CanvasComposer: Personalized Group Photo Generation via a Multi-Reference Canvas

1 Introduction

• Token-concatenation methods can swap, mix, duplicate, or omit identities as subject count grows, while their conditioning sequences increase linearly in length.Multi-human generation also requires preserving each identity, binding each reference to the correct generated person, and rendering subjects coherently.Source: §1 · p. 2
• CanvasComposer lets users place and resize separate subject references on a shared canvas, then renders them as one coherent, personalized group photo.References may overlap freely, have no fixed ordering, and can include an optional background.Source: §1 · p. 2
• Separate RGBA cutouts preserve subject content under input overlap, unlike flattened collages that occlude identity information.The canvas layout also supports user adjustment of relative subject locations.Source: §1 · p. 2, Figure 2 · p. 2
• Transparent latent pruning retains only non-transparent subject tokens, making conditioning length depend on visible reference content rather than growing linearly with subject count.The paper presents this as a way to reduce unnecessary conditioning tokens and improve efficiency.Source: §1 · p. 2
• Subject-wise cross-reference training samples humans and backgrounds from other images in the same identity-group set, with geometric and lighting augmentation.The strategy encourages disentangling pose and illumination differences so diverse inputs can be harmonized into one coherent image.Source: §1 · p. 2
• CanvasComposer combines interactive multi-human composition with transparent latent pruning and cross-reference training; experiments report state-of-the-art identity preservation and generation quality.The framework is built on the pretrained FLUX Kontext diffusion transformer.Source: §1 · pp. 2–3

2 Related Work

• Multi-subject personalization approaches use adapters, optimization, or in-context conditioning, but several remain costly as subject count grows or offer limited interactivity.Ar2Can [Borse et al. 2026] uses an Architect-predicted canvas and an Artist module, whereas CanvasComposer uses user-composed RGBA references that preserve identity under overlap.Source: §2 · p. 3
• Structural-conditioning methods guide generation with cues such as pose, depth, boxes, or masks, while layered-generation methods separate or render output layers.CanvasComposer instead conditions on a separate RGBA reference image per subject for identity personalization, including when inputs overlap.Source: §2 · p. 3
• ID-Patch [Zhang et al. 2025b] uses collaged face inputs but is limited to face-only generation and suffers from occlusion and copy-paste artifacts.CanvasComposer accepts full-body, portrait, or cropped-head references and uses subject-wise cross-reference training to mitigate copy-paste artifacts.Source: §2 · p. 3

3.1 Multi-Reference Canvas

• Each subject or optional background remains a separate RGBA canvas element, with alpha marking valid regions and no fixed element ordering.Source: §3.1 · pp. 3–4

3.2 CanvasComposer Pipeline

• The pretrained DiT processes pruned canvas tokens jointly with noisy image tokens, using 3D positional embeddings for element-aware conditioning.Source: §3.2 · pp. 4–5
• Transparent latent pruning retains tokens in non-transparent regions, making sequence length depend on canvas content rather than subject count.Source: §3.2 · pp. 4–5

3.3 CanvasComposer Training

• Training samples each subject's reference from another image in the same scene while using a different image as target, encouraging cross-image identity integration to mitigate copy-paste artifacts.Source: §3.3 · p. 5
• CanvasComposer wins over 84% of pairwise comparisons against personalization methods and over 60% against recent image-editing models across the 4P and 2P benchmarks.Source: Table 1 · p. 5

4.1 Experimental Setup

• Evaluation uses 128 FFHQ-in-the-wild identity inputs not included in training, with 32 prompts per benchmark at 1024×1024 resolution.CanvasComposer is evaluated with 28 denoising steps and no post-processing; the training data includes 32M in-house images across 6M scenes, and that dataset cannot be released.Source: §4.1 · pp. 5–6

