Five diffusion papers: July 14, 2026

Coverage window: July 13, 09:00, to July 14, 09:00 UTC-5. Intake is the active cs.CV/cs.LG arXiv new-listing pass visible at the scheduled scan; each entry below preserves its arXiv submission date, because several high-signal items surfaced in the listing with earlier v1 timestamps.

The ranked set favors papers where diffusion, rectified flow, or DiT internals are the main mechanism rather than a downstream decoration. The strongest pattern today is training-free control: reward signals, backbone intervention, motion transfer, long-video correction, and image editing all try to steer existing generators without retraining the whole model.

Decision: open SpectraReward first if you work on RL alignment for text-to-image models. It has the cleanest implementation story in today's set: instead of training a preference model or decomposing prompts into VQA checks, it asks a pretrained MLLM to read the prompt back from the generated image and uses average image-conditioned prompt log-likelihood as the reward. 1

Paper signal: Read It Back: Pretrained MLLMs Are Zero-Shot Reward Models for Text-to-Image Generation is arXiv 2607.11886, submitted on July 13, 2026. The project page lists Runhui Huang, Qihui Zhang, Zhe Liu, Yu Gao, Jie Wu, and Hengshuang Zhao, with affiliations at The University of Hong Kong, ByteDance Seed, and Peking University. 2 The page links code at huangrh99/AlphaGRPO. 2

Technical read: SpectraReward runs one teacher-forced MLLM forward pass over the original prompt conditioned on the generated image. A higher mean token log-likelihood means the image better supports the prompt. The self-reward variant uses a unified multimodal model's own understanding branch to score samples from its generation branch, so the reward model and policy share more of the same pretraining distribution. 2

Evidence and gaps: The project page reports that prompt likelihood beats scalar scoring and VQA-style reward decomposition on BAGEL: GenEval is 89.5 for prompt likelihood versus 84.0 for BAGEL without RL, 77.7 for scalar scoring, and 83.1 for VQA-score. It also reports TIIF-Short 85.1 and TIIF-Long 84.3 for prompt likelihood. 2 The broad study covers two diffusion models, three RL algorithms, nine reward MLLM backbones from four families, and five out-of-distribution T2I benchmarks. 1

Read it for: the reward-design shortcut. The first ablation to inspect is cross-family transfer: the self-reward result may partly come from reward-policy alignment rather than a universally better reward signal.

Decision: open DUNE if you work on diffusion sampling controls, artifact reduction, or cross-backbone diagnostics. It ranks high because it tries to locate a common failure point across U-Net and Transformer diffusion models rather than proposing a one-backbone patch. 3

Paper signal: Unified Backbone Refinement for Diffusion Models via Internal-Latent Analysis is arXiv 2607.09753, submitted on July 4, 2026, and marked accepted to ECCV 2026. 3 The full text lists Haksoo Lim, Myeongjin Lee, Wonjoon Chang, and Jaesik Choi from KAIST and INEEJI. 4

Technical read: DUNE, short for Diffusion Unified Network refiNEr, is training-free. It detects abrupt deviations in deep low-noise internal latents with an EMA-based rule, then suppresses the selected entries. In U-Nets, the intervention targets h-space through channel-aware masked scaling. In Transformer diffusion backbones, the same detect-suppress principle acts on deep self-attention latents through EMA blending. 4

Evidence and gaps: The full text reports tests on SDXL, LCM, Kandinsky 3, PixArt-Sigma, and Hunyuan-DiT, with comparisons to FreeU, ASCED, and PAG. DUNE improves SDXL FID from 18.86 to 18.75 and HADM from 0.227 to 0.074; on Kandinsky 3, FID moves from 21.36 to 20.76 and HADM from 0.250 to 0.095. 4 The cost is not zero: SDXL CLIP similarity drops from 1.000 to 0.925, and LCM drops to 0.940. 4

Read it for: the phase-aware latent analysis. The practical question is whether HADM captures the artifacts you actually care about, or whether the CLIP-similarity loss is visible in prompt adherence.

