Scaling behavior (Text-to-Image). As model size increases (290M → 420M → 625M → 1B parameters), the performance gap between our method and REPA widens. Our method effectively leverages increased compute, while REPA shows diminishing returns.
Multi-modal experiments. (a) We train a single model on three modalities with different weightings to control the trade-off between them. Self-Flow provides consistent improvements (shaded area) across all settings. Axes are inverted so that larger area indicates better performance. (b) Success rates for joint Video-Action prediction. Early on (30k), Self-Flow outperforms flow matching (FM) across all tasks and achieves success in all task categories, whereas FM fails entirely on Open and Place tasks. Later (100k), performance on single-object tasks (Pick Coke Can, Open/Close Drawer) converges, while Self-Flow maintains a significant advantage on complex multi-object and sequential tasks (Move Near, Open and Place).
All models use a 625M parameter backbone trained on 20M images. We benchmark against vanilla flow matching and leading approaches for improving diffusion representations. For external alignment methods, we compare against REPA with DINOv2 and REPA with SigLIP2. For methods without external models, we compare against SRA. Our method achieves superior visual quality and prompt adherence across diverse prompts.
"a parked blue Vespa scooter on a gravel path beside a tranquil body of water, with tall grasses, captured in a realistic photographic style."
"a professionally portrait of a woman with medium-length dark hair with bangs, a large white decorative bow with silver embellishments on the side of her head."
"a close-up, high-resolution photograph of a wheel and brake system, featuring a polished silver alloy wheel with a blue brake."
"a close-up of a monkey with distinctive facial markings, partially closed eyes, and thick fur, set against a blurred green forest background."
"a tractor, photographed in a vibrant red color, parked on a dirt path beside a grassy field under a partly cloudy sky, highlighting its vintage agricultural design."
"a trapper hat, featuring a red and black plaid pattern with a thick fur around the brim and ear flaps, displayed on a mannequin head against a light background."
"an elephant standing on a dirt path in a natural, forested area with its trunk extended downward, captured in high-definition photography."
"a softly lit portrait of a man with short hair and a beard, wearing a dark shirt with a plaid collar, against a light background with blurred circular elements."
"a stylized, cartoon figurine of a Shiba Inu dog, depicted in a friendly pose while holding a bowl of food, standing on a wooden base."
"a hand-painted, wooden fish-shaped decoration with horizontal white stripes, resting on a white shelf near a window."
"a bento box featuring a compartment of white rice decorated with nori into a cartoonish face, by another compartment containing grilled meat and vegetables."
"a donkey in a grassy, fenced enclosure stands close to the metal fence. The scene is captured in natural daylight."
"a close-up, well-lit portrait photograph of a man with light brown hair and a beard, wearing a blue and red plaid shirt, against a plain, light-colored background."
"a close-up, soft-lit photograph of a pug lying on its back on a red textured surface, with its large, open eyes looking directly at the camera."
"a midcentury-style terracotta sculpture of a hippopotamus, intricately detailed, seated with a decorative headdress against a plain white background."
"a close-up of the front of a car, highlighting its glossy black paint, chrome accents,, set against a snowy outdoor background."
"a donkey standing inside a stable, wearing a blue halter and a red quilted protective blanket over its back, with wooden walls visible in the background."
"a young boy tenderly hugging a white stuffed dog toy while standing indoors against a plain wall."
"a detailed, classical-style sculpted bust of a man in a serious expression, wearing a toga, encased in an oval frame and mounted on a textured, earthy-toned wall."
"a person wearing a black face mask with a white pattern, standing against a green and white gradient background, captured in a softly lit, balanced composition."
"a pair of pink sneakers, worn by someone standing on a concrete surface in an urban environment, close-up, low-angle photograph."
"a professionally lit portrait photograph of a woman with brown hair wearing a trench coat over a white shirt captured outdoors in natural light."
"a highly detailed and colorful traditional Japanese Beckoning Cat, depicted in a stylized manner with intricate patterns and vibrant colors."
"a photograph of a small, white church with blue trim and a prominent blue cross on its facade, featuring a tall steeple and situated on an elevated plot."
"a photograph of plush toys—a white teddy bear and a green rabbit—arranged together against a pink, patterned background."
"two white reindeer inside a wire-fenced enclosure, one facing the camera and the other turned to the side; the background includes a wooden structure."
"a photograph taken from a low angle of an, highlighting its sleek, modern design with metallic railings and overhead structural elements under bright lighting."
"a black cat in a blue jacket with red and white accents, wearing a tag, held by a visible arm, looking upwards with its mouth slightly open against a blue door."
"a man posing indoors with a shallow depth of field, wearing a dark blue shirt and sporting a friendly smile."
All models use a 625M parameter backbone trained on just 6M videos. We compare against vanilla flow matching, REPA with DINOv2 for external alignment, and SRA as the leading method without external models. Interestingly, DINOv2 remains the strongest external encoder for video generation, outperforming video-specific encoders such as V-JEPA 2 and advanced spatial learners such as Depth Anything 3 (see Sec. 4). Our method achieves superior results across all baselines.
All models use a 625M parameter backbone trained on the FMA music dataset. We compare against vanilla flow matching, REPA with MERT for external alignment, and SRA as the leading method without external models. Consistent with our findings on video, external alignment with MERT provides no benefit over vanilla flow matching on audio generation (see Sec. 4), demonstrating that external alignment fails to generalize beyond image-centric tasks. Our method achieves superior results without relying on any external representations.
Looking ahead, by bridging representation learning and generative modeling, our approach offers a path toward world models that harness the scalability and perceptual grounding of visual generative models without sacrificing the semantic abstraction required for planning and understanding. We present results obtained by fine-tuning our video-weighted multi-modal runs (with 675M parameters) for action predictions. We employ the SIMPLER simulator to evaluate the action predictions.
