README.md

metadata

title: Color Grade Transfer
emoji: 😻
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: 6.17.3
python_version: '3.12'
app_file: app.py
pinned: true
tags:
  - backyard-ai
  - an-adventure-in-thousand-token-wood
  - well-tuned
  - gradio
short_description: Apply the color grade from one image to another image

🌓 Color Grade Transfer

Transfer color grade directly from reference image to source image without manual color grading.

📖 Overview

It transfers the color grading from a reference image directly onto a source image.

🏆 Hackathon Patches

Patch	Status
🔬 Small Models Only — Qwen-Image-Edit-2511 at 20B parameters, well under the 32B limit	✅
🔌 Off the Grid — No cloud APIs, runs entirely on local GPU	✅
🎯 Well-Tuned — Custom LoRA ovi054/QIE-2511-Color-Grade-Transfer-LoRA published on Hub	✅
📓 Field Notes — Blog post on what I built and learned	✅

Key Features

• Direct Color Transfer: Copies styling from a reference image to a source image without manual grading.
• 4-Step Inference: Uses a lightning adapter for fast, low-step generation.
• Aspect Ratio Alignment: Uses an internal VAE dimension calculation 1MP to ensure the 2-D RoPE coordinate systems match, eliminating structural pixel shifts during inference.
• Interactive UI: Features a side-by-side image comparison slider.
• MCP Compliant: Work as a backend tool for Model Context Protocol clients.

---

🧠 LoRA Fine-Tuning Process

Curating a clean dataset for custom style mapping is often the bottleneck of training. To overcome this, a systematic data-generation technique was engineered to produce high-fidelity, bidirectional training pairs.

The Dataset Recipe

1. Palette Extraction: Two entirely different content images were chosen.
2. Homogeneous Grading: The exact same color palette was applied to both images so they shared an identical color profile.
3. Cross-Pair Mapping: * The newly graded version of Image B was assigned as the color-style reference for Image A.
   • The inputs were then flipped, mapping the graded version of Image A as the reference for Image B.
4. This instantly generated two highly consistent training pairs per asset set, allowing the dataset to scale efficiently without content bleeding or style loss.

Versatility & Flexibility

Because the dataset decoupling separated semantic structure from global grading parameters, the fine-tuned LoRA natively generalizes across all input configurations:

• 👤 ↔️ 👤 Character to Character
• 🏞️ ↔️ 🏞️ Scene to Scene
• 👤 ↔️ 🏞️ Character to Scene / Scene to Character

⚙️ Training Hyperparameters

The model was fine-tuned using the following precise configuration settings:

Configuration Category	Parameter	Value / Setting
Saving Settings	Save Precision	bf16
Learning Rate	Learning Rate	0.0001
	Optimizer	AdamW
Dataset Settings	Base Resolution	1024 * 1024
	Enable Bucket	True (open)
	Min Bucket Reso	128
	Max Bucket Reso	8192
	Bucket Reso Steps	64
Network Settings	Network Rank Dim (Rank)	16
	Network Alpha	16

---

⚙️ Tech Stack

• Base Model: Qwen/Qwen-Image-Edit-2511
• LoRA Adapters:
  • ovi054/QIE-2511-Color-Grade-Transfer-LoRA (Color grade transfer fine-tune)
  • lightx2v/Qwen-Image-Edit-2511-Lightning (4-step inference)
• Frontend: Gradio

---

🔗 Project Links

Live Demo

• Hugging Face Space: Color Grade Transfer

Blog

• Blog Post: Color Grade Transfer — Instant Cinematic Tone Matching

Social Media & Demo

• Twitter/X: https://x.com/ovi054/status/2064060651706925403
• Hugging Face Post: https://huggingface.co/posts/ovi054/229194823175599
• Youtube Video: https://youtu.be/DqlLDDgV3kQ

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference