Neural4D

What is Neural4D?

Neural4D is a cloud-based AI 3D content platform developed by DreamTech; it is used to convert text, photos, concept diagrams, and multi-angle references into editable 3D meshes.

In addition to text-to-3D and image-to-3D, it also offers PBR textures, model modification via dialogue, topology restructuring, automatic bone binding, animation, format conversion, 2D images, videos, and development interfaces.

Main functions

Generation of 3D from text and images

• Text to 3D accepts textual descriptions of objects, characters, materials, and styles, creates a basic mesh, and then allows the user to decide whether to proceed with generating textures.
• Image to 3D allows you to upload a single photo, a sketch, or multiple views of the same object; the system then infers the invisible surfaces and creates a complete model.
• Product images and concept diagrams with high contrast, clear subjects, even lighting, and simple backgrounds make it easier to obtain usable geometry; inputting data from multiple angles reduces the need for speculation about the back side.
• The processing of the basic grid usually takes around 90 seconds, while PBR textures require additional processing, so the total time may exceed two minutes; queue and task complexity can affect the waiting time.
• The resulting output may still suffer from issues related to proportions, internal structure, thin walls, topology, and texture seams; therefore, the claim of being “printable” cannot replace thorough checks of the slicing and structure.

PBR texture and format export

• AI Texture can extract color, pattern, and material characteristics from reference images, and generate PBR channels such as Normal, Roughness, and Metallic.
• Automatic UV unwrapping and material mapping are used to apply textures to surfaces; they are suitable for game props, product previews, and architectural concept materials.
• It supports OBJ, FBX, GLB, USDZ, STL, and BLEND – formats that are suitable for general editing, game engines, web and AR applications, printing, as well as Blender.
• Paid users can download models with complete textures, while free users can mainly download texture-free base meshes, and their use is limited to non-commercial purposes.
• Format compatibility does not mean that no adjustments are needed after import; Unity, Unreal, Maya, Blender, and slicing software may still require changes to the coordinate axes, units, materials, and number of faces.

Neural4D-2o interactive model modification

• Neural4D-2o retains the context of the current session within the chat interface, allowing successive requests to change the scale, local geometry, color, or surface texture.
• Users can submit both textual and visual references at the same time; for example, they can start by creating a basic model from a sketch, and then request that certain parts be extended, the style be changed, or the printing base be made thicker.
• Context memory is reduced, eliminating the need to regenerate everything from scratch; however, changes made to the natural language text may still affect unspecified areas. Critical dimensions should be verified using external modeling tools or CAD software.
• Studio offers conversation sessions, while the standard API does not support similar session functionality at the moment; the corresponding v2.0 interface is still part of future plans.

Redo topology, automatic binding, and animation

• Remesh allows you to set a target number of faces for the assets generated within Neural4D and to re-topologize them; it is useful for reducing the load on real-time applications or for preparing those assets for subsequent binding.
• Currently, Remesh does not accept external third-party meshes, and therefore cannot be regarded as a general online topology-reconstruction service.
• Auto-Rigging can identify the bone positions of humans, quadrupeds, and stylized characters, calculate skinning weights, and after exporting in FBX or GLB format, they can be integrated into the animation pipeline.
• The standard bone naming is used to ensure compatibility with Mixamo, Unity Mecanim, and Unreal redirection, but special limbs, costumes, tails, and complex topologies may still require manual weight adjustments.
• Animation is used to add movements to bound assets; in the free version, conversion, topology reshaping, binding, and animation are marked as Trial features, while full access is available only with a paid account.

Images, videos, and a unified agent

• Studio also includes text-to-image generation, AnimeArt, and video creation from text or reference images, making it suitable for developing dynamic presentations from concept art.
• You can choose between Veo 3.1, Seedance 2.0, Seedance 2.0 Fast, and Grok Imagine for the video; the resolution of the model varies, as does the quality of the original audio.
• Video settings include options from 480P to 4K, common aspect ratios, video lengths of 5 to 15 seconds, audio settings, and the number of videos to be generated at once; the output format is MP4.
• AI 3D Agents enable the generation of 3D models, images, or videos based on inputs provided from a fixed control panel; however, different media types are processed using their own respective models, settings, and power consumption levels.
• The names of third-party video models, the regions in which they are available, their content policies, and prices may change; the actual options available in an account take precedence over any static descriptions.

