Results · CSTD Route

Table 1 — mIoU and ARI

**Autonomous texture partitioning across four benchmarks.** Every method receives the same instruction and is told nothing about how many regions the image contains. No mask is completed by taking a complement, no proposal set is truncated to a known count, and images for which a method emits no mask are scored as zeros rather than dropped. Since the region count is withheld from every method, these numbers are not comparable to evaluations that supply it.
Route	Metric	LISA	GSVA	GLaMM	Sa2VA	Detecture
RWTD	mIoU	0.7541	0.3706	0.6225	0.5452	0.7843
RWTD	ARI	0.4413	0.4395	0.4499	0.4569	0.5993
RWTD-COCO	mIoU	0.6837	0.4099	0.6014	0.2617	0.8460
RWTD-COCO	ARI	0.3364	0.5001	0.4222	0.2362	0.7351
TextureADE	mIoU	0.5752	0.2010	0.3840	0.3092	0.7222
TextureADE	ARI	0.3869	0.2402	0.1851	0.1970	0.6863
CSTD	mIoU	0.8394	0.6578	0.7577	0.7510	0.8642
CSTD	ARI	0.6208	0.7656	0.6177	0.7155	0.7742

Table 2 — Region Covering and Boundary‑F

**The co-primary partition metrics, same runs.** Region Covering weights every ground-truth region by area whether or not a counterpart was proposed. Boundary‑F is computed on the dustbin-filled partition, so a method that emits a single region leaves no internal boundary and scores exactly zero.
Route	Metric	LISA	GSVA	GLaMM	Sa2VA	Detecture
RWTD	Cov.	0.5419	0.4230	0.4201	0.3987	0.7946
RWTD	B-F	0.0000	0.1179	0.0000	0.1025	0.4252
RWTD-COCO	Cov.	0.4817	0.4247	0.3558	0.2398	0.8536
RWTD-COCO	B-F	0.0000	0.1575	0.0000	0.1036	0.4226
TextureADE	Cov.	0.3155	0.2092	0.1674	0.1793	0.7479
TextureADE	B-F	0.0000	0.1795	0.0000	0.0516	0.5038
CSTD	Cov.	0.5094	0.7145	0.4844	0.5958	0.8835
CSTD	B-F	0.0000	0.3976	0.0000	0.2871	0.6499

Predicted region count on CSTD

Ground truth is exactly 2 on every image, marked on each bar.

Detecture 2.00

GSVA-7B 4.82

Sa2VA-4B 1.43

LISA-7B 0.99

GLaMM 0.87

05 regions

GSVA is the one baseline that always returns a partition; it simply returns the wrong number of parts. Its high CSTD ARI (0.7656, close to Detecture's 0.7742) is the reward for that, since over-fragmentation still agrees with the ground truth on most pixel pairs.

Two caveats stated plainly

These are 256 hand-verified images, not CSTD as published

CSTD's released region maps are the stitching masks fed into the generator, not annotations of what came out. Where the generator invented a third material or drifted from the mask, the ground truth silently stops being true. All 10,000 images were screened and ranked, 1,296 candidates were reviewed by eye, and 274 passed — an acceptance rate of 21%. The top 256 are what these numbers are computed on. Anyone rerunning CSTD off the shelf will get different numbers.

CSTD is synthetic

It enters only as an exactly-two-region control, not as evidence about real scenes. The real-world evidence is RWTD, RWTD-COCO and TextureADE.

Protocol

One protocol, unchanged across all five methods and all four routes: no ground-truth region count in the prompt, no inverse-mask completion, no truncation of proposals to a known K, and no dropping of images where a method returns nothing. Each baseline runs under its own pinned environment, the same ones that reproduce the RWTD, RWTD-COCO and TextureADE numbers.