SGE-Goal
A spatially grounded decomposed goal-image editing framework that first predicts initial and final 2D edit-region masks and then generates the goal image based on those masks, used to improve the control success rate of robotic manipulation policies.
Timeline (last 90 days)
The best closed-source editor achieves a success rate on the longest five-step chain of only about one-seventh that of the SGE-Goal fine-tuned model.
SGE-Goal is a spatially grounded decomposed editing framework that splits target image synthesis into two steps: first predicting the initial and final 2D edit region masks, then generating the target image based on the masks.
SGE-Goal, fine-tuned on FLUX.1 Kontext and FLUX.2 klein, ranks first on CALVIN.
SGE-Goal, fine-tuned on FLUX.1 Kontext and FLUX.2 klein, ranks second only to Gemini 3 Pro Image on SIMPLER-Bridge.
The paper releases code, datasets, checkpoints, and a human-aligned diagnostic evaluator.
The decomposed mask method decouples 'where to change' from 'what to change into,' with mask prediction responsible for spatial grounding and the generative model responsible for appearance synthesis, reducing the risk of hallucination or tampering in irrelevant regions.
The limitation of SGE-Goal lies in its dependence on the base model's editing capability, and the decomposed framework introduces an additional source of mask prediction error.
All Facts (7)
SGE-Goal, fine-tuned on FLUX.1 Kontext and FLUX.2 klein, ranks second only to Gemini 3 Pro Image on SIMPLER-Bridge.
50%UnverifiedSGE-Goal is a spatially grounded decomposed editing framework that splits target image synthesis into two steps: first predicting the initial and final 2D edit region masks, then generating the target image based on the masks.
50%UnverifiedSGE-Goal, fine-tuned on FLUX.1 Kontext and FLUX.2 klein, ranks first on CALVIN.
50%UnverifiedThe best closed-source editor achieves a success rate on the longest five-step chain of only about one-seventh that of the SGE-Goal fine-tuned model.
50%UnverifiedThe limitation of SGE-Goal lies in its dependence on the base model's editing capability, and the decomposed framework introduces an additional source of mask prediction error.
50%UnverifiedThe decomposed mask method decouples 'where to change' from 'what to change into,' with mask prediction responsible for spatial grounding and the generative model responsible for appearance synthesis, reducing the risk of hallucination or tampering in irrelevant regions.
50%UnverifiedThe paper releases code, datasets, checkpoints, and a human-aligned diagnostic evaluator.
50%