Input and output specifications

Complete usage process

1. Create a free account, decide whether the results will be used for personal testing or for commercial projects, and choose between Public or Private – the latter being available only with a paid account – before generating them.
2. When generating 3D from text, it is necessary to describe the contours, proportions, material, and intended use; when generating 3D from images, a clear subject should be used, along with references to the front, side, and back views whenever possible.
3. First, create a basic grid; use the online viewer to rotate it in order to examine the back side, bottom, thin walls, holes, proportions, and outline, and then decide whether to use Power to generate PBR textures.
4. Using Neural4D-2o, specific instructions can be used to make local modifications; when it comes to printing, the minimum wall thickness and dimensions must be specified, while for games, the number of faces and the binding requirements need to be indicated.
5. Run Remesh when real-time assets are needed, then use Auto-Rigging to bind the character and test joint bending, clothing interaction, and range of motion.
6. Select GLB, FBX, OBJ, USDZ, STL, or BLEND based on the target tool, and import it into Blender, Unity, Unreal, or slicing software for final technical verification.
7. Record the power consumption for each step, any refunds due to failures, and the settings used; before starting mass production, use a small number of samples to assess the cost associated with manual cleanup.

Power quota and billing logic

• According to the current English description of Free, 50 Power points are added each week, and only 1 parallel task is allowed; on pages in other languages, it was stated that 300 points per month could be obtained and that more parallel tasks were possible, but the information is inconsistent.
• The English homepage indicates that both Text to 3D and Image to 3D processes require 10 units of Power, but some localized versions state 15 or 20 units; the value shown before submitting the task should be taken as the correct one.
• Texture processing, format conversion, topology reshaping, binding, animation, as well as handling of images and videos all require additional power; the cost of video processing is further influenced by the model used, resolution, duration, audio elements, and the number of videos.
• Subscribers can top up their balances at a public price of $1 per 100 Power units; other accounts or promotions may have different prices.
• If there is a server error and no usable files are generated, the process will automatically return to Power; no refund is provided if the user cancels the process midway or is unsatisfied with the final result.

Price packages

Neural4D offers a free version, as well as 7-day passes, monthly and annual Pro subscriptions. The listed prices include both promotional rates and the regular prices; the actual discounts, taxes, and quotas are subject to those indicated on the payment page.

• The end date of the promotion has not been clearly specified; weekly, monthly, and annual subscriptions may be automatically renewed according to their respective renewal cycles.
• A yearly fee of 358.8 dollars does not exactly correspond to the figure of around 6.9 dollars per page; the actual amount to be paid should be based on the total settlement amount and the renewal cycle.
• The free quota, the price per Power unit, and the number of parallel tasks vary from one language and page to another; therefore, none of these versions can be considered as a uniform benefit applicable to all regions and accounts.

Subscription, cancellation, and refunds

• Subscriptions are automatically renewed at the end of the initial period for the same duration and at the then-current price; the terms require cancellation through Plans and Billing at least 30 days before renewal.
• After cancellation or downgrading, it can be used until the end of the current period; generally, the remaining fee for that period is not refunded.
• The terms also allow for cancellation and a refund of the fees for services that have not yet been provided, within 30 days of the start of the service or of the renewal period; this constitutes a conditional difference compared to the general rule that no refunds are given.
• Before making the payment, you should have the support team confirm whether the weekly pass allows for cancellation 30 days in advance, understand the method of refund calculation, and see what the price is for renewal as part of a promotion.
• No refund will be provided in the event of termination due to violations, and the content that has been published in public areas may remain available even after the account is terminated.

API integration

• The Neural4D API utilizes Bearer Tokens; it allows for the generation of keys from an account, and offers interfaces for creating 3D models from text or images, as well as for performing queries and format conversions.
• The format conversion interface can generate FBX, GLB, OBJ, STL, BLEND, or USDZ files based on the asset’s UUID; it is stated that 10 points are consumed for each conversion.
• The API uses an asynchronous workflow: the application should submit a task, save the UUID, poll for status, or wait according to the specified mechanism, and then request a download once the operation is successful.
• The target generation time for standard interfaces is less than 90 seconds, but there is no confirmation regarding publicly available SLAs, rate limits, retry guarantees, or data processing guidelines.
• The conversation state of Neural4D-2o has not yet been integrated into the standard API, so it is not possible to assume that the same session can be edited continuously using the existing interfaces.

API integration steps

1. Create a dedicated API Token within the account, store it in the server-side key system, and manage it separately for testing and production environments.
2. Select the endpoint for generating text or images, verify the file format, size, required fields, and necessary computing power; first test it with non-sensitive samples.
3. After submitting a task, save the UUID; use the query interface to handle situations such as queuing, success, failure, timeout, and insufficient balance.
4. Upon success, the target format conversion and download are requested; the integrity of the file is verified, and the original input data, relevant messages, the model version, as well as the status of the commercial license are recorded in the asset database.
5. Set limits on usage volume, implement retry mechanisms in case of failures, generate alerts when failures occur, and carry out key rotation. Before rolling out the solution on a large scale, it is necessary to obtain confirmation from sales teams regarding the rate, pricing, and scope of support.

