main is built on TRELLIS.2 and adds pixel-aligned projection conditioning. Compare the incremental conditioning, output quality, and operating cost—not two unrelated stacks.Hosted generation: input, output, and credits
Pixal3D.ai is an independent service running both models through Fal. Sign in, upload one image, choose Pixal3D or Trellis 2 in the model selector, and download the completed GLB.
| Item | Pixal3D | Trellis 2 |
|---|---|---|
| Input | One image | The same image |
| Deliverable | Textured GLB | Textured GLB |
| Default job credits | 220 | 160 |
| Credits across supported settings | 180-390 | 160-310 |
| Wait time | Depends on queue and settings | Depends on queue and settings |
Defaults are 1024 resolution, 2048 textures, and Balanced settings. Two default jobs total 380 credits and are charged separately. Higher texture, triangle, or token budgets can raise the cost; check the amount beside Generate.
How to judge the output for your task
Browser-only, private
Free GLB Inspector
Check GLB and glTF model structure locally before import, review, or handoff.
Free GLB Inspector
Runs locally in your browser, no upload. Load it to start.
Open the generator and choose a model · Inspect a downloaded GLB for free · Read the hosted and local workflow guide
The relationship most comparisons miss
TRELLIS.2 provides the 4B generative pipeline and O-Voxel representation. Current Pixal3D uses that foundation and explicitly back-projects image features into 3D. The fair question is whether that extra alignment improves fidelity for your input enough to justify the additional setup and model path.
| Dimension | Pixal3D main | TRELLIS.2 |
|---|---|---|
| Relationship | Current main is built on TRELLIS.2 | Upstream base project |
| Conditioning | Adds explicit pixel back-projection | Image-conditioned 4B DiT pipeline |
| Representation | View-aligned O-Voxel pipeline | O-Voxel sparse structured latent |
| Output | GLB with detailed geometry and PBR textures | GLB with PBR surface attributes |
| Official local baseline | Inherits TRELLIS.2 setup; low-VRAM mode exists | Linux + NVIDIA GPU with at least 24GB |
| Published timing | No universal consumer-GPU promise | 3s/17s/60s on H100 at 512³/1024³/1536³ |
| License | MIT; third-party terms still apply | MIT; third-party terms still apply |
Local benchmarks and hosted wait times
Run one controlled A/B before standardizing
Use the identical source image
Do not compare different crops, masks, or background removal.
Record branch and commit
Pixal3D main and paper are different implementations.
Record resolution and memory mode
A 1024 low-VRAM result is not the same test as a 1536 standard result.
Measure peak VRAM and wall time
Warm-up, downloads, and post-processing should be labeled separately.
Grade the full asset
Inspect hidden geometry, PBR channels, topology, file size, and destination import.
Practical selection
| Situation | Start with | Reason |
|---|---|---|
| You need the upstream baseline | TRELLIS.2 | It isolates the base model before Pixal3D conditioning |
| Visible details must align closely to the source | Pixal3D | Pixel back-projection is designed for direct correspondence |
| You are reproducing the SIGGRAPH paper | Pixal3D paper branch | The paper results belong to that branch |
| You have less than the official hardware baseline | Hosted evaluation | Validate value before maintaining an unsupported local port |
| You need a production decision | Both, one A/B harness | Use measured output and operating cost, not a generic winner |
How to read the credit rows
Three numbers appear above and they answer different questions. The default row is what one job costs at 1024 resolution with 2048 textures — the pair you get when you change nothing. The range row covers the supported settings, from the cheapest 1024/1024 combination to the most expensive 1536/4096 one, plus any decimation or token surcharge. The sum is what one fair comparison costs: one job per model on the same image, charged separately.
- Compare models on identical settings. "Cheaper" usually means "run at lower texture size", not a better price for the same job.
- Texture size is the lever that moves cost, so choose it from where the asset is used, not from the number at the top of the menu.
- Budget two jobs per comparison and read the actual charge from the task record afterwards; surcharges exist and are not guesses you can make in advance.
Wait time is the other number people quote, and it is the one neither model controls: queue depth does. The published TRELLIS.2 timings — 3, 17 and 60 seconds at 512³, 1024³ and 1536³ on an H100 — describe the model on dedicated hardware, not a hosted job at peak. Record your own wait from the task record and compare that, because a benchmark on someone else’s GPU is not a service-level promise.
Frequently asked
Is Pixal3D the same as TRELLIS 2?
No, but they share a foundation. The current Pixal3D main branch is built on TRELLIS 2 and adds explicit pixel back-projection conditioning. TRELLIS 2 is the upstream baseline; Pixal3D layers image-to-3D alignment on top of it.
Can I compare wait times from a review?
Only as a rough upper bound. Queue depth, region, and the settings chosen by the reviewer move the number more than the model does, and published H100 timings measure dedicated hardware. Run both once yourself on the same image and record the wall time from your own task records.
Which model should I choose for one photo?
If visible details must stay closely aligned to the source photo, start with Pixal3D for its pixel back-projection. If you need the upstream baseline, start with TRELLIS 2. For a production decision, run both through the same A/B harness and compare measured output and cost instead of a generic winner.
Can I run both locally on a consumer GPU?
The official TRELLIS 2 baseline is Linux with an NVIDIA GPU of at least 24 GB, and Pixal3D inherits that setup with a low-VRAM mode that has no guaranteed 6 GB minimum. The published timings were measured on an H100 and do not predict an RTX run. A hosted route removes the hardware requirement.
Do Pixal3D and TRELLIS 2 cost the same on Pixal3D.ai?
Both run through Fal and are charged per job in credits, with default jobs billed separately. Defaults are 1024 resolution, 2048 textures, and Balanced settings; higher texture, triangle, or token budgets can raise the cost, so check the amount beside Generate.