acceptodds
Withdrawn from ICLR 2027

DualCast: Dual-Image Conditioned Diffusion Transformers for Digital Human E-Commerce Video Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Digital human e-commerce video generation requires preserving product appearance, human identity, and overall video style during dynamic demonstrations. Although recent models support multiple reference images, their general conditioning schemes do not explicitly distinguish the asymmetric roles of “what is presented” and “who presents it.” Consequently, generated videos may favor a single reference, resulting in product distortion, identity drift, or single-reference collapse. To address this issue, we propose DualCast, a role-addressed dual-image conditioning framework for video generation. By explicitly modeling the distinct roles of product and human conditions, DualCast reduces role ambiguity and feature competition between references. First, a dual-spatiotemporal slot injection mechanism VAE-encodes the product and human images separately and places their latent features at the front and rear boundaries of the DiT sequence, enabling sequence-level condition separation and directed injection. Second, a mask-guided region-weighted loss strengthens supervision over product, face, and product–face overlap regions, improving generation fidelity in critical areas. Third, image condition dropout enables a single model to support dual-image, product-only, and human-only generation. In addition, we construct a curated dataset of 10K e-commerce videos for model training. Experimental results demonstrate the strong overall performance of DualCast.

Primary area:
generative models
Keywords:
E-commerce Video Generation, Dual-image Conditioning, Diffusion model

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_WmGxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

DualCast fine-tunes Wan2.2-I2V-14B to generate a presenter showing a product, given a product image, a presenter image and a script. The two images are VAE-encoded and attached to the video latent as an extra first frame (product) and an extra last frame (presenter), marked by an indicator channel; their CLIP features are appended to the text context. Training adds a mask-weighted flow-matching loss (overlap 3, product 2, face 2, background 0.5) and 50/25/25 dropout over dual/product-only/human-only inputs, and keeps Wan2.2's high/low-noise expert split. The 10K training clips are synthetic: first frames from GPT-Image-2, videos from Wan2.7, filtered by GPT-5.5. Evaluation uses ViCLIP, a CLIP-based subject-consistency score (S2V-C), Qwen3.5 judge ratings, VBench metrics and a 50-person user study.

Critical strengths

The task is practical and the design is light: two extra latent frames and one indicator channel on an existing I2V model, with one checkpoint serving dual, product-only and human-only inputs. The comparison covers four open-source and four commercial systems and includes a user study.

Critical weaknesses
  • The central claim is never tested. The thesis is that fixed front/rear positions remove role ambiguity and feature competition between product and presenter, but no experiment isolates position. "w/o Dual-End" (Table 2) is not defined, and App. F states that its identity drop "is not a positional-effect estimate"; Sec. 3.1 adds that roles are also carried by CLIP-token order and role-specific text. Temporal concatenation of VAE reference latents is already used in HuMo, which the paper cites, so position-as-role is the main new element, and it is the one left unverified. The obvious controls are missing: both references on one side, swapped sides, and a plain dual-reference fine-tune of the backbone (or VACE) on the same 10K clips. Without them, Table 1 cannot separate the conditioning design from domain fine-tuning.

  • The evaluation skips the task's key requirement and is confounded by the data. Product fidelity, the first requirement in the abstract, is absent from the main comparison: ProdID and ArcFace ID-Ret appear only in ablations on a different case set. The one objective subject metric in Table 1 (S2V-C 0.596) ties VACE and is below the base model (0.717), and the user study gives 52% product preference vs. VACE. The headline wins are LLM-judge scores on six mid-clip frames at ≤512 px, with margins around 0.02 (≈0.1 on the raw 0-5 scale) and no variance. Training targets are Wan2.7 outputs, and Wan2.7 is also a comparator. Training and test prompts come from the same generator (Qwen3.6-27B), and training clips were pre-filtered by GPT-5.5 for quality and prompt compliance, close to what the judge rates. After fine-tuning on 10K e-commerce clips, winning on "EcomStyle" and "Interaction" against general-purpose models is expected. The user study leaves out all commercial systems, and the signed-rank test specified in its protocol is not reported (App. A.4).

