InfiniHuman: Infinite 3D Human Creation with Precise Control

1 Introduction

• The paper identifies costly, limited-diversity captured datasets and imprecise control of garment appearance and body shape as key challenges.Existing training-free approaches such as SDS also have long optimization times and limited visual fidelity.Source: §1 · p. 2
• InfiniHuman is a fully automated framework that integrates existing vision-language, image-synthesis, pose-estimation, and diffusion models to generate annotated human identities.The framework is presented as a response to the high cost and limited diversity of manually captured datasets.Source: §1 · p. 2
• InfiniHumanData contains over 111K identities with multi-view images, fine-grained text, SMPL parameters, and explicit clothing representations.Examples span ethnicities, ages, clothing styles, hair types, and skin tones, and are described as visually indistinguishable from real scan renderings.Source: §1 · p. 2, Figure 2 · p. 2
• InfiniHumanGen synthesizes avatars jointly conditioned on text, clothing images, and body shape.The two complementary models support rapid generation and high-resolution, visually detailed creation.Source: §1 · p. 2
• InfiniHumanGen outperforms prior works in visual quality, speed, and attribute controllability, with significantly lower computational cost.Gen-Schnell targets rapid Gaussian-splat generation, while Gen-HRes produces high-resolution, photorealistic textured meshes.Source: §1 · p. 2

2.1 3D Human Generation.

• Existing generation approaches support text or body-shape conditioning, but not direct control over detailed clothing alongside both, a gap InfiniHuman addresses with multimodal conditioning.The paper connects this control gap to personalized-appearance applications such as digital fashion and virtual fitting rooms.Source: §2.1 · p. 3

2.2 Large-Scale 3D Datasets.

• Commercial scan repositories are expensive, while public human datasets are often limited in scale, demographic diversity, and coverage of clothing, age, and body morphology.Commercial scans often cost about 100 USD per identity; recruitment, scanning logistics, and privacy constrain public datasets.Source: §2.2 · p. 3
• Multi-view capture datasets can be restricted to fixed viewpoints and controlled lighting, while video-diffusion datasets can have view inconsistencies and lack true 3D geometry.The passage attributes the latter limitations to neighbor-only attention and the absence of explicit 3D supervision.Source: §2.2 · p. 3
• InfiniHumanData targets dataset limitations with diverse identities and fine-grained identity-level annotations, including text and clothing images, for controllable 3D human generation.The authors also describe releasing the automatic pipeline and dataset to support further expansion.Source: §2.2 · p. 3, Table 1 · p. 3

3.1 InfiniHumanData - Generation by Reconstruction

• InfiniHumanData generates structured descriptions, clothing images, SMPL body shapes and keypoints, and controlled-lighting orthographic multi-view images.The orthographic views are designed for 3D lifting.Source: §3.1 · p. 3, Figure 3 · p. 4
• Captions are expanded into diverse variants and summarized at ten granularities, from 40 words to 5 words.The hierarchy provides both coarse and fine-grained semantic cues.Source: §3.1 · p. 3
• FLUX is fine-tuned on uniformly lit orthographic scan renderings to generate scan-like images compatible with downstream 3D reconstruction.The stylization retains FLUX’s diversity while aligning generated images with the reconstruction domain.Source: §3.1 · p. 4
• Instruct-Virtual-TryOff extracts clean garment images from full-body images, while GPT-4o rejects poor garment matches among four generated candidates.The selection considers color, texture, length, and details such as zippers and pockets.Source: §3.1 · p. 4
• SMPL pose is fitted by aligning projected joints with OpenPose 2D joints, following NLF regression from orthographic views.Per-joint weights and regularization are adjusted for pixel-level matching.Source: §3.1 · p. 4
• Orthographic multi-view diffusion generates front, left, right, and back body and head images, using row-wise view attention and cross-attention between corresponding body and head views.At inference, the model takes an orthographic image and SMPL normal maps.Source: §3.1 · p. 5

