using ZeroGPU

README.md

@@ -9,6 +9,9 @@ app_file: app.py
 pinned: false
 license: openrail
 short_description: Testing generate Images from Controlnet
+preload_from_hub:
+  - runwayml/stable-diffusion-v1-5
+  - monster-labs/control_v1p_sd15_qrcode_monster
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -1,137 +1,94 @@
+import io, random, spaces
 import gradio as gr
-import torch
-from diffusers import StableDiffusionPipeline
 from PIL import Image
 import qrcode
 from qrcode.constants import ERROR_CORRECT_H
 
-# ========= device/dtype =========
-device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
-dtype = torch.float16 if device != "cpu" else torch.float32
-
-# ========= SD 1.5 (prompt-only) =========
-sd_pipe = StableDiffusionPipeline.from_pretrained(
-    "runwayml/stable-diffusion-v1-5",
-    torch_dtype=dtype
-).to(device)
-
-def sd_generate(prompt, negative, steps, guidance, seed):
-    gen = torch.Generator(device=device).manual_seed(int(seed)) if int(seed) != 0 else None
-    def run():
-        return sd_pipe(
-            prompt,
-            negative_prompt=negative or "",
-            num_inference_steps=int(steps),
-            guidance_scale=float(guidance),
-            generator=gen
-        ).images[0]
-    if device in ("cuda", "mps"):
-        with torch.autocast(device):
-            return run()
-    return run()
+import torch
+from diffusers import StableDiffusionControlNetPipeline, ControlNetModel, DPMSolverMultistepScheduler
+
+SD15 = "runwayml/stable-diffusion-v1-5"
+QR_MONSTER = "monster-labs/control_v1p_sd15_qrcode_monster"  # v2 lives in a subfolder
+
+# ----- Load models on first GPU call (saves ZeroGPU minutes) -----
+pipe = None
+def get_pipe(dtype=torch.float16):
+    global pipe
+    if pipe is not None:
+        return pipe
+    controlnet = ControlNetModel.from_pretrained(
+        QR_MONSTER, subfolder="v2", torch_dtype=dtype
+    )
+    pipe = StableDiffusionControlNetPipeline.from_pretrained(
+        SD15, controlnet=controlnet, torch_dtype=dtype, safety_checker=None
+    )
+    pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+    pipe.enable_model_cpu_offload()  # better mem usage on H200 slice
+    return pipe
 
-# ========= QR Maker =========
-def make_qr(url: str = "http://www.mybirdfire.com", size: int = 512, border: int = 4) -> Image.Image:
+def make_qr_img(data: str, module_size: int = 16, border: int = 4) -> Image.Image:
     qr = qrcode.QRCode(
         version=None,
-        error_correction=ERROR_CORRECT_H,  # highest EC
-        box_size=10,
-        border=border
+        error_correction=ERROR_CORRECT_H,
+        box_size=module_size,
+        border=border,
     )
-    qr.add_data(url.strip())
+    qr.add_data(data)
     qr.make(fit=True)
-    img = qr.make_image(fill_color="black", back_color="white").convert("RGB")
-    return img.resize((size, size), resample=Image.NEAREST)
-
-# ========= SD1.5 ControlNet stylizer (canny + tile) =========
-from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
-from diffusers.schedulers.scheduling_euler_discrete import EulerDiscreteScheduler
-from controlnet_aux import CannyDetector
-
-BASE_15 = "runwayml/stable-diffusion-v1-5"
-CN_CANNY_15 = "lllyasviel/sd-controlnet-canny"
-CN_TILE_15  = "lllyasviel/control_v11f1e_sd15_tile"
-
-_cn = {"pipe": None}
-def _load_sd15_dual():
-    if _cn["pipe"] is None:
-        canny = ControlNetModel.from_pretrained(CN_CANNY_15, torch_dtype=dtype)
-        tile  = ControlNetModel.from_pretrained(CN_TILE_15,  torch_dtype=dtype)
-        pipe = StableDiffusionControlNetPipeline.from_pretrained(
-            BASE_15, controlnet=[canny, tile], torch_dtype=dtype, safety_checker=None
-        ).to(device)
-        pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)  # Sampler: Euler
-        pipe.enable_attention_slicing()
-        pipe.enable_vae_slicing()
-        _cn["pipe"] = pipe
-        _cn["canny_aux"] = CannyDetector()
-    return _cn["pipe"], _cn["canny_aux"]
-
-NEG_DEFAULT = "lowres, low contrast, blurry, jpeg artifacts, worst quality, bad anatomy, extra digits"
-
-def stylize_qr_sd15(prompt: str, negative: str, steps: int, guidance: float, seed: int,
-                    canny_low: int, canny_high: int, border: int):
-    # Make fresh QR each time
-    qr_img = make_qr("http://www.mybirdfire.com", size=512, border=int(border))
-    pipe, canny = _load_sd15_dual()
-    edges = canny(qr_img, low_threshold=int(canny_low), high_threshold=int(canny_high))
-
-    gen = torch.Generator(device=device).manual_seed(int(seed)) if int(seed) != 0 else None
-
-    # Control weights (canny, tile). Tune if too “lego” or too artsy.
-    cn_scales = [1.2, 0.6]
-
-    def run():
-        return pipe(
-            prompt=str(prompt),
-            negative_prompt=negative or NEG_DEFAULT,
-            image=[edges, qr_img],                              # canny first, tile second
-            controlnet_conditioning_scale=cn_scales,
-            num_inference_steps=int(steps),
-            guidance_scale=float(guidance),
-            generator=gen
-        ).images[0]
-
-    if device in ("cuda", "mps"):
-        with torch.autocast(device):
-            return run()
-    return run()
+    # Gray background (#808080) improves blend/readability for v2
+    img = qr.make_image(fill_color="black", back_color="#808080").convert("RGB")
+    return img
+
+@spaces.GPU(duration=120)
+def generate(
+    qr_text, prompt, negative_prompt, steps, guidance_scale,
+    controlnet_scale, width, height, seed
+):
+    pipe = get_pipe()
+    if seed is None or seed < 0:
+        seed = random.randint(0, 2**31 - 1)
+    generator = torch.Generator(device="cuda").manual_seed(seed)
+
+    control_img = make_qr_img(qr_text, module_size=16, border=4).resize((width, height), Image.NEAREST)
+
+    out = pipe(
+        prompt=prompt,
+        negative_prompt=negative_prompt,
+        num_inference_steps=steps,
+        guidance_scale=guidance_scale,
+        controlnet_conditioning_scale=controlnet_scale,  # higher => more readable
+        image=control_img,
+        width=width,
+        height=height,
+        generator=generator,
+    )
+    img = out.images[0]
+    return img, seed
 
