Supporting Page — Interactive Web Studio

Free Live Mage-Flow Studio

Experience real-time 4-step distilled image generation and instruction editing directly inside your web browser. Powered by Hugging Face Spaces and Microsoft's open 4B model checkpoints.

Live Interactive Mage-Flow Studio

Powered by Hugging Face Cloud GPUs & Microsoft Open Weights

100% Free Access
If the embedded interface takes a few seconds to load, the cloud GPU container may be warming up.

Interactive Playground User Guide

The Mage-Flow web playground connects your browser directly to cloud GPU hardware running Microsoft's official microsoft/Mage-Flow-Turbo model weights. Follow these guidelines to achieve high quality outputs during live experiments:

1. Selecting Generation Modes

The studio supports two distinct operation modes accessible via the top tab navigation:

  • Text-to-Image Generation: Enter descriptive text prompts to synthesize new images from gaussian noise. Select target resolutions (1024x1024, 1280x720, 720x1280, or 2048x512).
  • Instruction-Based Image Editing: Upload an existing reference image and provide action instructions (e.g. "Make the sky dark with shining stars"). The model modifies target areas without altering background geometry.

2. Prompt Parameter Settings

ParameterRecommended RangeFunction & Effect
Inference Steps4 Steps (Turbo) / 20-30 Steps (Base)Number of denoising evaluation iterations. 4 steps is optimal for Turbo models.
Guidance Scale (CFG)1.5 to 3.5Controls prompt adherence strength. Lower values produce softer, natural blending.
Aspect Ratio1:1, 16:9, 9:16, 4:1Selects native positional encoding grid without cropping target compositions.
Random SeedNumeric Integer (-1 for Random)Locks random noise generation for exact composition reproducibility across edits.

3. Prompt Presets for Fast Testing

Need inspiration? Try these pre-tested prompt templates inside the playground:

Photorealistic Architecture
"Modern glass villa cantilevered over a misty pine forest cliff, golden hour sunlight reflecting off floor-to-ceiling windows, 35mm architectural photography."
Macro Character Concept
"Detailed macro photograph of a mechanical hummingbird made of polished brass and emerald glass, resting on a wet lotus flower, soft studio bokeh."
Ultra-Wide 4:1 Panorama
"Wide panoramic horizon view of an ancient desert valley with towering carved stone monoliths under a starry galaxy, twilight sky, 4:1 aspect ratio."
Sci-Fi Vehicle Studio Render
"Studio product photograph of an autonomous electric rover with matte obsidian finish, copper accents, dramatic studio rim lighting, 8k resolution."

4. Step-by-Step Interactive Workflow

  1. Select Task Tab: Choose between Text-to-Image for new synthesis or Image Editing for existing image modification.
  2. Enter Prompt Description: Type a natural language description detailing primary subjects, environment lighting, and style.
  3. Adjust Dimension Settings: Pick target aspect ratios matching your project needs (square, widescreen, story vertical, or ultra-wide panorama).
  4. Click Submit: Trigger sub-second execution on cloud GPU hardware. Download generated PNG images directly to your local machine.

5. Cloud GPU Hosting Infrastructure

The live playground is hosted on public GPU infrastructure provided by Hugging Face Spaces. By deploying microsoft/Mage-Flow-Turbo within a containerized Gradio interface, requests are processed efficiently without queuing bottlenecks.

Infrastructure LayerProvider / TechnologySpecification / Benefit
Model Weights HostHugging Face HubOfficial microsoft/Mage-Flow-Turbo weights repository
Web Runtime EngineGradio Web SDK (v6.15)Lightweight web components embedded via custom HTML elements
Hardware AcceleratorNVIDIA A10G / A100 Cloud GPUSub-second tensor execution with FP16 precision
Cost to End User100% Free Open AccessNo paywalls, registration fees, or credit subtractions

6. Troubleshooting Playground Latency & Errors

If the embedded Gradio studio experiences temporary connection delays:

  • Container Cold Starts: If the Space has been idle, allow 10 to 15 seconds for cloud GPU weights to load into VRAM.
  • Queue Wait Times: During high community traffic, generation requests queue automatically and execute sequentially within seconds.
  • Local Execution Alternative: For heavy batch production, refer to our Developer SDK Guide to run the model locally on your own GPU hardware.

7. In-Depth Prompt Engineering Examples Table

CategoryTarget Aspect RatioSample Prompt InputExpected Composition
Portrait Photography9:16 Vertical"Cinematic portrait of a veteran craftsman in a wood workshop, warm lantern light, 85mm lens"Shallow depth of field with sharp facial details and soft wooden background blur
Landscape Architecture16:9 Widescreen"Futuristic eco-city with vertical gardens and glass skybridges at dawn, soft morning sunlight"Expansive urban perspective with architectural details and atmospheric fog
Product Concept1:1 Square"Minimalist ceramic teapot on a dark slate tabletop, dramatic side rim lighting, macro focus"Clean studio product framing with rich surface texture contrast
Panoramic Environment4:1 Panorama"Ultra wide horizon of an alien crystal desert under a twin moon night sky, 4:1 aspect ratio"Continuous horizontal landscape horizon across wide panoramic canvas

8. Instruction Editing Workflow Guidelines

When using the instruction editing tab inside the playground:

  • Upload Clear Source Images: Provide high-resolution reference images (1024x1024 or 1280x720) with distinct subject boundaries.
  • Use Action-Oriented Text: Phrase commands directly: "Change shirt color to bright red" or "Add a coffee cup on the desk".
  • Avoid Vague Phrases: Avoid ambiguous prompts like "make it look cooler"; specify exact target changes (materials, colors, objects, lighting).

9. Local Web UI Deployment Alternative

If you prefer hosting a private web studio on your local workstation without external cloud dependencies, launch the local Gradio interface using Python:

# Launch local Gradio web studio interface
from diffusers import MageFlowPipeline
import gradio as gr
import torch

pipe = MageFlowPipeline.from_pretrained("microsoft/Mage-Flow-Turbo", torch_dtype=torch.bfloat16).to("cuda")

def generate_fn(prompt, steps, cfg):
    return pipe(prompt=prompt, num_inference_steps=steps, guidance_scale=cfg).images[0]

interface = gr.Interface(
    fn=generate_fn,
    inputs=[gr.Textbox(label="Prompt"), gr.Slider(1, 10, value=4), gr.Slider(1.0, 5.0, value=1.5)],
    outputs=gr.Image(label="Output Image"),
    title="Local Private Mage-Flow Studio"
)
interface.launch(server_name="127.0.0.1", server_port=7860)

10. Summary of Live Web Playground Benefits

The MageFlow AI interactive studio provides instant, zero-cost web experimentation for Microsoft's 4B generative foundation models, allowing researchers and creators to test prompts, aspect ratios, and instruction edits without commercial paywalls. From testing quick ideas via 4-step Turbo sampling to building private local Gradio interfaces, our platform offers an unconstrained environment for creative exploration.

Compare Quantitative Metrics

See how Mage-Flow compares against 20B and 32B models in quantitative benchmarks.