Speed up SDXL: LCM-LoRA (5 steps) + batched image generation

Replace 18-step CFG=7 SDXL with LCM-LoRA scheduler (~5 steps, guidance 1.2)
on the same base model. Batch target+2 decoys into a single GPU call instead
of 4 sequential calls. ~4-6x faster asset generation.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app.py

@@ -52,9 +52,14 @@ llm = AutoModelForCausalLM.from_pretrained(
     trust_remote_code=True,
 )
 
-from diffusers import StableDiffusionXLPipeline
+from diffusers import StableDiffusionXLPipeline, LCMScheduler
 
 IMAGE_MODEL_ID = "stabilityai/stable-diffusion-xl-base-1.0"
+LCM_LORA_ID = "latent-consistency/lcm-lora-sdxl"
+
+# Nombre de steps et guidance optimisés pour LCM (au lieu de 18 / 7.0 en SDXL classique)
+IMAGE_STEPS = 5
+IMAGE_GUIDANCE = 1.2
 
 pipe = StableDiffusionXLPipeline.from_pretrained(
     IMAGE_MODEL_ID,
@@ -63,7 +68,12 @@ pipe = StableDiffusionXLPipeline.from_pretrained(
     variant="fp16",
 )
 pipe.to("cuda")
-pipe.enable_attention_slicing()
+
+# LCM-LoRA : distillation par consistency model -> ~4-5 steps suffisent.
+# On réutilise le même SDXL base (déjà téléchargé), on ajoute juste le LoRA + le scheduler LCM.
+pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
+pipe.load_lora_weights(LCM_LORA_ID)
+pipe.fuse_lora()
 
 
 # ---------------------------------------------------------------------------
@@ -199,55 +209,59 @@ def pil_to_b64(img: Image.Image) -> str:
     return base64.b64encode(buffer.getvalue()).decode("utf-8")
 
 
-def generate_image(prompt, width=512, height=512, steps=18):
-    key = (prompt, width, height, steps)
+def generate_images_batch(prompts, width, height, steps=IMAGE_STEPS):
+    """Génère plusieurs images de MÊME taille en un seul appel GPU (batch).
 
-    if key in ASSET_CACHE:
-        return ASSET_CACHE[key]
+    Les prompts déjà en cache sont renvoyés directement ; seuls les manquants
+    sont générés, puis l'ordre d'origine est reconstruit.
+    """
+    results = [None] * len(prompts)
+    todo_idx, todo_prompts = [], []
 
-    with torch.no_grad():
-        image = pipe(
-            prompt=prompt,
-            negative_prompt=NEGATIVE_PROMPT,
-            width=width,
-            height=height,
-            num_inference_steps=steps,
-            guidance_scale=7.0,
-        ).images[0]
+    for i, p in enumerate(prompts):
+        key = (p, width, height, steps)
+        if key in ASSET_CACHE:
+            results[i] = ASSET_CACHE[key]
+        else:
+            todo_idx.append(i)
+            todo_prompts.append(p)
+
+    if todo_prompts:
+        with torch.no_grad():
+            out = pipe(
+                prompt=todo_prompts,
+                negative_prompt=[NEGATIVE_PROMPT] * len(todo_prompts),
+                width=width,
+                height=height,
+                num_inference_steps=steps,
+                guidance_scale=IMAGE_GUIDANCE,
+            ).images
 
-    ASSET_CACHE[key] = image
-    return image
+        for slot, p, img in zip(todo_idx, todo_prompts, out):
+            ASSET_CACHE[(p, width, height, steps)] = img
+            results[slot] = img
+
+    return results
 
 
 def generate_assets(spec: VisualMiniGameSpec):
     style = spec.visual_style
 
-    background = generate_image(
-        f"{spec.background_prompt}, {style}, no text, game background",
+    # Background (768x512) -> 1 appel
+    (background,) = generate_images_batch(
+        [f"{spec.background_prompt}, {style}, no text, game background"],
         width=768,
         height=512,
-        steps=18,
     )
 
-    target = generate_image(
+    # Target + 2 decoys (512x512, même taille) -> 1 seul appel batché
+    object_prompts = [
         f"{spec.target_prompt}, {style}, isolated object, white background, no text",
-        width=512,
-        height=512,
-        steps=18,
-    )
-
-    decoy_1 = generate_image(
         f"{spec.decoy_prompts[0]}, {style}, isolated object, white background, no text",
-        width=512,
-        height=512,
-        steps=18,
-    )
-
-    decoy_2 = generate_image(
         f"{spec.decoy_prompts[1]}, {style}, isolated object, white background, no text",
-        width=512,
-        height=512,
-        steps=18,
+    ]
+    target, decoy_1, decoy_2 = generate_images_batch(
+        object_prompts, width=512, height=512
     )
 
     return {