3.2.1 Joint Conditional Distribution.

• Gen-Schnell conditions multi-view diffusion on SMPL normal maps and clothing images, then uses rendered 3D Gaussian splats to enforce consistency across views.It produces a 3D-GS in about 12 seconds, though 256×256 resolution leaves detailed features blurry.Source: §3.2.1 · p. 5
• Gen-HRes fine-tunes a multi-image translation model and uses aligned views for normal estimation and SMPL-driven volumetric carving into high-fidelity meshes.It offers higher resolution, visual fidelity, and detailed text prompting than Gen-Schnell, with generation taking about 4 minutes.Source: §3.2.1 · p. 6

4.1 Implementation Details

• The orthographic multiview diffusion model builds on SD2.1-unclip and generates four orthogonal body views and four head views from one orthographic input.It was trained on 6,000 scans using eight H100 GPUs for two days.Source: §4.1 · p. 6

4.2 Evaluation Benchmark

• Most participants preferred our method for appearance quality and text alignment, and it achieved SOTA CLIP and FID T2I metrics.Table 2 reports these comparison outcomes.Source: Table 2 · p. 6
• Gen-Schnell is compared with feed-forward multi-view methods, while Gen-HRes is compared with SDS-based avatar methods and Chupa.The evaluation also assesses the realism of generated InfiniHumanData identities.Source: §4.2 · p. 6

4.2.3 InfiniHumanData Evaluation.

• Real scan renderings received 746 votes and InfiniHumanData images 765, indicating users found the generated identities similarly realistic.The study asked participants to choose the more realistic image or “both” when they could not tell.Source: §4.2.3 · pp. 7–8

4.3 Fine-grained Controllability

• With initial Gaussian noise fixed, Gen-HRes preserves subject identity while reproducing input garments in try-on results.The passage describes high fidelity to clothing images and identity-preserving garment changes.Source: Figure 6 · p. 7, §4.3.1 · p. 8
• Gen-HRes follows SMPL-specified pose and body shape, while text inputs control attributes such as age, ethnicity, gender, and accessories.Fine-grained text editing maintains identity consistency.Source: §4.3.2 · p. 8, §4.3.3 · p. 8

4.4 Application

• Applications include extracting garments from photographs, reanimating meshes with SMPL motion data, and producing watertight meshes for freestanding 3D-printed figurines.Reanimation uses barycentric interpolation of SMPL skinning weights onto the generated mesh.Source: §4.4.1 · p. 8, §4.4.2 · p. 8, §4.4.3 · p. 8

4.5 Ablation Study

• Qualitative ablations examine orthographic multiview diffusion, scan-like image generation, additional SMPL fitting, and tolerance to inaccurate SMPL for children.The cited examples appear in Figures 14–16.Source: §4.5 · p. 8

5 Limitations and Future Works

• Gen-HRes can perform high-fidelity generation but remains slower than Gen-Schnell; Gen-Schnell cannot generate faithful facial details because its pretrained MVDream resolution is 256×256.Limited training resources prevented directly training a higher-resolution Gen-Schnell; the authors publicly release 768×768 InfiniHumanData with multi-modal labels.Source: §5 · p. 8
• GPT-4o refuses to identify unmatched famous-person samples for privacy reasons, while multi-view mesh carving can artifact textures in self-occluded regions.The authors suggest another vision-language model and data-driven mesh reconstruction as possible future directions.Source: §5 · p. 8

6 Conclusion

• InfiniHuman combines automated data generation with InfiniHumanData’s 111K diverse identities and InfiniHumanGen’s fine-grained, multimodally controlled avatar synthesis.The conclusion reports controls over appearance, shape, pose, and clothing and says the data, models, and pipeline will be released.Source: §6 · p. 8

A.1 InfiniHumanGen

• InfiniHumanGen examples show independent edits to identity, clothing, body shape, pose, and accessories using text, SMPL, and clothing-image inputs.Examples include changing clothing while maintaining identity and editing pose or shape with consistent appearance.Source: §A.1 · p. 13