4.2 Baseline Comparisons

• Over 96% of pairwise comparisons favored CanvasComposer on average against 4P personalization baselines; it also outperformed the image-editing baselines in identity preservation and user preference.The 4P evaluation reports the highest identity preservation among personalization baselines while maintaining competitive text-image alignment and aesthetic quality.Source: §4.2 · pp. 6–7
• In 2P personalization, CanvasComposer achieves the best identity preservation and strong user preference, while producing coherent images that preserve both reference identities.Competing methods may omit or duplicate a subject, mix identities, or retain facial identity poorly.Source: §4.2 · p. 7, Figure 6 · p. 7
• For 1P personalization, CanvasComposer follows varied prompts while preserving reference identity, whereas competing methods tend to limit pose or expression variation.Examples include prompts for smiling and laughing.Source: §4.2 · p. 7, Figure 7 · p. 8

4.3 Ablation Study and Analysis

• Removing subject-wise cross-reference training produces copy-paste artifacts and lower image quality; the full model preserves identities more naturally and avoids copying input expressions.The ablation attributes this difference to disentanglement introduced by the training strategy.Source: §4.3 · p. 7, Figure 9 · p. 9
• Under overlap, the single-collage variant loses fine details and identity fidelity, whereas separate canvas subjects preserve references; canvas layout controls generated subjects’ spatial arrangement.The overlap benchmark finds the collage-input variant competitive on standard 2P personalization but substantially less robust under overlap.Source: §4.3 · pp. 7–8, Table 3 · p. 8
• Transparent latent pruning cuts average conditioning tokens by over 85% and reduces 4P inference latency from 91.8s to 21.2s.The efficiency analysis is reported for the 4P Personalization Benchmark.Source: §4.3 · p. 8, Table 2 · p. 8
• CanvasComposer also generates people interacting naturally with optional backgrounds, while results beyond four people degrade and FLUX Kontext does not reliably follow prompts with more than four people.The paper also shows anecdotal canvas examples with a non-human object as an additional reference.Source: §4.3 · p. 8, Table 4 · p. 9, Table 5 · p. 9

Reasoning Limitations

• Strong interaction reasoning can fail: subjects may sit unnaturally on chairs, producing copy-paste-like outputs close to the inputs.The authors suggest integrating Vision Language Models as a promising remedy.Source: §5 · p. 9

Beyond 4P Generation Limitations

• Performance degrades beyond four people, reflecting both limited training data and FLUX Kontext’s unreliability at that scale.The authors cap training data at four people and suggest that more diverse, higher-quality samples and access to raw base weights could improve robustness.Source: §5 · p. 9

Output Layout and Occlusion Control

• Canvas indices separate subjects but encode no depth order; output occlusion ordering follows the base model.The canvas preserves reference pixels when subjects overlap in the input.Source: §5 · p. 9

Base Model Dependence

• Results use a frozen FLUX Kontext base with a lightweight LoRA, so performance may vary with other base models.The paper ties this dependence to the backbone’s strengths and weaknesses, including generation beyond four people.Source: §5 · p. 9

Single Flattened Output

• The output is a single harmonized image, not layered or transparent; layered RGBA generation is proposed as future work.The authors suggest combining the input canvas with transparent-image generation techniques.Source: §5 · pp. 9–10

Results Beyond Humans

• At inference, the model can produce coherent interactions between people and non-human objects despite training only on human-centric data.The paper demonstrates this in results beyond humans.Source: Figure 10 · p. 10

Qualitative Results on the 2P Overlap Benchmark

• On the 2P Overlap Benchmark, the collage baseline loses the occluded reference, while CanvasComposer preserves both identities.The input illustration’s first row shows the composited canvas; its second row shows references side by side for demonstration.Source: Figure 11 · p. 10

4P Personalization with Background

• With four people and an optional background, the model produces natural interactions and coherent lighting, such as people leaning against a tree.The canvas in this example contains five separate images: four people and one background.Source: Figure 12 · p. 10

6 Conclusion

• CanvasComposer improves identity preservation and generation quality over existing methods by keeping references separate, pruning transparent tokens, and using cross-reference training to mitigate copy-paste artifacts.The conclusion describes the multi-reference canvas as an effective interface for adapting pretrained diffusion models to multi-human personalization.Source: §6 · p. 10

A User Study

• The pairwise user study collected 14,720 choices from 460 participants across 23 baselines and three benchmarks, using 32 prompts per benchmark.Each comparison involved 20 random participants; each participant completed one study, with method order shuffled for every question.Source: A User Study · p. 13

