Pixal3D 3D AI Guide

3D AI turns text and images into 3D assets without a modeling suite. Compare Pixal3D, Trellis 2, and Hunyuan Motion, then generate a model on pixal3d.ai.

What "3D AI" covers

"3D AI" is the umbrella term for models that produce three-dimensional output from a text prompt or a source image. Some convert a single image into a textured mesh. Some generate motion from text. Some specialize in reconstructing a photo into a scan-like asset. They share one promise: you skip manual modeling and reach a reviewable draft much faster.

This page is the category guide. It answers which lane to choose and what each lane can produce. If you already know you want the image-to-3D lane and only need the steps, read the AI 3D generator workflow instead; if your question is specifically about how the image stays aligned to the surface, read 3D pixel. Preparing the input itself is covered by the photo to 3D model guide for photographs and the 2D image to 3D model guide for illustrations and renders.

How 3D AI generation works

Every lane on this site runs a research model behind a hosted interface. The three mechanisms below are genuinely different, which is why the output differs too.

Image-to-3D with pixel back-projection

The upstream Pixal3D project generates assets from a single image and, in its own words, "explicitly lifts pixel features into 3D through back-projection" instead of loosely injecting image features through attention. That direct pixel-to-3D correspondence is what keeps the visible side of the result close to your photograph. The official repository keeps two implementations: main is the current TRELLIS.2-based version, and paper is the Direct3D-S2-based version that produced the published results. See the official TencentARC/Pixal3D repository.

A second image-to-3D read

The Trellis 2 lane takes the same kind of input through a different implementation, so it can read thin parts and structure differently from Pixal3D. That is the whole reason to run it: a second geometry interpretation of the same source image rather than a better or worse model in the abstract. The upstream Pixal3D README describes its own main branch as the TRELLIS.2-based version, and the Trellis 2 baseline is documented by Microsoft's TRELLIS.2 repository.

Text-to-motion

The Hunyuan Motion lane is not image-to-3D at all. It reads a text description of movement and returns FBX plus motion JSON instead of a static mesh, which is why it appears under text to motion in the workspace rather than under image to 3D.

The three lanes in one workspace

The AI 3D Generator puts all three behind one authenticated task pipeline, one history, and one credit ledger.

Lane	Input	Output	Use it when
Pixal3D (`fal-ai/pixal3d`)	one image	GLB with PBR textures	you want the pixel-faithful default for a product shot, prop, or character
Trellis 2 (`fal-ai/trellis-2`)	one image	GLB	you want a second read on the same image before approving the result
Hunyuan Motion (`fal-ai/hunyuan-motion`)	text prompt	FBX + motion JSON	you need movement, not geometry

20 criteria for choosing a 3D AI path

Each row isolates one decision factor. Use them to pick between the image-to-3D lanes and the text-to-motion lane before you queue a job.

#	Criterion	What it decides
1	Input type	A source image routes to image-to-3D; a text description alone can only drive motion or prompt-led generation
2	Desired output	Need a mesh? Choose GLB. Need animation? Choose the FBX motion lane
3	Geometry fidelity	Pixel-aligned reconstruction preserves the source silhouette better than a generic depth lift
4	Texture quality	PBR albedo from the image is included on the Pixal3D GLB lane
5	Viewpoint control	The model output is anchored to the reference view, so pick a reference that matches your final pose
6	Model selection	Compare Pixal3D and Trellis 2 outputs when the first pass is not review-ready
7	Motion intent	Text-to-motion is for movement, not for building a static mesh
8	Credit budget	Each lane has a different credit range; read the in-workspace estimate before submitting
9	Turnaround time	Longer tasks queue on the provider; do not refresh away from an in-flight job
10	Output format	GLB is the delivery format for 3D; FBX and motion JSON are the motion outputs
11	Downstream engine	Confirm the target (Blender, Unity, Unreal) can import GLB before you commit to a run
12	Source delivery	A public direct image URL is fetched by the provider; a file you upload is stored and used the same way. A login-walled or HTML page link fails
13	Subject complexity	A single well-lit subject beats a busy multi-object photo on the first try
14	Texture resolution	The texture-size setting caps the maps the run can produce; a soft source stays soft at any size
15	Prompt specificity	For motion, a concrete action beats a vague one and reduces retries
16	Review surface	Prefer a lane that shows an in-page preview so you can reject before downloading
17	Re-roll cost	Factor in the cost of a second run when the first output is close but not perfect
18	License awareness	Upstream Pixal3D is released under the MIT License; check the provider's terms before shipping generated assets commercially
19	Account history	Use a workspace that keeps task history so you can re-download without regenerating
20	Fallback plan	If the hosted path errors, verify the source image and retry before changing the lane

