File size: 6,695 Bytes
58b2866
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
"""
Depth ControlNet Logo Generator
--------------------------------
- Nimmt ein hochgeladenes Logo/Referenzbild
- Berechnet daraus eine Depth Map (Midas)
- Generiert per Stable Diffusion 1.5 + ControlNet-Depth ein neues,
  strukturell ähnliches Bild nach Text-Prompt
- Läuft auf HuggingFace ZeroGPU (dynamisch zugewiesene GPU pro Request)
"""

import spaces
import gradio as gr
import numpy as np
import torch
from PIL import Image
from diffusers import (
    StableDiffusionControlNetPipeline,
    ControlNetModel,
    UniPCMultistepScheduler,
)
from transformers import pipeline as hf_pipeline

# ---------------------------------------------------------------------------
# Konfiguration
# ---------------------------------------------------------------------------
SD_MODEL_ID = "runwayml/stable-diffusion-v1-5"
CONTROLNET_ID = "lllyasviel/sd-controlnet-depth"
DEPTH_MODEL_ID = "Intel/dpt-hybrid-midas"  # gleicher Preprocessor wie beim ControlNet-Training
RESOLUTION = 512  # native SD1.5 Auflösung

device = "cuda"
dtype = torch.float16

# ---------------------------------------------------------------------------
# Modelle laden (einmalig beim Space-Start)
# Bei ZeroGPU: .to("cuda") hier ist ok, tatsächliche GPU wird erst bei
# @spaces.GPU-Aufrufen zugewiesen.
# ---------------------------------------------------------------------------
print("Lade Depth-Estimator ...")
depth_estimator = hf_pipeline("depth-estimation", model=DEPTH_MODEL_ID)

print("Lade ControlNet + Stable Diffusion Pipeline ...")
controlnet = ControlNetModel.from_pretrained(CONTROLNET_ID, torch_dtype=dtype)
pipe = StableDiffusionControlNetPipeline.from_pretrained(
    SD_MODEL_ID,
    controlnet=controlnet,
    torch_dtype=dtype,
    safety_checker=None,
)
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
pipe = pipe.to(device)

try:
    pipe.enable_xformers_memory_efficient_attention()
except Exception:
    pass


# ---------------------------------------------------------------------------
# Hilfsfunktionen
# ---------------------------------------------------------------------------
def preprocess_image(image: Image.Image, resolution: int = RESOLUTION) -> Image.Image:
    """Logo auf quadratische Zielauflösung bringen."""
    image = image.convert("RGB")
    image = image.resize((resolution, resolution), Image.LANCZOS)
    return image


def get_depth_map(image: Image.Image) -> Image.Image:
    """Depth Map aus dem Logo berechnen (das ist das ControlNet-Kontrollbild)."""
    depth = depth_estimator(image)["depth"]
    depth = np.array(depth)
    depth = depth[:, :, None]
    depth = np.concatenate([depth, depth, depth], axis=2)
    return Image.fromarray(depth)


# ---------------------------------------------------------------------------
# Haupt-Generierungsfunktion (läuft auf der ZeroGPU-Instanz)
# ---------------------------------------------------------------------------
@spaces.GPU(duration=60)
def generate(
    logo_image,
    prompt,
    negative_prompt,
    controlnet_scale,
    num_steps,
    guidance_scale,
    seed,
    progress=gr.Progress(track_tqdm=True),
):
    if logo_image is None:
        raise gr.Error("Bitte zuerst ein Logo-/Referenzbild hochladen.")
    if not prompt or prompt.strip() == "":
        raise gr.Error("Bitte einen Prompt eingeben.")

    logo_image = preprocess_image(logo_image)
    depth_image = get_depth_map(logo_image)

    seed = int(seed)
    if seed < 0:
        generator = None  # zufälliger Seed
    else:
        generator = torch.Generator(device=device).manual_seed(seed)

    result = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        image=depth_image,
        num_inference_steps=int(num_steps),
        guidance_scale=float(guidance_scale),
        controlnet_conditioning_scale=float(controlnet_scale),
        generator=generator,
    ).images[0]

    return result, depth_image


# ---------------------------------------------------------------------------
# Gradio UI
# ---------------------------------------------------------------------------
with gr.Blocks(title="Depth ControlNet Logo Generator") as demo:
    gr.Markdown(
        """
        # 🎨 Depth ControlNet — Logo Generator
        Lade ein bestehendes Logo hoch, gib einen neuen Stil-Prompt ein.
        Die **Tiefenstruktur / Silhouette** des Logos bleibt erhalten,
        während Stil, Farben und Textur komplett neu generiert werden.
        """
    )

    with gr.Row():
        with gr.Column():
            logo_input = gr.Image(
                label="1️⃣ Logo / Referenzbild hochladen",
                type="pil",
                height=300,
            )
            prompt = gr.Textbox(
                label="2️⃣ Prompt",
                placeholder="z.B. vintage japanese emblem logo, ink brush style, minimal, black and red, flat vector, white background",
                lines=3,
            )
            negative_prompt = gr.Textbox(
                label="Negative Prompt",
                value="blurry, low quality, watermark, text, extra elements, photo, 3d render",
                lines=2,
            )

            with gr.Accordion("⚙️ Erweiterte Einstellungen", open=False):
                controlnet_scale = gr.Slider(
                    0.0, 2.0, value=1.0, step=0.05,
                    label="ControlNet Conditioning Scale (Struktur-Treue)",
                )
                num_steps = gr.Slider(10, 50, value=25, step=1, label="Inference Steps")
                guidance_scale = gr.Slider(1.0, 20.0, value=7.5, step=0.5, label="Guidance Scale")
                seed = gr.Slider(-1, 999999, value=-1, step=1, label="Seed (-1 = zufällig)")

            run_button = gr.Button("🚀 Generieren", variant="primary")

        with gr.Column():
            output_image = gr.Image(label="Ergebnis", height=400)
            depth_preview = gr.Image(label="Erkannte Depth Map (Kontrollbild)", height=200)

    run_button.click(
        fn=generate,
        inputs=[
            logo_input,
            prompt,
            negative_prompt,
            controlnet_scale,
            num_steps,
            guidance_scale,
            seed,
        ],
        outputs=[output_image, depth_preview],
    )

    gr.Markdown(
        """
        ---
        💡 **Tipps:**
        - Höherer *ControlNet Conditioning Scale* = Form/Struktur des Original-Logos wird strenger befolgt.
        - Niedrigerer Wert = mehr kreative Freiheit, weniger Ähnlichkeit zur Vorlage.
        - Für saubere Vektor-Optik: Begriffe wie "flat vector", "clean lines", "white background" im Prompt verwenden.
        """
    )

demo.queue(max_size=20)
demo.launch()