Suitable for users and scenarios

• Indie games and studios: Quickly create props, basic character meshes, PBR materials, and bindable assets from concept art.
• 3D printing creators: Convert sketches or photos into STL files, and then use slicing software to check the wall thickness, supports, dimensions, and integrity of the model.
• E-commerce and XR teams: Generate GLB or USDZ previews of products, but manual calibration is required for their actual size, color, and structure.
• Film, television, and animation teams: use automatic binding and animations for preliminary testing, followed by fine-tuning in Blender, Maya, or gaming engines.
• Developers: Integrate image-to-3D, text-to-3D, and format conversion functions into asset pipelines, custom tools, or user applications via APIs.
• Researchers: They use publicly available Direct3D and Direct3D-S2 code along with relevant weights to study graph-based 3D generation, without relying on commercial studios.

Advantages and capabilities boundaries

• The advantage is that it integrates generation, dialogue modification, textures, topology restructuring, binding, animation, images, video, and API integration into a single asset workspace.
• Six common formats are available for games, web AR, DCC, and printing; however, complex assets still require professionals to examine their topology, UV mapping, skeleton structure, and physical manufacturability.
• AI can infer the back and internal structures that are not visible in the input, but generating a single image does not guarantee an accurate representation of a product, a person, or an engineering component.
• Capabilities such as “water resistance,” “direct printing,” and “no need for repair” should be verified as target characteristics; however, actual printing is still affected by factors such as model scale, thin walls, overhangs, and slicing settings.
• High-density triangular meshes may not be suitable for deformation animations, CAD parameter editing, or real-time rendering on mobile devices; they require topology restructuring and manual refinement.
• Videos and third-party models are deployed quickly; the current model, clarity, audio quality, availability by region, and Power costs may change at any time.

The difference between open-source models and commercial products

• Neural4D Studio, the cloud API, accounts, the community, billing, and the complete production pipeline are proprietary hosted services; they cannot be considered fully open source just because the related research code is made public.
• DreamTechAI has made the Direct3D inference code and models available; the code library is licensed under the Apache 2.0 license.
• Direct3D-S2 also makes the inference code and models available; the code library is licensed under the MIT license. Inference at a resolution of 1024 requires approximately 24 GB of video memory.
• Open-source repositories mainly focus on research related to 3D generation from images; they do not contain the complete source code for Neural4D-2o, cloud-based texture processing, binding mechanisms, video handling, billing systems, or APIs.
• When deploying it manually, it is also necessary to verify the model weights, dependencies, training data, and the licenses for third-party components, as it is not sufficient to only consider the license for the main code.

Copyright, Accessibility, and Commercial Use

• The terms of service do not claim ownership over the content entered and submitted by users, but users must ensure that the images, messages, and other materials they upload have legitimate rights.
• Paid Pro or enterprise users receive commercial rights to use these generation models without the need for attribution, and they can be applied to legitimate projects such as games, films, products, and NFTs.
• Free assets are intended solely for personal, non-commercial testing; they are created by default in the Public category, can be made available in the community, and allow other users to view or remix them.
• The assets presented in the public Showcase are available under the CC BY 4.0 license; when used for commercial purposes, credit must be given to the original creator. This is different from the right to use assets created privately by paid users without any attribution required.
• Choosing Public will allow the platform to display the generated content, while choosing Private will make it available only to those with a paid subscription; sensitive prototypes and customer projects must have their settings confirmed before submission.
• Similar outputs, trademarks, character portraits, and protected designs can still give rise to disputes; the commercial license provided through subscription cannot replace a thorough assessment of third-party rights.

Privacy, training, and data retention

• The platform handles accounts, contacts, transactions, devices, networks, activity logs, as well as images, videos, audio files, messages, works, and related metadata uploaded by users.
• The terms specify that the input and output data can be used for tagging, classification, content moderation, and model training and improvement, granting long-term, global, transferable, and sublicensable usage rights.
• Users who do not wish for their confidential assets to be involved in such processing should not rely solely on Private visibility; instead, they must obtain commitments regarding the exclusion of such data from training processes and regarding its handling through corporate contracts.
• The service shares the information needed to achieve its objectives with hosting, IT, support, payment, marketing, analytics, related parties, and business partners.
• The privacy policy states that data may be processed and stored in the United States, but the dates and locations still contain placeholder text; companies should request confirmation regarding the actual legal entities, data centers, and cross-border mechanisms.
• Closing an account results in the deletion of the assets associated with it, but the content in public areas may remain intact indefinitely; the retention period for personal information depends on its purpose, applicable laws, as well as considerations related to disputes and risks, and there is no fixed number of days for this retention.
• Users can update their information in their accounts and request that they be closed or deleted; technical, organizational, and physical measures are employed to reduce risks, but the service does not guarantee absolute security.

Summary

Neural4D is suitable for creators and development teams who need to obtain 3D assets quickly from text or images, in order to proceed with tasks such as adding textures, modifying models, binding elements, creating animations, exporting files, or integrating APIs.

Before official adoption, it is necessary to verify the geometry and printing quality using one’s own samples, check the real-time power usage and billing details, and specifically assess the rules regarding free access, commercial licensing, training permissions, as well as any conflicts regarding page quotas.