Генеративные модели теперь можно обучать сквозным способом: секрет — цикл for

Researchers, whose work is analyzed by Jiqizhixin, show that generative models can be trained end-to-end — using the same principle that revolutionized computer vision in 2012. The key mechanism turned out to be surprisingly simple — an ordinary `for` loop.

What AlexNet proved in 2012

Before AlexNet, computer vision tasks were solved through manual multi-step pipelines: researchers manually designed edge, texture, and shape detectors, connecting them in a sequential chain. In 2012, AlexNet handed the entire process over to a neural network, training it on the task as a whole, and crushed competitors at the ImageNet competition by a margin of more than 10 percentage points.

The principle of end-to-end training has since prevailed in every new computer vision task:

• 2012 — image classification: AlexNet outperforms manual methods by more than 10%
• 2013–2015 — object detection: R-CNN and YOLO replace composite pipelines
• 2015–2017 — semantic segmentation: FCN and U-Net are trained end-to-end

Each time, a single trainable system outperformed the manual pipeline. The logic is the same: instead of optimizing each stage separately, it is better to let the model find a solution across the entire task as a whole.

Why generative models remained the exception

Modern image generators — diffusion models — produce results iteratively: over tens or hundreds of steps, they transform random noise into an image. Autoregressive models work differently, but also step by step — token by token.

Passing gradients through such an iterative process was either computationally prohibitive or required special architectural solutions. As a result, the components of generative systems were traditionally trained separately and combined manually at inference time — precisely the scheme that AlexNet proved suboptimal for recognition tasks.

The problem is fundamental: each diffusion step depends on the previous one, and the chain of dependencies quickly becomes enormous. Storing the entire computational graph for backpropagation through hundreds of steps leads to memory explosion and unstable gradients.

How a for loop changes the approach to training

The authors show that the iterative generation process can be differentiated by simply organizing it as a standard `for` loop. The entire sequence of steps — from initial noise to the final image — is joined into a single computational chain through which the gradient passes.

"AlexNet proved something that has been confirmed again and again ever

since: handing the entire task over to end-to-end model training is almost always better than a carefully human-designed step-by-step pipeline," — as stated in the Jiqizhixin article.

The analogy with AlexNet is direct: before 2012, features for image recognition also passed through a multi-stage pipeline, and it seemed right to optimize each stage separately. AlexNet showed that if you allow the entire chain to be differentiated as a single whole, the model finds the best strategy on its own. Now the same applies to the generation loop.

No new architecture or exotic optimizer is required. Standard backpropagation, implemented through a basic programming language construct, allows the entire generation process to be trained as a unified system. Modern frameworks — PyTorch and JAX — natively support differentiation of operations within a loop.

What this means

If end-to-end training scales to large generative models, it could change the standards for their training just as AlexNet changed approaches to recognition tasks. Instead of disparate, sequentially trained components — a unified system that independently optimizes the generation strategy. The history of deep learning shows: when a model is given the opportunity to learn on the task as a whole, it almost always finds a solution better than one designed by hand.