# FrameForge: Harnessing Priors of Video Foundation Models for Unified Image Generation

**Authors:** Yijing Lin, Zhendong Mao, Mengqi Huang, Yuxin Wang, Yongdong Zhang  
**Lab:** FrameX-AI · **Published:** 26 Oct 2026 · **Upvotes:** 112 · **Level:** Intermediate  
**Links:** [arXiv](https://arxiv.org/abs/2610.36545) · [PDF](https://arxiv.org/pdf/2610.36545) · [Code](https://github.com/Lyne1/FrameForge) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2610.36545.html)

## Curator's notes

**TL;DR:** FrameForge uses adaptive virtual temporal relations from video models to improve unified conditional image generation.

- **What's new:** It introduces Mixture of Temporal Experts and Chain of Temporal Reasoning to dynamically model condition-target distances and intermediate transformations.
- **Key result:** FrameForge achieves superior performance on DreamBench, COCO2017, and GEdit-Bench across subject-driven, canny-to-image, and image editing tasks.
- **Why it matters:** Builders can leverage video priors to handle diverse image generation conditions with a single model.

## Abstract

Unified conditional image generation requires a single model to preserve different aspects of a visual input while producing a target image that follows textual guidance. Pretrained video models provide a promising foundation by encoding visual transformations across time, yet existing approaches typically assign the condition and target to fixed temporal positions. We observe that condition-target relationships vary substantially across generation tasks, while video models naturally encode visual transformations at different temporal scales. This motivates us to model conditional generation through input-adaptive virtual temporal relations. We propose FrameForge, which captures two complementary aspects of such relations. Mixture of Temporal Experts (MoTE) models how far the target lies from the condition by dynamically selecting temporal RoPE experts with different virtual intervals. Chain of Temporal Reasoning (CoTR) models how the condition is transformed into the target through a latent intermediate state, with a reasoning-bottleneck attention mask encouraging information propagation through this pathway. Experiments on DreamBench, COCO2017, and GEdit-Bench show that FrameForge achieves superior performance in subject-driven generation, canny-to-image generation, and image editing among the compared methods.

## What you could build

1. Create a text‑guided photo editor that edits images via latent intermediate states
2. Develop a sketch‑to‑photo generator using adaptive temporal intervals
3. Build a subject‑driven image synthesis tool that swaps objects based on textual prompts

## Quiz yourself

How does FrameForge model the distance between condition and target images?

**Answer:** It uses a Mixture of Temporal Experts that dynamically selects temporal RoPE experts with different virtual intervals.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.