iMontage: Unified, Versatile, Highly Dynamic Many-to-many Image Generation

1. Introduction

• Most unified image models remain specialized for single-input, single-output tasks, leaving many-to-many generation for academic and open-source exploration.The passage notes that a commercial model had recently extended unified generation to multi-input, multi-output settings.Source: §1 · p. 2
• Image-only models can generate diverse outputs but struggle with temporal consistency, whereas video models’ contiguous-clip training limits abrupt transitions and large motions.These limitations constrain video models’ transfer to highly dynamic content and versatile tasks.Source: §1 · p. 2
• iMontage adapts a pretrained video model for variable numbers of reference and output images, using a RoPE strategy, curated data, and multi-task training.Its RoPE strategy distinguishes discrete image sets from continuous video while aiming to preserve temporal coherence.Source: §1 · p. 2
• iMontage reports strong qualitative performance across one-to-one editing, many-to-one generation, and many-to-many generation, with state-of-the-art quantitative metrics on their benchmarks.The benchmarks cover image editing, in-context learning, and storyboard generation.Source: §1 · p. 2

2.1. Unified Generation and Editing Models

• Unified image systems such as OmniGen [59] and ACE++ [35] combine generation and editing, but predominantly target single-input, single-output tasks.Later systems use multimodal language models to interpret instructions and guide diffusion decoders.Source: §2.1 · p. 3

2.2. From One-to-one to Many-to-many Generation

• Video-backbone models such as UniReal [10], RealGeneral [30], and Frame2Frame [45] support variable image sets, but contiguous-clip training limits highly dynamic, discontinuous content.The passage contrasts these models with any-to-any approaches such as BAGEL [60] and OmniGen [59], trained on interleaved multimodal data.Source: §2.2 · p. 3

3.1. Model Design

• The MMDiT and 3D VAE encode reference images while a language model encodes instructions; clean reference tokens join noisy target tokens, with variable-length image attention and task-specific prompts.The model preserves spatial RoPE and assigns input images early temporal positions and outputs late positions, separating them with a margin to reduce positional interference.Source: §3.1 · pp. 3–4

3.2. Dataset Creation

• Dataset construction separates pretraining data from supervised fine-tuning data, organizing foundational examples and task-oriented subsets.The overview describes four data sources arranged in two stages.Source: §3.2 · p. 4, Figure 3 · p. 4

3.2.1. Pretraining Dataset

• 5M image-edit pairs and 15M video frame pairs form the pretraining data, with high-motion sampling and cross-transition pairs diversifying video supervision.The two pools provide paired editing instructions and captioned video frames; frame-pair processing preferentially retains or upweights high-motion samples.Source: §3.2.1 · pp. 4–5

3.2.2. Multi Task Dataset

• Around 90k CRef, 50k conditioned CRef, 35k SRef, 100k multi-turn, 90k multi-view, and 29k storyboard samples form task-specific supervision.The tasks span reference-conditioned generation and editing, camera-view variation, and storyboards designed for high inter-panel diversity.Source: §3.2.2 · pp. 5–6

3.3. Training Scheme

• Training compares FlatMix, StageMix, and difficulty-ordered CocktailMix for SFT, selects CocktailMix, and follows pretraining with brief all-task high-quality fine-tuning and learning-rate annealing.Pretraining uses aspect-ratio-aware resolution bucketing; the high-quality stage uses manually reviewed and VLM-assisted subsets.Source: §3.3 · p. 6

4.1. One-to-one Editing

• On GEdit and ImgEdit, iMontage is reported to perform strongly against twelve baselines, with superior motion-aware editing, temporal consistency, and motion priors.The paper attributes these motion-related gains to inherited world-dynamics knowledge from its pretrained video backbone, reinforced with high-dynamics video–frame pretraining.Source: §4.1 · pp. 6–7

4.2. Many-to-one Generation

• Qualitative examples show context retention across diverse tasks: multi-reference cue fusion, conditioned human-detail preservation, and style transfer retaining identity.The examples include complex instructions and detailed backgrounds in Multi-CRef, and scene-centric and human/object-centric inputs for SRef.Source: §4.2 · p. 7