Formats and limits

The hosted workspace is narrower than the upstream repositories, and it is worth knowing exactly where the edges are before you spend credits.

Input: one source image per job. Uploads accept PNG, JPEG, WebP, AVIF, and GIF up to 10 MB, and a public direct image URL works without an upload.
Output: GLB with geometry and PBR textures for the image-to-3D lanes; FBX plus motion JSON for the text-to-motion lane.
Hardware: none on your side. The models run on the provider, so no local GPU is required.
Not available here: multi-view conditioning, local weights, and GGUF builds. The upstream repository released multi-view inference code in September 2026 for people running it locally, and its official model files are .safetensors checkpoints rather than a GGUF distribution.

Honest comparison: this site vs the upstream project

pixal3d.ai is an independent hosted service. It is not affiliated with the Pixal3D research team, and it runs third-party provider infrastructure rather than the research codebase.

Question	Upstream repository	pixal3d.ai workspace
Setup	Linux, CUDA, compiled dependencies	browser sign-in
Hardware	starts from the TRELLIS.2 official baseline	none required
Control	full code, weights, and configuration	managed lanes and settings
Multi-view	yes, via the repository's multi-view script	no, one source image per job
Best for	research and reproducible local pipelines	evaluating a lane and shipping a GLB

Getting the most out of 3D AI

Start with the lane that matches your input, not the one that is newest. A photo in hand points to image-to-3D. A movement you want to test points to text-to-motion. After generating, use the free GLB Inspector to verify the mesh before import, and read Pricing to plan credit usage across multiple runs.

Where to go next

Each of these pages takes one part of the 3D AI workflow further than a category guide can:

Deciding what to make: 3D object creator covers which single-object images convert predictably, and 3D asset generator covers the handoff checklist for a pipeline asset.
Working with the output: image to mesh explains what the GLB geometry contains, and AI generated 3D models is the review checklist before you use a result.
Non-mesh output: text to motion covers the Hunyuan Motion lane, which returns FBX motion and motion JSON instead of geometry.
Manufacturing: 3D printing AI follows the route from one photo through GLB and STL to a slicer.
Budget and provenance: image to 3D model free states exactly what the signup credits cover, and open source image to 3D separates the public research models from this hosted service.

FAQ

What counts as 3D AI?

3D AI is the umbrella term for models that turn an image or a text prompt into 3D output. On pixal3d.ai that is the Pixal3D and Trellis 2 image-to-3D lanes plus the Hunyuan Motion text-to-motion lane, all reached through one workspace and one credit ledger.

Which 3D AI lane should I run first?

Start from your input, not from the newest model. A single photo points to the Pixal3D image-to-3D lane, a second read on that same photo points to the Trellis 2 lane, and a text description of movement points to Hunyuan Motion, which produces motion instead of a static mesh.

What files does 3D AI output here?

The image-to-3D lanes return a GLB mesh with PBR textures that you preview and download in the browser. The text-to-motion lane returns FBX plus motion JSON instead of a mesh.

Do I need a GPU to run 3D AI?

Not for the hosted workspace, which runs the models on the provider side. Running the upstream Pixal3D repository locally does need a Linux CUDA environment, and the official installation starts from the TRELLIS.2 baseline documented in the upstream README.

Is 3D AI the same thing as the AI 3D generator?

3D AI is the category; the AI 3D generator is the hosted workspace that runs it on pixal3d.ai. This guide covers which lane to choose, and the AI 3D workflow page covers the steps from a source image to an exported GLB.