TEXT-CONTROLLED 3D PART SEGMENTATION

FlexPartText-Controlled Semantics-Aware 3D Part Segmentation
at Flexible Granularity

Yiqun Zhao1,2*Jiachen Xu2Yihao Zhi3,2*Yunfei Zhao2Zibo Zhao2Xintong Han2Chunchao Guo2†Shenghua Gao1,4†
1 The University of Hong Kong2 Tencent Hunyuan3 The Chinese University of Hong Kong, Shenzhen4 Shenzhen Loop Area Institute

* Work done during an internship at Tencent Hunyuan.   † Corresponding authors.

Paper PDFVideo 1080pExplore results

One model. Different prompts. The partition you ask for.

WHAT YOU ASK FOR IS WHAT YOU GET

Your mesh.
Your parts. Your level of detail.

A 3D asset admits many useful decompositions. FlexPart lets text specify both how finely to divide a mesh and which semantic parts to extract. A control prompt lists part names with location and shape descriptions; the model assigns discrete part identities directly to the mesh faces.

One segmentation model supports manually authored prompts, automatic coarse-to-fine partitions, and selective extraction. An optional visual planner completes the prompt from mesh renderings and a simple user request.

01 / METHOD

Text controls a native 3D part-identity field.

A shared representation connects geometry, semantic part descriptions, and the desired partition.

FlexPart method from Figure 2: a part-aware VAE encodes position, normal and part IDs; a text-conditioned diffusion transformer generates the part-identity field from geometry and control prompts
Method overview. A part-aware VAE represents geometry and part identity. A diffusion transformer conditions on geometry and encoded text to generate the part-ID field; querying mesh face centers yields the segmentation. Full and selective decompositions differ in their control prompts, with an explicit background item for selective inputs. Figure 2 in the paper ↗

Describe the partition. Part names, locations and shapes specify the desired label space.

Generate the identity field. Text and geometry jointly condition the same diffusion model.

Label the original mesh. Face-center queries assign discrete part IDs directly in 3D.

02 / TEXT CONTROL

Three ways to ask. One model to answer.

Geometry + a control prompt → FlexPart → a 3D part partition.

Unsegmented off-road kart
Input geometry
Control prompt

An off-road kart with 14 parts:

1. seat <location> <shape>
2. axle <location> <shape>
3. roll cage <location> <shape>
…

FlexPart
The same kart partitioned into colored semantic parts
Requested part partition

Write the part inventory for precise control. Semantic names, locations and shapes tell the model what each label means.

Full paper teaser showing the three FlexPart workflows in the original paper layout
The three workflows in the paper teaser. The visual planner writes prompts; FlexPart performs the 3D segmentation.
03 / EXPERIMENTS

Semantic parts, at the requested scale.

Selected main-paper results on full, multi-granularity and selective segmentation.

Part segmentation on PartObjaverse-Tiny

Recover the requested semantic parts across diverse object shapes.

Figure 3: four PartObjaverse-Tiny examples comparing GT, FlexPart-GT, FlexPart-Auto, SegviGen-GT, SegviGen-Auto, P3-SAM, PartField, SAMPart3D and Find3D
Full-partition comparison. “-GT” uses annotated conditions: part descriptions for FlexPart and a rendered guidance map for SegviGen. “-Auto” uses planner-completed prompts or predicted maps; other baselines use their native inputs. Figure 3 ↗

Planner-completed prompts achieve 0.681 mIoU, compared with 0.594 for SegviGen supplied with ground-truth guidance, under the paper’s watertight evaluation protocol. Table 1 ↗

Multi-granularity control on FlexPartBench

Change the item list to move from broad assemblies to fine semantic parts.

Figure 4: granularity-control comparisons; each method block shows a large fine partition and smaller coarse and medium partitions
The same geometry, different prompts. Within each method block, fine is the large panel, coarse is at the upper right, and medium is at the lower right. Baselines receive the inputs described in the paper, including ground-truth part counts for PartField and guidance maps for SegviGen. Figure 4 ↗

FlexPart scores 0.763 / 0.635 / 0.511 mIoU at coarse / medium / fine granularity, averaging 0.637 on the held-out benchmark. Table 2 ↗

Selective segmentation and design choices

Extract the named parts and assign the remainder to one background region.

Figure 5: selective segmentation at K=4 and K=8 comparing FlexPart with annotated or planner prompts against SegviGen and P3-SAM with clicks and Find3D with names
Selective segmentation. The two FlexPart columns use the same model with annotated or planner-written descriptions. Click-driven baselines receive points inside the requested parts; Find3D receives names. Unrequested parts share one background region. Figure 5 ↗
Figure 6: ablations comparing part-ID prediction with RGB-palette targets and training without background dropping
Why part IDs and background dropping? The paper’s ablations illustrate fragmentation with RGB targets and reduced accuracy without drop-to-background training. Figure 6 ↗
04 / GENERALIZATION TO AIGC ASSETS

Flexible parts on generated geometry.

Four Hunyuan3D 3.1 assets from the paper, each shown at coarse, medium and fine granularity.

Coarse → Medium → Fine

Tracked machine

Coarse Medium Fine

Sci-fi blaster

Coarse Medium Fine

Armored character

Coarse Medium Fine

Winged robot

Coarse Medium Fine

Three granularities, side by side. A fixed camera and synchronized model rotations make the parts easy to compare.

05 / PAPER

FlexPart

Text-Controlled Semantics-Aware 3D Part Segmentation at Flexible Granularity

Read the paper Download video

what you ask for
is what you get