-# ========= UI =========
 with gr.Blocks() as demo:
-    gr.Markdown("## Stable Diffusion + QR Code + ControlNet (SD1.5)")
-
-    with gr.Tab("Stable Diffusion (prompt → image)"):
-        prompt = gr.Textbox(label="Prompt", value="Sky, Moon, Bird, Blue, In the dark, Goddess, Sweet, Beautiful, Fantasy, Art, Anime")
-        negative = gr.Textbox(label="Negative Prompt", value="lowres, bad anatomy, worst quality")
-        steps  = gr.Slider(10, 50, value=30, label="Steps", step=1)
-        cfg    = gr.Slider(1, 12, value=7.0, label="Guidance Scale", step=0.1)
-        seed   = gr.Number(value=0, label="Seed (0 = random)", precision=0)
-        out_sd = gr.Image(label="Generated Image")
-        gr.Button("Generate").click(sd_generate, [prompt, negative, steps, cfg, seed], out_sd)
-
-    with gr.Tab("QR Maker (mybirdfire)"):
-        url   = gr.Textbox(label="URL/Text", value="http://www.mybirdfire.com")
-        size  = gr.Slider(256, 1024, value=512, step=64, label="Size (px)")
-        quiet = gr.Slider(0, 8, value=4, step=1, label="Border (quiet zone)")
-        out_qr = gr.Image(label="QR Code", type="pil")
-        gr.Button("Generate QR").click(make_qr, [url, size, quiet], out_qr)
-
-    with gr.Tab("QR Stylizer (SD1.5 canny + tile, Euler)"):
-        s_prompt = gr.Textbox(label="Style Prompt", value="Sky, Moon, Bird, Blue, In the dark, Goddess, Sweet, Beautiful, Fantasy, Art, Anime")
-        s_negative = gr.Textbox(label="Negative Prompt", value=NEG_DEFAULT)
-        s_steps  = gr.Slider(10, 50, value=28, label="Steps", step=1)
-        s_cfg    = gr.Slider(1, 12, value=7.0, label="CFG", step=0.1)
-        s_seed   = gr.Number(value=1470713301, label="Seed", precision=0)
-        canny_l  = gr.Slider(0, 255, value=80, step=1, label="Canny low")
-        canny_h  = gr.Slider(0, 255, value=160, step=1, label="Canny high")
-        s_border = gr.Slider(2, 10, value=6, step=1, label="QR border")
-        out_styl = gr.Image(label="Stylized QR")
-        gr.Button("Stylize").click(
-            stylize_qr_sd15,
-            [s_prompt, s_negative, s_steps, s_cfg, s_seed, canny_l, canny_h, s_border],
-            out_styl
-        )
+    gr.Markdown("# 🧩 QR‑Blended Image (ZeroGPU + ControlNet QR Monster v2)")
+
+    with gr.Row():
+        with gr.Column():
+            qr_text = gr.Textbox(label="QR Content (URL or text)", value="https://example.com")
+            prompt = gr.Textbox(
+                label="Prompt",
+                value="a detailed isometric cyberpunk city with neon lights, intricate geometry, depth, high detail, trending artstation"
+            )
+            negative = gr.Textbox(label="Negative Prompt", value="lowres, blurry, deformed, distorted, extra limbs, watermark, signature")
+            steps = gr.Slider(10, 50, value=30, step=1, label="Steps")
+            guidance = gr.Slider(1.0, 12.0, value=7.0, step=0.5, label="CFG Scale")
+            cn_scale = gr.Slider(0.2, 2.0, value=1.0, step=0.05, label="ControlNet Guidance (↑=more scannable)")
+            width  = gr.Slider(512, 1024, value=768, step=64, label="Width")
+            height = gr.Slider(512, 1024, value=768, step=64, label="Height")
+            seed = gr.Number(value=-1, precision=0, label="Seed (-1=random)")
+            go = gr.Button("Generate", variant="primary")
+        with gr.Column():
+            out_img = gr.Image(label="Result", interactive=False)
+            out_seed = gr.Number(label="Used Seed", interactive=False)
+
+    go.click(generate, inputs=[qr_text, prompt, negative, steps, guidance, cn_scale, width, height, seed],
+             outputs=[out_img, out_seed])
 
 if __name__ == "__main__":
     demo.launch()

requirements.txt

@@ -1,10 +1,11 @@
-torch
-diffusers
-transformers
-accelerate
+torch==2.4.0
+diffusers>=0.27.2
+transformers>=4.42.0
+accelerate>=0.31.0
 safetensors
-gradio
+xformers
+gradio>=4.29.0
 qrcode[pil]
-controlnet-aux
-mediapipe
-timm
+Pillow
+huggingface-hub
+spaces