Two new Replicate surfaces illustrate a practical pattern in visual systems: watermarking and Flux fine-tuning are exposed as separate callable operations. One protects or labels an image artifact; the other creates a specialized generative model. Neither needs to be hidden inside one large “visual AI” endpoint.
That separation makes the workflow easier to reason about. Input validation, deterministic transforms, training, generation, and output checks can each have their own cost, retry policy, and artifact receipt. A failed watermark pass should not force an expensive model run to repeat, and a training artifact should remain addressable independently of later generations.
The earlier basket also attached an Amazon semantic-segmentation paper from 2018 as if it were current news. It is useful background, but it is not part of this fresh signal and is intentionally excluded here. The supported update is narrower: hosted visual tooling is becoming more composable.