Image generation splits in two: learn semantics, then render, and a 3B model reaches GenEval 0.86 in 6,600 H100 hours

Image generation no longer asks one objective to do two jobs: autoregressive pretraining acquires high-level visual knowledge, and diffusion post-training renders it into pixels. Autoregressive models and latent or pixel diffusion models have usually been compared as competing paradigms, each asked to solve both knowledge acquisition and detailed rendering. With final generators of about 400M parameters and the same cumulative number of training images, the autoregressive-then-diffusion route reached 64.8 on GenEval after 100k post-training updates, above 61.2 for pixel diffusion trained for 400k iterations; further post-training brought it to 65.5 on GenEval and 71.1 on DPG-Bench, versus 57.3 on GenEval for direct latent diffusion and 56.6 for latent-diffusion pretraining plus diffusion post-training.

What pretraining should learn is semantics, not pixel reconstruction. Conventional tokenizers preserve low-level appearance for image reconstruction, so the autoregressive representation is required to support reconstruction or standalone image generation. The Recon tokenizer with better reconstruction yielded lower final generation, 54.7 on GenEval and 63.0 on DPG-Bench, below Semantic-Recon at 65.5 and 71.1; Semantic-LM, trained through language supervision alone with no image decoder or reconstruction loss, reached 64.8 on GenEval and 67.7 on DPG-Bench after post-training.

Each stage gets its own fitted scaling law, which then sets how compute is divided between pretraining and rendering. Earlier scaling studies mainly scaled a single learning objective; separating the stages adds the question of how much budget each receives and how each balances model size against data. The autoregressive pretraining law was fitted on 18 models spanning 55.71M to 2.43B parameters and the post-training law on 14 models spanning 25.1M to 2.03B; the derived allocation assigns about 60% of the training budget to knowledge acquisition and 40% to rendering, and stays nearly invariant as total compute grows.

The scaling-guided recipe was carried to a roughly 3B-parameter text-to-image model. The recipe first assigns the total compute budget to autoregressive pretraining, then uses the pretraining law to set backbone size and data and the post-training law to set rendering data. Trained in about 6,600 H100 GPU-hours, the model scored 0.86 on GenEval and 87.78 on DPG-Bench; among the compared baselines only Qwen-Image and Seedream 3.0 reported higher DPG-Bench scores, and its GenEval is close to Qwen-Image and BLIP3o-NEXT SFT at 0.87 and SeFi-Image-5B at 0.88.

The scope is text-to-image generation: controlled comparisons use generators of about 400M parameters, the scaling laws cover 55.71M to 2.43B parameters for autoregressive pretraining and 25.1M to 2.03B for post-training, and the final validation is a roughly 3B-parameter model trained in about 6,600 H100 GPU-hours. The source qualifies cross-system comparisons as approximate because text encoders, training data and evaluation settings differ.

The source notes that the asymptotic-loss comparison between post-training and pixel diffusion trained from scratch evaluates the pipeline as a whole, including pretrained weights and conditioning tokens extracted from the target image, and does not isolate their individual contributions; the asymptotic term is an extrapolated trend under the fitted model. It also states that the final model shows competitive generation quality within the stated training budget without establishing that the compute allocation is optimal.