A.2 InfiniHumanData

• InfiniHumanData examples pair clothing assets with head and body RGB views and show view-consistent generated images across subjects.The examples illustrate the dataset’s diversity and annotation richness.Source: §A.2 · p. 13

A.3 InfiniHumanGen vs. SOTA

• Compared with SDS-based methods, InfiniHumanGen produces higher visual fidelity, more consistent geometry, and better prompt alignment, avoiding their slow convergence and common visual artifacts.The reported artifacts include over-smoothed textures and structural inaccuracies.Source: §A.3 · p. 13
• InfiniHumanGen handles complex prompts and clothing-image conditioning that Chupa does not support, while Chupa accurately follows pose.The comparison describes Chupa as failing to generalize to complex text prompts.Source: §A.3 · p. 13
• Against feed-forward approaches, InfiniHumanGen shows stronger text following, more accurate pose and clothing reproduction, and improved realism.The passage contrasts these results with competitors’ geometry distortion, incorrect clothing, and weak text following for head avatars.Source: §A.3 · p. 13

A.4 InfiniHumanData vs. SOTA

• Compared with IDOL [Zhuang et al. 2025], InfiniHumanData produces sharper, more view-consistent renderings with smoother transitions and coherent shape, texture, and lighting.The passage links this consistency to downstream uses such as reconstruction, animation, and virtual try-on.Source: §A.4 · p. 13

B.1 Scan Captioning

• A structured GPT-4o prompt extracts gender, age, ethnicity, pose, clothing, and body parts directly from scan images.The same prompt is used for every sample to minimize hallucination and irrelevant details.Source: §B.1 · p. 13

B.2 Multi-granularity Text Generation

• GPT-4o expands each detailed scan caption into increasingly concise versions, from about 20 words to five words.A summarization prompt generates the caption sequence.Source: §B.2 · p. 13
• The multiple caption granularities let models learn from fine-grained and coarse cues and improve robustness to user input.Source: §B.2 · p. 13

• The pipeline generates several clothing candidates per subject and uses GPT-4o to select the best match.Source: §B.3 · p. 13
• GPT-4o compares candidates by color, texture, pattern, length, and details such as pockets and zippers.Source: §B.3 · p. 13
• If none of the candidates match the subject's clothing, GPT-4o returns “No”.Source: §B.3 · p. 13

• The orthographic multi-view diffusion model generates high-quality, view-consistent head views and full-body views.Source: Figure 18 · p. 15
• The appendix includes qualitative appearance and geometry comparisons with SDS-based and feed-forward text-to-3D avatar approaches.Source: Figure 19 · p. 16, Figure 20 · p. 17
• InfiniHumanData is reported to achieve better visual realism and multi-view consistency than the IDOL Dataset.Source: Figure 21 · p. 18
• OpenPose 2D joints refine SMPL pose parameters with reprojection error, resolving image-body misalignment, especially in face and hand regions.The alignment is described as crucial for SMPL-guided multi-view generation.Source: §B.4 · p. 22

C.1 Gen-Schnell.

• The modified MVDream input convolution expands from four to 12 channels, with added-channel weights zero-initialized to preserve pretrained behavior at initialization.Fine-tuning gradually teaches the model to use the new SMPL and clothing conditions.Source: §C.1 · p. 22
• InfiniHumanData received 765 user-study votes, compared with 746 votes for scan subjects.Participants assessed randomly paired dataset subjects and scan subjects for realism.Source: Figure 27 · p. 23

C.2 Gen-HRes.

• Gen-HRes conditions full-body image generation on a SMPL normal map, clothing image, and detailed text; it treats the map as spatially aligned and clothing as non-aligned.The design assigns dynamic positioning tokens of 0 to the normal map and -48 to the clothing image.Source: §C.2 · p. 23
• Gen-HRes uses a fine-tuned orthographic multi-view diffusion model to generate high-resolution head views directly, while PSHuman generates a low-resolution single head view and resizes it.The paper says direct high-resolution views allow greater head detail, including eyes, ears, and hair.Source: §C.2 · p. 23