4.3. Many-to-many Generation

• In a 50-professional-participant user study, iMontage achieved the best instruction-following and identity-preservation performance, while producing coherent, diverse outputs across all three settings in one forward pass.The study reports storyboard ratings for twenty samples.Source: Table 4 · p. 8, §4.3 · p. 8

4.4. Ablation Study

• Marginal RoPE assigns input and output images to separated ends of the temporal index range, whereas Even RoPE spaces them uniformly; Even RoPE converged later under equal training steps.The ablation uses a low-resolution subset of editing data.Source: §4.4 · p. 8, Figure 5 · p. 8
• CocktailMix, which schedules by task difficulty, outperformed StageMix on harder settings and yielded a 12.6% gain on OmniContext; FlatMix loss oscillated and drifted toward easier tasks.StageMix groups training by task type; FlatMix remained unstable despite inverse-size reweighting.Source: §4.4 · p. 8

• iMontage is a unified many-to-many image generation model designed to create dynamic content while preserving temporal and content consistency.Source: §5 · p. 8
• Long-context many-to-many settings remain unexplored because of data and compute constraints, and quality is currently best with up to four inputs and four outputs.Source: §5 · p. 8

• Training uses pretraining, supervised fine-tuning, and high-quality annealing, with dynamic resolution grouping in pretraining and fixed task resolutions in SFT.Pretraining resolution buckets are based on 512, 768, and 1024 pixel square sizes with ±32-pixel variants; inference generalizes to arbitrary resolutions.Source: §6 · p. 12
• All DiT blocks remain trainable across stages, while the VAE and text encoders are frozen.Source: §6 · p. 12
• Inference uses text classifier-free guidance at 6.0 with 50 steps, and training drops captions with probability 0.1 for unconditional alignment.Source: §6 · p. 12

• Additional visualizations cover image editing, multi-reference, and multi-view results.Source: §7 · p. 12

8.1. Storyboard Generation Evaluation

• Storyboard evaluation uses foreground DINO and CLIP feature similarities to assess identity preservation and mean pairwise similarity across generated images for temporal consistency.The process segments foreground characters and computes similarities from masked, feature-extracted generated and reference character embeddings.Source: §8.1 · p. 12
• A GPT-4o judge receives input and output images and rates identity preservation and temporal consistency using task-specific prompts and scoring criteria.The prompts assess identity across images and against reference images, with attention to facial features, body shape, clothing, and identity mapping.Source: §8.1 · pp. 12–13

8.2. User Study

• The user study involved 50 participants rating 20 storyboard samples with one or two reference characters and two to four generated images.Samples varied across three racial groups, two genders, and prompt complexity.Source: §8.2 · p. 13
• Participants rated instruction following, identity preservation, temporal consistency, and overall quality on a five-point scale; competing models were anonymized and presentation order randomized.The same participant rated all competing models for each sample, and mean ratings were reported.Source: §8.2 · pp. 13–14

8.3. Training Scheme Ablation

• On OmniContext, StageMix underperforms CocktailMix by 12.6%, while storyboard generation has roughly 0.2 higher training loss than multi-editing under a shared setup.The authors report worse visualizations for StageMix on other tasks and say difficulty-ordered mixing improves optimization stability and generalization, especially on harder tasks.Source: §8.3 · p. 14

9.1. Concurrent Works

• iMontage casts image synthesis as a unified many-to-many problem, distinct from ChronoEdit’s edit-focused video framing and UniVid’s broad visual-sentence task adaptation.The authors describe this as a practical technical pathway toward a more capable, fully unified system.Source: §9.1 · p. 14

9.2. Observed Failure Case

• Chinese-character rendering is near-zero, SRef can leak backgrounds, and some generations show head-detail mismatches.The authors link these issues to the HunyuanVideo backbone’s limited text-rendering supervision, distilled SRef training data, and insufficient coverage of high-detail heads and faces.Source: §9.2 · p. 14
• The authors suggest identity modules using face embeddings and more high-quality head-focused data as complementary remedies for head-detail preservation.The proposed remedies target identity information and fine-grained facial-detail coverage, respectively.Source: §9.2 · p. 14