Research map / Image generation / Text-to-image diffusion

TKG-DM: Training-free Chroma Key Content Generation Diffusion Model

CVPR 2025 · Image generation › Text-to-image diffusion · approach: Text-driven synthesis and semantic guidance · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TKG-DM, a training-free method that shifts channel means of initial diffusion noise (blended via a Gaussian mask) so Stable Diffusion produces foreground subjects over a uniform chroma key background.

Problem

Text-to-image diffusion models like Stable Diffusion struggle to produce foreground objects over a specifiable chroma key background, so separating foreground and background usually needs fine-tuning or large datasets.

Approach

Manipulate the color statistics of the initial Gaussian noise (channel mean shift keeping std fixed) to induce a uniform background color, then blend shifted and normal noise with a Gaussian mask so the foreground follows the text prompt, with no training.

Key results

• On SD1.5, TKG-DM reaches FID 56.32 and m-FID 40.75 vs 85.00 and 63.54 for SD1.5 with GBP, the best training-free baseline.
• On SDXL, LayerDiffuse gets the best FID (29.34) while TKG-DM (41.81 FID, 31.43 m-FID) stays competitive with it and beats SDXL with GBP (45.32, 39.17) on FID and m-FID.
• In a 100-participant user study, preference was roughly even against LayerDiffuse (about 50/50) and favored TKG-DM 53.9% vs SDXL with GBP on image quality/prompt alignment.

Measured effects reported in the paper

Surprising findings

• LayerDiffuse achieves the best SDXL FID, which the authors attribute to its evaluation setup overlaying foregrounds on the same lime green background as the ground truth.

Limitations

• Primarily generates foreground content; cannot generate landscape or background-focused content.
• Fails to generate the foreground when the requested object size (Gaussian sigma) is too small or the prompt lacks object information.
• Evaluation relies on a subset of MAGICK and on ground truth built by overlaying alpha images, which can favor some baselines.

Experimental setup

Datasets: MAGICK dataset (3,000 images subset)

Benchmarks: MAGICK dataset subset (3,000 images)

Metrics: FID m-FID CLIP-I CLIP-S user preference

Baselines: SD1.5 with GBP SDXL with GBP DeepFloyd with GBP GreenBack LoRA LayerDiffuse ControlNet with GBP

Builds on: Stable Diffusion ControlNet AnimateDiff MAGICK

Closest papers by meaning

• RealCompo: Balancing Realism and Compositionality Improves Text-to-Image Diffusion Models (NeurIPS 2024 · similarity 0.86)
• CtrLoRA: An Extensible and Efficient Framework for Controllable Image Generation (ICLR 2025 · similarity 0.86)
• Rethinking Training for De-biasing Text-to-Image Generation: Unlocking the Potential of Stable Diffusion (CVPR 2025 · similarity 0.86)
• DiffuseHigh: Training-Free Progressive High-Resolution Image Synthesis Through Structure Guidance (AAAI 2025 · similarity 0.85)
• Efficient Video Diffusion Models via Content-Frame Motion-Latent Decomposition (ICLR 2024 · similarity 0.85)
• Spectral Progressive Diffusion for Efficient Image and Video Generation (NeurIPS 2026 · similarity 0.85)