Questions for authors
  • Can you run position-controlled ablations on the same data (same-side slots, swapped slots, roles given only by text/CLIP order) and a plain dual-reference fine-tune of the backbone? This would change my assessment more than anything else.
  • Please add ProdID, a product-region similarity (e.g., DINO features on the segmented product) and ArcFace ID-Ret for every method in Table 1, with per-case variance. Please also report how often each method drops one of the two subjects, since that is the failure mode the paper is built around.
  • App. A.4 says the signed-rank test cannot be computed "without participant/case records and test outputs." Do you have the raw study data? Why were the commercial systems, Wan2.7 in particular, excluded from the study?
  • How are frame-level masks aggregated onto temporally compressed latent positions? App. E describes the issue but not the procedure used.
  • How was the single-image Wan2.2-I2V given two references in Table 1, and what exactly does "w/o Dual-End" keep?
Reviewer_QJyEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

DualCast generates digital human e-commerce videos from human-only, product-only, or dual-image inputs. Its main components are Dual-Spatiotemporal Slot Injection, Region-Weighted Supervision, and flexible image conditioning. The paper also introduces a curated dataset of 10K synthetic e-commerce videos conditioned on real and synthetic reference assets.

Critical strengths
  1. DualCast explicitly distinguishes the asymmetric roles of product and human references.
  2. The paper introduces a task-specific e-commerce video dataset.
  3. The figures and tables are clear.
Critical weaknesses
  1. Unverified dataset contribution. No controlled data ablation establishes the benefits of the proposed dataset.
  2. Incomplete baseline coverage. The paper lacks quantitative and qualitative comparisons with the closely related MAGREF [1], despite citing it.
  3. Insufficient conditioning ablations. The choice of the product–video–human order is not experimentally justified. Controlled comparisons with different orders are needed.
  4. Untested mask robustness. Region-Weighted Supervision relies on externally generated masks. The paper does not experimentally assess robustness to mask quality.
  5. Unexamined presenter bias and identity risks. In product-only mode, the presenter is generated from the script and model prior. The paper does not analyze demographic or appearance biases in these generated presenters, or whether they reproduce training identities. Its general privacy discussion does not empirically address these risks.
  6. Insufficiently established technical novelty. The novelty of Dual-Spatiotemporal Slot Injection and Region-Weighted Supervision relative to existing reference-conditioning and regional-supervision methods is not convincingly established. The missing comparisons and controlled ablations further weaken the attribution of gains to these designs.

[1] MAGREF: Masked Guidance for Any-Reference Video Generation with Subject Disentanglement.

Questions for authors
  1. Can you provide controlled data ablations assessing dataset sources under the same backbone and training budget?
  2. Can you add quantitative and qualitative comparisons with MAGREF?
  3. Can you compare swapped product/video/human positions?
  4. Can you evaluate robustness to segmentation masks of varying quality?
  5. Can you analyze presenter biases and potential training-identity reproduction in product-only mode, and discuss mitigations?
Reviewer_4gtxOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes DualCast for generating promotional videos from a product image, a presenter image, and a text prompt. It adapts Wan2.2 by placing the two reference latents at opposite sequence boundaries, applying region-weighted supervision, and using condition dropout to support missing references. It also constructs a dataset of 10,000 generated demonstration clips.

Critical strengths

The paper addresses a relevant application that requires preserving both human identity and product appearance. Supporting three input configurations within one checkpoint is practically useful. The evaluation includes several competing systems, component ablations, and full-video human judgments. The authors also acknowledge several limitations of their comparisons and metrics.

Critical weaknesses
  1. How sensitive are the multimodal judge’s scores to sampling only six middle frames? Could the authors report agreement with full-video human ratings and confidence intervals for the improvements?
  2. How are frame-level masks aligned with the temporally compressed video latents? Does additional weighting of product–face overlap improve results compared with weighting only the product and face regions?
  3. Does DualCast generalize beyond the Wan2.7-generated training distribution? Could the authors evaluate it on an independently collected benchmark with different presentation styles and challenging product manipulation?
Questions for authors

See Weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.