add txt2img tab + infer_txt2img endpoint (SD 1.5)

app.py

@@ -38,12 +38,20 @@ def _get_adapter() -> AnimateDiffAdapter:
     return _adapter
 
 
+PIXEL_ART_LORA_REPO = "artificialguybr/pixelartredmond-1-5v-pixel-art-loras-for-sd-1-5"
+PIXEL_ART_LORA_WEIGHT_FILE = "PixelArtRedmond15V-PixelArt-PIXARFK.safetensors"
+PIXEL_ART_LORA_ADAPTER = "pixart"
+# Trigger words: "pixel art, PixArFK" should appear in the prompt for the LoRA
+# to engage. The probe + UI include them by default.
+
+
 def _get_sd_pipe():
-    """Lazily load SD 1.5 txt2img pipeline (runs inside @spaces.GPU context).
+    """Lazily load SD 1.5 txt2img pipeline + PixelArtRedmond LoRA (runs inside
+    @spaces.GPU context).
 
-    Cached at module scope so warm calls skip re-loading. Cold start is ~30s
-    including weight download on first call ever (cached in persistent storage
-    for subsequent cold starts).
+    Cached at module scope so warm calls skip re-loading. Cold start is ~30-60s
+    including weight downloads on first call ever (cached in persistent storage
+    for subsequent cold starts). LoRA weight is set per-call via set_adapters().
     """
     global _sd_pipe
     if _sd_pipe is not None:
@@ -59,6 +67,16 @@ def _get_sd_pipe():
     pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
     pipe = pipe.to("cuda")
     pipe.set_progress_bar_config(disable=True)
+    try:
+        pipe.load_lora_weights(
+            PIXEL_ART_LORA_REPO,
+            weight_name=PIXEL_ART_LORA_WEIGHT_FILE,
+            adapter_name=PIXEL_ART_LORA_ADAPTER,
+        )
+        pipe.set_adapters([PIXEL_ART_LORA_ADAPTER], adapter_weights=[0.9])
+        print(f"loaded pixel-art LoRA: {PIXEL_ART_LORA_REPO}", flush=True)
+    except Exception as e:
+        print(f"WARN: could not load LoRA {PIXEL_ART_LORA_REPO}: {e}", flush=True)
     _sd_pipe = pipe
     return pipe
 
@@ -106,15 +124,24 @@ def infer_txt2img(
     height: int,
     width: int,
     seed: int,
+    lora_weight: float,
 ) -> str:
     """Generate a single sprite from a text prompt. Returns path to a PNG.
 
-    Defaults tuned for pixel-art sprites: 512×512, 25 steps, guidance 7.5.
-    Caller downscales to target res (256×256 is the Dicer sprite size).
+    Defaults tuned for pixel-art sprites: 512×512, 25 steps, guidance 7.5,
+    LoRA strength 0.9. Caller downscales to target res (256×256 is the Dicer
+    sprite size).
+
+    Prompt must include "pixel art, PixArFK" to engage the LoRA. The UI
+    pre-populates these tokens; the probe always includes them.
     """
     import torch
 
     pipe = _get_sd_pipe()
+    try:
+        pipe.set_adapters([PIXEL_ART_LORA_ADAPTER], adapter_weights=[float(lora_weight)])
+    except Exception as e:
+        print(f"WARN: set_adapters failed: {e}", flush=True)
     g = torch.Generator(device="cuda").manual_seed(int(seed))
     out = pipe(
         prompt=prompt,
@@ -169,17 +196,17 @@ with gr.Blocks(title="Venture-Studio") as demo:
                 with gr.Column():
                     t2i_prompt = gr.Textbox(
                         value=(
-                            "pixel art, 16-bit fantasy goblin warrior, empty hands, "
-                            "no weapon, standing pose, full body, centered, "
-                            "white background, sharp pixels, blocky, retro game sprite"
+                            "pixel art, PixArFK, fantasy goblin warrior, green skin, "
+                            "leather armor, empty hands, unarmed, standing pose, "
+                            "full body, centered, white background, retro game sprite"
                         ),
-                        lines=3, label="prompt",
+                        lines=3, label="prompt (include 'pixel art, PixArFK' for LoRA)",
                     )
                     t2i_neg = gr.Textbox(
                         value=(
                             "sword, weapon, dagger, axe, staff, blurry, soft, "
-                            "anti-aliasing, smooth, photorealistic, 3d render, "
-                            "extra limbs, distorted"
+                            "photorealistic, 3d render, extra limbs, distorted, "
+                            "multiple characters"
                         ),
                         lines=2, label="negative_prompt",
                     )
@@ -189,6 +216,7 @@ with gr.Blocks(title="Venture-Studio") as demo:
                         t2i_height = gr.Slider(256, 768, value=512, step=64, label="height")
                         t2i_width = gr.Slider(256, 768, value=512, step=64, label="width")
                         t2i_seed = gr.Number(value=0, precision=0, label="seed (0 = random)")
+                        t2i_lora = gr.Slider(0.0, 1.5, value=0.9, step=0.05, label="LoRA weight (PixelArtRedmond)")
                     t2i_run = gr.Button("Generate sprite", variant="primary")
                 with gr.Column():
                     t2i_out = gr.Image(label="Generated sprite", height=512)
@@ -196,7 +224,7 @@ with gr.Blocks(title="Venture-Studio") as demo:
             t2i_run.click(
                 infer_txt2img,
                 inputs=[t2i_prompt, t2i_neg, t2i_steps, t2i_guidance,
-                        t2i_height, t2i_width, t2i_seed],
+                        t2i_height, t2i_width, t2i_seed, t2i_lora],
                 outputs=t2i_out,
                 api_name="infer_txt2img",
             )

studio/backends/hf_space.py

@@ -172,3 +172,50 @@ def _gradio_client_available() -> bool:
         return True
     except ImportError:
         return False
+
+
+def generate_sprite_via_space(
+    prompt: str,
+    *,
+    space_id: str,
+    negative_prompt: str = "",
+    num_inference_steps: int = 25,
+    guidance_scale: float = 7.5,
+    height: int = 512,
+    width: int = 512,
+    seed: int = 0,
+    lora_weight: float = 0.9,
+    hf_token: Optional[str] = None,
+    api_name: str = "/infer_txt2img",
+) -> PILImage.Image:
+    """Call the Space's txt2img endpoint and return the generated sprite.
+
+    This is a free-standing helper (not bound to the cursor/motion abstraction).
+    A sprite is a fresh artifact, not a transform — wrapping it in a PixelCursor
+    would add ceremony without benefit.
+
+    Prompt should include the PixelArtRedmond trigger words "pixel art, PixArFK"
+    to engage the loaded LoRA. lora_weight scales LoRA strength (0.0 disables,
+    1.5 is the upper end; default 0.9).
+
+    Raises ImportError if gradio_client is missing locally.
+    """
+    try:
+        from gradio_client import Client
+    except ImportError as e:
+        raise ImportError(INSTALL_HINT) from e
+
+    token = _resolve_hf_token(hf_token)
+    client = Client(space_id, token=token)
+    png_path = client.predict(
+        prompt,
+        negative_prompt,
+        int(num_inference_steps),
+        float(guidance_scale),
+        int(height),
+        int(width),
+        int(seed),
+        float(lora_weight),
+        api_name=api_name,
+    )
+    return PILImage.open(png_path).convert("RGB")