Decision: open HALO if you study video DiTs, motion transfer, or interpretable control in generation. It has a useful claim for model analysis: motion and structure are not just abstract properties of the output; the paper says they can be localized to distinct attention-head subsets. 5

Paper signal: Controlling Motion Transfer in Diffusion Transformers via Attention Heads is arXiv 2607.11081, submitted on July 13, 2026, and accepted to ECCV 2026. 5 The authors are Sunyoung Jung, Jiwoo Park, Yoonseok Choi, Kyobin Choo, Ming-Hsuan Yang, and Seong Jae Hwang. The arXiv page and project page do not expose a complete affiliation list. 5 6

Technical read: HALO first identifies temporal heads that produce displacement maps aligned with reference motion and low-entropy structure heads that preserve layout. It then refines motion cues with semantic correspondence guidance and preserves structure by selectively injecting value features from the reference. The method is training-free and is built on CogVideoX, with validation also claimed on Wan. 6

Evidence and gaps: On the standard motion-transfer benchmark, the project page reports HALO at CLIP 31.7, TC 87.5, MF 66.2, and FTD 19.4, beating DiTFlow, RoPECraft, GWTF, and U-Net-based baselines on motion fidelity. On the Movie Scene Dataset, HALO reports CLIP 30.5, TC 88.9, and MF 52.6. The user study reports motion accuracy 93.3, temporal consistency 87.8, and edit accuracy 86.5. 6

Read it for: the head-level intervention. The full paper should be checked for how stable the head selection is across prompts, seeds, and different video DiT checkpoints.

Decision: open Cycle-World if long-horizon video generation or world models are on your roadmap. The paper attacks the autoregressive failure mode directly: small frame-to-frame deviations compound until the sequence drifts or collapses. 7

Paper signal: Cycle-World: Mitigating Error Accumulation in Long-term Video World Models via Reverse-Prediction Cycle Consistency is arXiv 2607.11836, submitted on July 13, 2026, and accepted to ECCV 2026. 7 The full text lists Shanghai Jiao Tong University, Zhejiang University, and Nanyang Technological University as affiliations. 8 A project page is live at szhcz.github.io/projects/Cycle-World. 7

Technical read: Cycle-World uses a reverse-prediction model in both training and inference. During training, cycle-consistent learning pushes the forward generator to produce sequences that remain reversible. At inference, the frozen reverse model becomes a runtime corrector: gradient-based cycle guidance refines generated latent states before their errors enter the historical context. 8

Evidence and gaps: On 60-second VBench evaluation, the full text reports Cycle-World at Total 82.88, Quality 84.39, and Semantic 76.86, slightly ahead of LongLive's Total 82.62 and Context Forcing's 82.45. On physical consistency, it reports PC 70.90, PACE 75.42, and average 73.16, above LongLive's 67.58 average and Infinity-RoPE's 67.98. 8

Read it for: the inference-time corrector. The key implementation check is whether CGI's extra gradient steps still make sense when wall-clock generation cost is part of the product constraint.

Decision: open h-Flow if you care about principled image editing in rectified-flow models. It gives a mathematical framing for the familiar editing tension: stay close to the source image while moving toward the target prompt. 9

Paper signal: h-Flow: Flexible Flow-based Image Editing via Doob's h-Transform is arXiv 2607.10800, submitted on July 12, 2026. 9 The full text lists East China Normal University and The Hong Kong University of Science and Technology as affiliations. 10 Code is not live yet; the abstract says it will be released soon. 9

Technical read: h-Flow reformulates image editing as conditional generation under multiple terminal events: source consistency and target alignment. It extends Doob's h-transform from SDE models to deterministic rectified flow by constructing an equivalent SDE with identical marginals. Dedicated h-functions provide closed-form reconstruction guidance and velocity-based semantic editing signals, while velocity orthogonal decomposition separates reconstruction and editing directions. 10

Evidence and gaps: The full text reports h-Flow's average rank as first across nine metrics on PIE-Bench and PIE-Bench++. On PIE-Bench, h-Flow reports PSNR 23.88, LPIPS 0.110, SSIM 0.840, CLIP-Entire 25.80, CLIP-Edited 22.80, HPS 27.79, and AS 4.98. On PIE-Bench++, it reports PSNR 24.30, LPIPS 0.1090, SSIM 0.850, CLIP-Entire 25.48, CLIP-Edited 24.54, HPS 27.67, and AS 4.92. 10

Read it for: the separation of reconstruction and editing velocity. The limitation to inspect is explicit in the paper: the method still relies on a forward inversion trajectory, and complex multi-object edits may need the proposed decompose-then-project extension. 10

For T2I alignment and post-training, start with SpectraReward, then h-Flow. SpectraReward gives a reward signal for RL; h-Flow gives a controllable edit trajectory after the base model exists.

For model internals, read DUNE before HALO. DUNE asks where artifact-prone internal latents emerge across backbones; HALO asks which video DiT heads carry motion and structure.

For long video, read Cycle-World after HALO. HALO is about transferring reference motion in a clip; Cycle-World is about keeping autoregressive video stable over 60 seconds.