- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper proposes splitting text-to-image generation into two stages. The first stage is Knowledge Acquisition where a text conditioned AR model is pretrained to predict discrete visual tokens. In the second stage, the AR backbone is frozen and pixel space diffusion models reuse the attention projections and are post trained with flow matching.
Critical strengths
- The framing of the paper is useful. Instead of asking AR vs Diffusion, the authors focus on specific strengths and use them in collaboration.
- The dissociation between reconstruction or AR only quality and final quality is a practically useful observation for tokenizer design
- A decoder free tokenizer trained only through language supervision is an interesting idea.
Critical weaknesses
-
The post training renderer receives codes extracted from the same image it denoises. The from scratch pixel diffusion model only receives text. A lower velocity MSE is therefore expected from the extra information about the target alone, independent of any benefit from pre training. Section 3.3 acknowledges that the comparison does not isolate the two factors, but the abstract and conclusion still attribute the lower floor to AR pretraining.
-
The fitted asymptotes also come without uncertainty estimates. The post-training allocation exponent ranges from 0.45 to 0.70 across estimators (Tab 9), which suggests parametric fit is weakly identified.
-
There are inconsistencies in proposed tokenizer and what the evidence favors. The Ar -> PDM row in Table 1 (65.51/71.11) is the semantic-recon configuration and not the semantic-lm as shown in table 2. Semantic-LM is actually lower on both geneval and DPG.
-
Semantic quality determines final performance is asserted rather than measured. No measure of semantic quality is reported (eg. linear probing, retrieval, or VQA), and conclusion rests on three tokenizers that differ in several ways at once.
-
There is a plausible alternative explanation of why recon gets worse after post training. Recon codes nearly determine the image(rFID 1.04), so renderer trained on GT codes may learn to copy them and hence is brittle to AR errors. This would be consistent with recon getting worse post training. Coarser semantics may leave more to the renderer and may simply tolerate rollout errors better. The paper does not discuss or distinguish between these explanations.
-
The compute-efficiency claim does not consider the tokenizer training (A.1). More fundamentally, a substantial part of the "knowledge" comes from Qwen2.5-VL-32B features rather than the AR pre-training.
Questions for authors
-
Can you separate pretraining from conditioning in the asymptote comparison? Either of these would do: i) a from-scratch pixel diffusion model conditioned on the same ground-truth codes, or ii) post-training loss evaluated with AR-generated codes.
-
Which tokenizer does the final model use? If it is Semantic-Recon, how should readers interpret the emphasis on Semantic-LM?
-
What is the measured compute for tokenizer training and for the frozen AR forward pass during post-training, and what is the measured wall-clock time of the final run?
-
GenEval and DPG-Bench measure prompt alignment, not fidelity. Since rendering quality is a central claim, an FID, human-preference, or text-rendering evaluation would help