Krea 2 Raw + Turbo ZeroGPU demo

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+diffusers-0.39.0.dev0-py3-none-any.whl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,13 +1,14 @@
 ---
 title: Krea 2
-emoji: 📈
-colorFrom: yellow
-colorTo: red
+emoji: 🎨
+colorFrom: indigo
+colorTo: purple
 sdk: gradio
-sdk_version: 6.19.0
-python_version: '3.12'
+sdk_version: 6.15.1
 app_file: app.py
-pinned: false
+short_description: Krea 2 text-to-image (Raw + Turbo)
+python_version: "3.12"
+startup_duration_timeout: 1h
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Private demo of Krea 2 (Raw and Turbo distilled) text-to-image, both loaded on ZeroGPU.

app.py

@@ -0,0 +1,107 @@
+import os
+import random
+
+import spaces
+import torch
+import gradio as gr
+from huggingface_hub import login
+
+if os.environ.get("HF_TOKEN"):
+    login(token=os.environ["HF_TOKEN"])
+
+from diffusers import Krea2Pipeline
+
+DTYPE = torch.bfloat16
+RAW_REPO = "krea/Krea-2-Raw"
+TURBO_REPO = "krea/Krea-2-Turbo"
+MAX_SEED = 2**31 - 1
+
+# Both models are loaded at global scope. They share the architecture, so Turbo
+# reuses Raw's text encoder / tokenizer / VAE and only loads its own transformer.
+pipe_raw = Krea2Pipeline.from_pretrained(RAW_REPO, torch_dtype=DTYPE)
+pipe_turbo = Krea2Pipeline.from_pretrained(
+    TURBO_REPO,
+    text_encoder=pipe_raw.text_encoder,
+    tokenizer=pipe_raw.tokenizer,
+    vae=pipe_raw.vae,
+    torch_dtype=DTYPE,
+)
+pipe_raw.to("cuda")
+pipe_turbo.to("cuda")
+
+PIPES = {"Raw": pipe_raw, "Turbo": pipe_turbo}
+DEFAULTS = {
+    "Raw": {"steps": 28, "guidance": 4.5},
+    "Turbo": {"steps": 8, "guidance": 0.0},
+}
+
+
+def _duration(prompt, negative_prompt, model, steps, guidance, width, height, seed, randomize, progress=None):
+    return int(steps) * 2 + 25
+
+
+@spaces.GPU(duration=_duration, size="xlarge")
+def generate(
+    prompt,
+    negative_prompt,
+    model,
+    steps,
+    guidance,
+    width,
+    height,
+    seed,
+    randomize,
+    progress=gr.Progress(track_tqdm=True),
+):
+    if randomize:
+        seed = random.randint(0, MAX_SEED)
+    seed = int(seed)
+    generator = torch.Generator("cuda").manual_seed(seed)
+    pipe = PIPES[model]
+    image = pipe(
+        prompt=prompt,
+        negative_prompt=(negative_prompt or None) if guidance > 0 else None,
+        height=int(height),
+        width=int(width),
+        num_inference_steps=int(steps),
+        guidance_scale=float(guidance),
+        generator=generator,
+    ).images[0]
+    return image, seed
+
+
+def on_model_change(model):
+    d = DEFAULTS[model]
+    return (
+        gr.update(value=d["steps"]),
+        gr.update(value=d["guidance"]),
+        gr.update(interactive=d["guidance"] > 0),
+    )
+
+
+with gr.Blocks() as demo:
+    gr.Markdown("# Krea 2\nText-to-image with **Krea 2 Raw** (CFG) and **Krea 2 Turbo** (distilled, few-step).")
+    with gr.Row():
+        with gr.Column():
+            model = gr.Radio(["Raw", "Turbo"], value="Raw", label="Model")
+            prompt = gr.Textbox(label="Prompt", lines=3, placeholder="a fox in the snow")
+            negative_prompt = gr.Textbox(label="Negative prompt", lines=1)
+            run = gr.Button("Generate", variant="primary")
+            with gr.Accordion("Advanced", open=False):
+                steps = gr.Slider(1, 50, value=28, step=1, label="Steps")
+                guidance = gr.Slider(0.0, 10.0, value=4.5, step=0.1, label="Guidance scale")
+                with gr.Row():
+                    width = gr.Slider(512, 1536, value=1024, step=16, label="Width")
+                    height = gr.Slider(512, 1536, value=1024, step=16, label="Height")
+                with gr.Row():
+                    seed = gr.Slider(0, MAX_SEED, value=0, step=1, label="Seed")
+                    randomize = gr.Checkbox(value=True, label="Randomize seed")
+        with gr.Column():
+            output = gr.Image(label="Result", format="png")
+
+    model.change(on_model_change, model, [steps, guidance, negative_prompt])
+    inputs = [prompt, negative_prompt, model, steps, guidance, width, height, seed, randomize]
+    run.click(generate, inputs, [output, seed])
+    prompt.submit(generate, inputs, [output, seed])
+
+demo.launch()

diffusers-0.39.0.dev0-py3-none-any.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a0593cfc51465e7414e1dc2c4c6acdf14fadc4e346d49a19d6116e2d50adf856
+size 5594675

requirements.txt

@@ -0,0 +1,4 @@
+./diffusers-0.39.0.dev0-py3-none-any.whl
+transformers>=4.57.0
+accelerate
+sentencepiece