Hugging Face published a TRELLIS.2 fine-tuning guide on November 12, 2026, showing how to adapt the existing image-to-3D model using a collection of 3D assets. For technical artists and small studios, it offers a concrete starting point for tackling recurring shape problems—not a new model launch.
The strongest case for a small pilot is a reusable mesh library, a specific geometric defect and enough separate designs to evaluate the result. A one-off asset request is harder to justify: there is less opportunity to recover the preparation and evaluation work.
Start with meshes, not just product photographs
This training recipe needs 3D meshes as reference geometry: files that describe an object’s surfaces. A photo catalogue alone is insufficient. The example dataset pairs mesh-derived training data with 16 rendered views per object. Images guide generation, but the meshes provide the shapes to learn from.
LoRA trains small additions, known as adapters, while keeping the existing model’s original weights fixed. The guide adapts coarse structure and detailed shape; the texture generator is left unchanged.
That distinction should decide whether this recipe belongs on your shortlist. If the recurring problem is colour or material appearance, this experiment does not demonstrate a solution. A better-shaped object and a better-looking surface are separate goals.
What the nine-object result establishes
The authors report using 188 footwear objects: 179 for adapter training and nine held out for evaluation. On those nine, their mean geometric F-score rose from 0.276 to 0.803. This measures surface similarity at a chosen tolerance; it does not mean 80.3% of generated assets were production-ready.
The published evaluation code uses rendered input views and rescales each mesh before comparison. The result therefore cannot certify real-world dimensions or manufacturing fit. Nor does this small footwear experiment establish benefits for other object categories, ordinary product photos or animation-ready mesh structure.
For your own library, reserve entire designs and their variants for evaluation. The repository’s default split groups variants using naming rules designed for Google Scanned Objects; other filenames can receive a random split. Check that two versions of one design do not end up on opposite sides. This is a precaution for a new dataset, not a finding of leakage in the authors’ experiment.
About $29.50 in hardware is not a project budget
The authors report roughly one hour of preparation, followed by 7.2 hours of coarse-structure training and 3.6 hours of detailed-shape training, using A100 80 GB hardware.
As of November 12, Hugging Face Jobs lists that A100 at $2.50 an hour. Applying that rate to one pass through the reported stages gives this illustrative calculation:
(1 + 7.2 + 3.6) GPU-hours × $2.50 = approximately $29.50.
That is a calculation from the authors’ timings, not a quote or a measured bill. It excludes additional startup and download time, evaluation, retries, storage, labour, integration and applicable taxes. Jobs bills by the minute during Starting and Running states, so charged time can exceed the duration of training itself.
The maintainers describe the 24 GB configuration as smoke-tested on an L4—a short check that it runs. It is not evidence of a completed equivalent experiment. Do not substitute the L4’s $0.80 hourly rate into the A100 timings: a full L4 runtime and comparable result are not established.
Clear the access and data requirements first
Training also needs access to Meta’s gated DINOv3 image encoder, a separate model that processes the input images. Its own licence and access conditions apply; access is not supplied merely by downloading the training software. Resolve that dependency before paying for a long run.
For the cloud route, Jobs is available to users or organisations with a positive credit balance. A paid subscription is not the only entry route. The companion software also supports local Docker execution on suitable hardware.
For confidential assets, check privacy separately for source meshes, prepared datasets, adapter repositories and the tracking dashboard. The repository documents separate privacy controls; private inputs do not automatically make every output private. Confirm that your assets may be used for the intended training, too.
Decide what would make a pilot worth continuing
Before committing to a full run, turn the recurring defect into an acceptance decision. For a footwear project, that might mean judging the sole against the reference mesh, rather than choosing whichever whole-object render looks more attractive.
Compare like with like. Use the same evaluation inputs and generation settings for the original and adapted models. The supplied evaluation script uses matched random seeds. Inspect the target defect alongside the numerical scores.
Check the downstream task. If dimensions, animation or simulation matter, assess those requirements separately. A shape-similarity score is not an acceptance test for all of them.
Measure the work that remains. Record preparation, review and manual cleanup time. Compare the total effort with your existing workflow across the volume of assets you actually expect to produce. The published experiment supplies no measured labour saving to assume in advance.
Continue only if the adapted outputs improve the defect on designs excluded from adapter training and the remaining work makes the overall cost worthwhile. If the original model already meets the brief, custom training adds a project without a demonstrated benefit.