FlowLess enables image generation of different concepts from a provided visual abstraction set. We provide further fine-grained control over the abstractness level of the generated image (top), and over the composition of the visual elements using additional provided visual details (bottom).
While large-scale text-to-image models excel at generating high-fidelity content from semantic descriptions, users often lack precise control over the of non-semantic abstract elements and the models expressive freedom. Existing structural conditioning methods typically impose rigid spatial layouts or global style constraints, failing to address the need for fine grained creative control via mixture of visual primitives. We introduce a novel self-supervised framework that enables granular control over image generation through a visual abstraction set—a collection of disjoint shapes, textures and composition details. By fine-tuning a pre-trained flow model, we allow users to construct diverse semantic concepts from the same underlying visual kit. Our method offers control over the level of abstraction: by leveraging a training strategy based on varying geometric augmentations and per-part modulation, users can dictate whether the model should preserve input shapes or interpret them loosely to fit the narrative. Furthermore, we enable control over the generation’s expressiveness, allowing for the creation of minimal compositions. We demonstrate that our method provides a richer, more flexible paradigm for creative design compared to state-of-the-art baselines across diverse styles and compositions.
Our method enables per part abstraction control. On top, we specify the preservation of a single shape (yellow outline) while other elements (green outline) are combined freely into the generated image. On bottom we set high fidelity with respect to all four elements.
Our method can complete input drawings (red boxes) to generate different concepts using additional (optional) abstract elements.
Using different visual details (left) with a fixed set of base elements (top) enables further creative extrapolation as users can define specific inter-relations between curves, shapes, or materials to influence the overall look of the generated images.
By increasing the coverage score, from left to right, we reduce he number of elements in the generated and images, constraining the model to use only the abstraction set input (left).