B Additional Results

• The evaluation adds image-editing methods on the 4P Personalization Benchmark and single-subject personalization approaches on the 1P benchmark.The main paper focuses on comparisons against multi-subject personalization baselines.Source: B Additional Results · p. 13

B.1 Comparison with Image Editing Pipelines in 4P Personalization

• The qualitative comparison covers FLUX Kontext, Overlay Kontext, Qwen-Image-Edit, and Nano-Banana, using collaged inputs and identical prompts.Overlay Kontext prompts also include its official trigger phrase, “Place it.”Source: §B.1 · p. 13
• In a complex interaction example, subjects fail to sit naturally on chairs and the output closely resembles the inputs.This is presented as a failure sample for scenarios requiring strong interaction reasoning.Source: Figure III · p. 13
• CanvasComposer handles the occlusions common with more personalized subjects, whereas baseline approaches often fail in these cases.The passage points to Fig. II for the comparison.Source: §B.1 · p. 13
• CanvasComposer’s multi-reference representation and cross-reference training improve performance over its FLUX Kontext base model.Quantitative comparisons for image-editing baselines are reported in the main paper’s table.Source: §B.1 · p. 13

• CanvasComposer produces coherent, natural generations that follow text prompts while maintaining identity, unlike competitors that often inject identity with limited pose and expression variation.The comparison includes IP-Adapter, PuLID, and InfiniteYou, all built upon FLUX.1 dev.Source: §B.2 · p. 13
• 0.893 VQAScore is the highest prompt adherence among identity-preserving baselines, and user win rates range from 57.0% to 85.0%.ID-Patch's 0.911 VQAScore is paired with ArcFace 0.124, so the passage says the reference identity is not preserved.Source: §B.2 · p. 13
• A failure example in a complex reasoning scenario complements the paper's main-text Limitations section.Source: §B.2 · p. 13
• CanvasComposer captures varied expressions and challenging states, including smiling, relaxing with closed eyes, and eating, which involves complex poses or expressive gestures.Source: Figure I · p. 14, §B.2 · p. 14

C Additional Ablations

• Rank 512 gives the strongest ArcFace score on the 4P Personalization Benchmark while retaining competitive HPSv3 and VQAScore, so it is the default.Source: C Additional Ablations · p. 16

D Full 3D RoPE Formulation

• CanvasComposer reuses FLUX Kontext’s factorized 3D RoPE without modification, with token positions comprising an element index j and latent-grid coordinates (x, y).Source: D Full 3D RoPE Formulation · p. 16
• The RoPE rotation is block-diagonal across element, x, and y axes, making attention logits depend on relative offsets while repurposing j for canvas-element identity.Source: D Full 3D RoPE Formulation · p. 16

• OmniGen2 uses an in-context prompt naming which image contains each person; all other methods, including CanvasComposer, receive the prompt directly.Source: §E.1 · p. 16
• CanvasComposer receives the multi-reference canvas, whereas image-editing pipelines receive its composited visualization because they do not support that canvas representation.Source: §E.1 · p. 16
• Previous personalization methods receive individual input images, as required by their respective methods.Source: §E.1 · p. 16

E.2 Automated Benchmarking Pipeline

• For 4P and 2P benchmarking, the pipeline generates a prior image, detects people and faces, segments input subjects, matches them, and assembles the multi-reference canvas; 1P uses face landmarks to create a single-subject RGBA image.Source: §E.2 · p. 16

F Ethics and Broader Impact

• In-house training data cannot be released under internal data policies, though a comparable dataset can be curated from public videos.Source: F Ethics and Broader Impact · p. 16
• Personalized human generation could be misused to create images of real people without consent, including for impersonation or misinformation; the authors encourage consent and provenance signals.Source: F Ethics and Broader Impact · p. 16
• Generation quality and identity preservation may vary across demographic groups because of biases inherited from training data and the frozen base model.The authors recommend evaluating and mitigating disparities before downstream-product deployment.Source: F Ethics and Broader Impact · p. 16