RegalRinse/roomviz-wall-editor

🏠 RoomViz Wall Editor

RoomViz Wall Editor is a domain-adapted image editing model built on top of Qwen/Qwen-Image-Edit-2511, configured and optimized for transforming under-construction Indian room images into fully finished wall visualizations with user-specified colors.

---

🎯 Use Case

This model is designed for a specific real-world application:

• Input: A photograph of an under-construction or rough Indian room (exposed brick, unplastered walls, raw concrete)
• Instruction: A natural language color instruction from the user
  (e.g. "paint the walls light beige")
• Output: The same room rendered with smooth, fully plastered, and painted walls in the desired color — with floors, ceilings, doors and windows preserved exactly

---

🔧 Intended Prompt Format

Make the walls smooth, fully plastered and painted [COLOR].
Keep the floor, ceiling, doors, windows and furniture exactly as they are.

Example prompts:

• Make the walls smooth, fully plastered and painted off-white. Keep the floor, ceiling, doors, windows and furniture exactly as they are.
• Make the walls smooth, fully plastered and painted sage green. Keep the floor, ceiling, doors, windows and furniture exactly as they are.
• Make the walls smooth, fully plastered and painted warm beige. Keep the floor, ceiling, doors, windows and furniture exactly as they are.

---

🚀 How to Use

from diffusers import FlowMatchEulerDiscreteScheduler
from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel
import torch
from PIL import Image

dtype = torch.bfloat16
device = "cuda"

pipe = QwenImageEditPlusPipeline.from_pretrained(
    "RegalRinse/roomviz-wall-editor",
    transformer=QwenImageTransformer2DModel.from_pretrained(
        "RegalRinse/roomviz-transformer",
        torch_dtype=dtype,
        device_map="cuda"
    ),
    torch_dtype=dtype
).to(device)

image = Image.open("your_room.jpg").convert("RGB")

result = pipe(
    image=[image],
    prompt="Make the walls smooth, fully plastered and painted light beige. Keep the floor, ceiling, doors, windows and furniture exactly as they are.",
    negative_prompt="worst quality, low quality, blurry, unrealistic",
    height=1024,
    width=1024,
    num_inference_steps=4,
    true_cfg_scale=3.5,
).images

result.save("output.png")

---

⚙️ Technical Details

Property	Value
Base Model	Qwen/Qwen-Image-Edit-2511
Transformer	RegalRinse/roomviz-transformer (Lightning-distilled, 4-step)
Inference Steps	4 (Lightning)
Recommended CFG Scale	3.5
Input Resolution	Up to 1024×1024
Dtype	bfloat16
Task	Text-guided image-to-image editing
Domain	Indian under-construction rooms

---

🏗️ Architecture

This model uses the QwenImageEditPlus pipeline with a Lightning-distilled transformer that enables high-quality image editing in just 4 inference steps instead of the standard 50, making it suitable for real-time deployment on shared GPU infrastructure.

---

📦 Deployment

This model is deployed as part of the RoomViz application stack:

• Model Hosting: Hugging Face Model Hub
• Inference: Hugging Face ZeroGPU Space (free H200)
• API Layer: FastAPI on Render
• Frontend: React on Vercel

---

⚠️ Limitations

• Works best on images where walls are the dominant surface
• May not perfectly isolate walls if room has complex geometry
• Output quality depends on input image clarity and lighting
• Optimized for Indian construction room aesthetics