- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
CONFLUX separates semantic conditioning in a shared backbone from iterative image rendering in a flow transformer. It applies content–query factorization to text and images, using separate states for observed content and prediction conditions. Experiments with a trainable Qwen3-0.6B backbone report improved language retention and visual understanding, faster image generation, and generation-quality improvements from scaling the flow head.
Critical strengths
Reducing the computational cost of unified multimodal models is practically meaningful. The paper provides comparisons with matched initialization, data, and update budgets, and reports improvements in language retention and visual understanding. The attention rules and experimental limitations are clearly described.
Critical weaknesses
-
The incremental benefit over existing fixed-conditioning query interfaces is insufficiently isolated. The authors appropriately acknowledge DreamLLM, MetaQueries, ScaleRAE, and XLNet, and their contribution is the particular modality-general integration. However, the main comparison is against iterative-backbone designs. A matched image-interface comparison—ordinary independently learned queries with query-to-query attention versus CONFLUX’s shared-modality, content-only-K/V queries—would clarify the value of the proposed interface while retaining the same backbone, image inputs, and renderer. Table 13 isolates the text interface but retains CONFLUX’s image-query design. The current results therefore leave the advantage over the closest architectural alternatives unclear.
-
The specific interleaved visibility benefit remains unvalidated. The teacher-forced [P, I, S] example is a legitimate architectural motivation, but Appendix D.1 states that actual training uses separate T2I/I2T microbatches and reports no mixed-sequence benchmark. Moreover, all baselines receive complete, clean I2T inputs. The understanding gains and local attention ablations do not directly establish the benefit of preserving clean image context in a pass where the image is also a prediction target. This distinction limits the empirical support for that part of the contribution.
Questions for authors
-
Under matched backbone, image inputs, and renderer, what additional benefit comes from the specific image-query interface over standard learned-query conditioning, including its training and inference cost?
-
Can the interleaved visibility advantage be evaluated directly, comparing appropriate conventional masking/query baselines in the same teacher-forced setting?
-
The paper explicitly separates common-setting throughput from FID-selected quality. Can the speed–quality frontier be reported at common operating points to quantify the practical trade-off?
Image-query design, interleaved visibility, and generation efficiency
We sincerely thank you for your careful review and insightful questions on the image-query interface, interleaved visibility, and efficiency. To clarify the role of content-query (CQ) factorization beyond semantic-rendering separation, we added a jointly trained one-pass Learned-Query Diffusion (LQD) baseline, direct query-interface controls, interleaved training, and speed/cost profiles. References below are to the revised paper, with key updates in blue (pp = percentage points).
1. The image-query interface and query interaction (W1, Q1)
The default shared-query interface improves language and visual understanding over independent queries with causal interaction, with CQ, bidirectional image-content attention, and the 200.7M renderer held fixed (Table 18 and Appendix D):
Table 18 crosses CQ versus AR text prediction and bidirectional versus causal image content across both query designs (eight configurations, including LQD), showing two key patterns:
All Table 1 and 18 models share the Qwen3-0.6B initialization, clean image inputs, data order, and 66.7B-token budget. Cross-attention replaces image-condition AdaLN in CONFLUX to match LQD's renderer (preserving CQ, one-pass conditioning, and time AdaLN), improving FID from 3.671 to 3.216 with comparable language and understanding performance (Table 13). SLD, NTD, and MAD use distinct native heads, so this change to CONFLUX's renderer does not affect their architectures or implementations.
2. Interleaved target/context use (W2, Q2)
After interleaved training beyond ImageNet, CONFLUX-B leads LQD (the strongest QA baseline) by +4.15/+4.68/+3.45 pp on GQA/VQAv2/OK-VQA (53.20%/49.80%/20.14%) and +188.41 on MME perception (853.56), supporting CQ's visibility design, while reaching 62.42% GenEval—exceeding SLD/NTD/MAD by +19.49/+53.35/+25.86 pp and comparable to LQD (63.56%; Tables 23-24). Training uses interleaved sequences (COCO/Flickr images), complete multi-turn image-text sequences, and BLIP3-o compositional text-image examples (Appendix E), supervising images and subsequent text in one teacher-forced sequence under native visibility (noisy/masked for SLD/MAD, causal clean for NTD/LQD, bidirectional clean for CONFLUX).
3. Quality-throughput frontier and training/inference costs (Q1, Q3)
CONFLUX-B improves the generation speed-quality trade-off against both baseline families (Figure 6; Table 15):
We hope these results clarify the image-query interface, interleaved modeling, and practical efficiency, and would appreciate your updated assessment of the revised manuscript.