor...", flush=True)
@@ -624,7 +623,7 @@ with gr.Blocks() as demo:
     seed = gr.Slider(minimum=0, maximum=MAX_SEED, step=1, value=0, elem_id="gradio-seed", elem_classes="hidden-input", container=False)
     randomize_seed = gr.Checkbox(value=True, elem_id="gradio-randomize", elem_classes="hidden-input", container=False)
     guidance_scale = gr.Slider(minimum=1.0, maximum=10.0, step=0.1, value=1.0, elem_id="gradio-guidance", elem_classes="hidden-input", container=False)
-    steps = gr.Slider(minimum=1, maximum=50, step=1, value=4, elem_id="gradio-steps", elem_classes="hidden-input", container=False)
+    steps = gr.Slider(minimum=1, maximum=50, step=1, value=3, elem_id="gradio-steps", elem_classes="hidden-input", container=False)
     mode = gr.Textbox(value="fast", elem_id="gradio-mode", elem_classes="hidden-input", container=False)
     gpu_duration = gr.Slider(minimum=10, maximum=120, step=5, value=30, elem_id="gradio-gpu-duration", elem_classes="hidden-input", container=False)
     result = gr.Image(elem_id="gradio-result", elem_classes="hidden-input", container=False, format="png")

dimensions.py

@@ -1,5 +1,5 @@
 MAX_OUTPUT_DIM = 2048
-MAX_OUTPUT_DIM_FAST = 1024
+MAX_OUTPUT_DIM_FAST = 768
 
 
 def max_dim_for_mode(mode): return MAX_OUTPUT_DIM_FAST if mode == "fast" else MAX_OUTPUT_DIM

static/mode_toggle.js

@@ -1,18 +1,24 @@
 () => {
 window.__selectedMode = 'fast';
 var MODE_GPU_DURATION = { fast: 30, high_detail: 60 };
+var MODE_STEPS = { fast: 3, high_detail: 4 };
+
+function applyModeSliderDefault(sliderId, valueMap, mode) {
+    var val = valueMap[mode];
+    if (val === undefined) return;
+    var sl = document.getElementById(sliderId);
+    var vl = document.getElementById(sliderId + '-val');
+    if (sl) { sl.value = val; sl.dispatchEvent(new Event('input', {bubbles: true})); }
+    if (vl) vl.textContent = val;
+}
+
 window.__setMode = function(m) {
     window.__selectedMode = m;
     var fast = document.getElementById('mode-btn-fast');
     var hd   = document.getElementById('mode-btn-hd');
     if (fast) fast.classList.toggle('mode-btn-active', m === 'fast');
     if (hd)   hd.classList.toggle('mode-btn-active',   m === 'high_detail');
-    var dur = MODE_GPU_DURATION[m];
-    if (dur !== undefined) {
-        var sl = document.getElementById('custom-gpu-duration');
-        var vl = document.getElementById('custom-gpu-duration-val');
-        if (sl) { sl.value = dur; sl.dispatchEvent(new Event('input', {bubbles: true})); }
-        if (vl) vl.textContent = dur;
-    }
+    applyModeSliderDefault('custom-gpu-duration', MODE_GPU_DURATION, m);
+    applyModeSliderDefault('custom-steps', MODE_STEPS, m);
 };
 }

templates/app.html

@@ -139,8 +139,8 @@
                     </div>
                     <div class="slider-row">
                         <label>Steps</label>
-                        <input type="range" id="custom-steps" min="1" max="50" step="1" value="4">
-                        <span class="slider-val" id="custom-steps-val">4</span>
+                        <input type="range" id="custom-steps" min="1" max="50" step="1" value="3">
+                        <span class="slider-val" id="custom-steps-val">3</span>
                     </div>
                 </div>
             </div>

tests/test_dimensions.py

@@ -1,5 +1,5 @@
 import pytest
-from dimensions import compute_output_dimensions, MAX_OUTPUT_DIM
+from dimensions import compute_output_dimensions, max_dim_for_mode, MAX_OUTPUT_DIM, MAX_OUTPUT_DIM_FAST
 
 
 def aspect_ratio_error(w_in, h_in, w_out, h_out):
@@ -97,3 +97,14 @@ def test_custom_max_dim():
     nw, nh = compute_output_dimensions(1920, 1080, max_dim=1024)
     assert nw == 1024
     assert nh % 8 == 0
+
+
+# --- mode -> max dimension mapping ---
+
+def test_fast_mode_uses_fast_max_dim():
+    assert max_dim_for_mode("fast") == MAX_OUTPUT_DIM_FAST
+
+
+@pytest.mark.parametrize("mode", ["high_detail", "anything_else", None])
+def test_non_fast_modes_use_default_max_dim(mode):
+    assert max_dim_for_mode(mode) == MAX_OUTPUT_DIM