TMP: Tree-structured Mixed-policy Pruning for Large-scale Image Generation and Editing

A new pruning framework called TMP can compress large-scale image generation models by up to 75 percent, enabling an 80-billion-parameter model to run on a single consumer GPU, according to a preprint posted to arXiv on June 25, 2026 [1]. The framework, named Tree-structured Mixed-policy Pruning, targets the growing computational cost of high-fidelity image synthesis models. The authors note that modern image generation models have rapidly expanded in size, making them “unaffordable for their enormous parameter consumption and computation budget” [1]. TMP is designed to generalize across two prevalent image tasks — text-to-image generation and text-guided image editing — and two major architectures: Mixture-of-Experts and Diffusion transformer [1]. The most dramatic result reported involves HunyuanImage-3.0, an open-source state-of-the-art model. TMP reduced its parameter count from 80 billion to 20 billion, a 75 percent reduction, while “sacrificing limited generation quality” [1]. Through additional engineering, the pruned 20-billion-parameter version was made to run inference on a single NVIDIA RTX 4090 GPU with 24 GB of memory [1]. The inference script and model weights have been integrated into the existing HunyuanImage-3.0 GitHub and Hugging Face repositories [1]. A second experiment applied TMP to Z-Image turbo, a smaller efficient model. The framework compressed it from 6 billion to 4 billion parameters, a 33 percent reduction, with what the authors describe as “negligible degradation” [1]. The paper states that the framework is also compatible with step-distilled models and can serve as a final-stage compression step [1]. The work lands as the broader research community continues to pursue methods for shrinking diffusion models without destroying output quality. A separate 2025 preprint introduced TreeQ, a quantization framework for Diffusion transformers that achieved a Fréchet Inception Distance score of 6.92 at 4-bit precision, close to the full-precision baseline [5]. That effort used a tree-structured search to navigate mixed-precision configurations efficiently, a conceptual parallel to the structured merging strategy in TMP [5]. The TMP paper was submitted to arXiv’s computer-vision and pattern-recognition section. arXiv, founded in 1991, now hosts over two million e-prints and receives roughly 24,000 new submissions per month, serving as the primary distribution channel for pre-publication research in machine learning and related fields [9].