Datasets:

igfbench-neurips2026
/

IGF-Bench

model stringclasses 7 values	n_samples int64 487 3.6k	blender_abs_rel_mean float64 0.06 0.27	blender_abs_rel_std float64 0.21 0.39 ⌀	blender_rmse_mean float64 0.12 0.56	blender_rmse_std float64 0.25 0.35 ⌀	blender_si_rmse_mean float64 0.06 0.28	blender_si_rmse_std float64 0.12 0.2 ⌀	blender_delta125_mean float64 0.65 0.98	blender_delta125_std float64 0.1 0.25 ⌀	l_plane_residual_mean float64 0.06 0.18	l_plane_residual_std float64 0.08 0.15 ⌀	l_plane_inlier_ratio_mean float64 0.71 0.9	l_plane_inlier_ratio_std float64 0.06 0.1 ⌀	l_ortho_mean float64 0.15 0.36	l_ortho_std float64 0.16 0.2 ⌀	l_edge_f1_mean float64 0.03 0.61	l_edge_f1_std float64 0.11 0.32 ⌀	l_vp_deg_mean float64 0 18.7	l_vp_deg_std float64 0 10.8 ⌀	depthpro_abs_rel_mean float64 0.12 0.31 ⌀	depthpro_abs_rel_std float64 0.24 0.38 ⌀	depthpro_rmse_mean float64 0.33 0.69 ⌀	depthpro_rmse_std float64 0.4 0.45 ⌀	depthpro_si_rmse_mean float64 0.15 0.32 ⌀	depthpro_si_rmse_std float64 0.17 0.2 ⌀	depthpro_delta125_mean float64 0.58 0.86 ⌀	depthpro_delta125_std float64 0.23 0.25 ⌀	delta_l_plane_mean float64 0.01 0.12 ⌀	delta_l_plane_std float64 0.13 0.17 ⌀	delta_l_ortho_mean float64 0.02 0.19 ⌀	delta_l_ortho_std float64 0.18 0.22 ⌀	delta_l_edge_mean float64 -0.6 -0.36 ⌀	delta_l_edge_std float64 0.31 0.33 ⌀	delta_l_vp_mean float64 10.2 18.7 ⌀	delta_l_vp_std float64 7.4 10.8 ⌀	fid float64 56.3 103 ⌀	lpips_mean float64 0.59 0.65 ⌀	lpips_std float64 0.08 0.12 ⌀	level stringclasses 3 values	edge_class stringclasses 3 values
GT_baseline	3,600	0.055982	0.224619	0.122449	0.254923	0.060945	0.115274	0.976117	0.099322	0.05641	0.099598	0.900665	0.061788	0.166335	0.18125	0.606062	0.323899	0	0	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null	null
sd15	3,600	0.243166	0.318001	0.502138	0.351133	0.254967	0.190163	0.691091	0.247296	0.068756	0.084992	0.877144	0.070989	0.307996	0.196132	0.066791	0.108817	17.737395	10.477742	0.270724	0.338166	0.59881	0.452054	0.283308	0.194529	0.645307	0.25107	0.012254	0.126874	0.141106	0.213067	-0.539271	0.321161	17.737395	10.477742	56.647955	0.652943	0.091196	null	null
sd15	1,200	0.228911	null	0.516158	null	0.238282	null	0.699394	null	0.065546	null	0.892343	null	0.33545	null	0.027676	null	18.673136	null	0.256659	null	0.642923	null	0.274164	null	0.641113	null	0.009608	null	0.179855	null	-0.467045	null	18.673136	null	null	null	null	L0_empty	null
sd15	1,200	0.237724	null	0.497909	null	0.248106	null	0.699057	null	0.068058	null	0.874416	null	0.318624	null	0.079876	null	17.497314	null	0.268182	null	0.593515	null	0.277542	null	0.654743	null	0.009938	null	0.155512	null	-0.550177	null	17.497314	null	null	null	null	L1_basic	null
sd15	1,200	0.262928	null	0.492315	null	0.278591	null	0.674768	null	0.072713	null	0.864489	null	0.269026	null	0.092822	null	17.033391	null	0.287342	null	0.559954	null	0.298225	null	0.640069	null	0.01728	null	0.086563	null	-0.600591	null	17.033391	null	null	null	null	L2_full	null
sd15	1,702	0.26978	null	0.560646	null	0.273666	null	0.653482	null	0.067951	null	0.88406	null	0.361733	null	0.037221	null	18.184869	null	0.314209	null	0.693767	null	0.316803	null	0.584675	null	0.007634	null	0.154275	null	-0.482668	null	18.184869	null	null	null	null	null	sparse
sd15	1,411	0.218485	null	0.44492	null	0.236418	null	0.729988	null	0.069104	null	0.869686	null	0.262215	null	0.101384	null	17.814671	null	0.235244	null	0.509358	null	0.254979	null	0.702327	null	0.015525	null	0.132963	null	-0.600714	null	17.814671	null	null	null	null	null	moderate
sd15	487	0.221879	null	0.463922	null	0.243513	null	0.709525	null	0.070504	null	0.875046	null	0.257922	null	0.069912	null	16.00753	null	0.221724	null	0.526509	null	0.24846	null	0.691757	null	0.018555	null	0.120172	null	-0.55907	null	16.00753	null	null	null	null	null	dense
sdxl	3,600	0.168757	0.212842	0.375286	0.311226	0.194502	0.159851	0.796742	0.223563	0.095711	0.120509	0.840022	0.083673	0.227463	0.169537	0.193328	0.202417	15.569834	10.82304	0.197037	0.243344	0.475141	0.432822	0.223783	0.170602	0.748005	0.243387	0.039209	0.144136	0.060925	0.177997	-0.412735	0.315054	15.569834	10.82304	64.227121	0.594109	0.115818	null	null
sdxl	1,200	0.175216	null	0.398208	null	0.190265	null	0.785861	null	0.108201	null	0.835213	null	0.242561	null	0.085793	null	16.024754	null	0.207502	null	0.542915	null	0.229791	null	0.723595	null	0.051904	null	0.08765	null	-0.408929	null	16.024754	null	null	null	null	L0_empty	null
sdxl	1,200	0.164714	null	0.369885	null	0.191045	null	0.806364	null	0.08697	null	0.845839	null	0.226829	null	0.208176	null	16.351852	null	0.195457	null	0.466099	null	0.220657	null	0.758301	null	0.028773	null	0.063935	null	-0.421877	null	16.351852	null	null	null	null	L1_basic	null
sdxl	1,200	0.166335	null	0.357707	null	0.20222	null	0.798006	null	0.09187	null	0.839033	null	0.212606	null	0.286014	null	14.32424	null	0.188142	null	0.416353	null	0.220897	null	0.762141	null	0.036865	null	0.030353	null	-0.407398	null	14.32424	null	null	null	null	L2_full	null
sdxl	1,702	0.213286	null	0.458244	null	0.22993	null	0.728288	null	0.120745	null	0.816746	null	0.288977	null	0.099318	null	18.589079	null	0.25925	null	0.599357	null	0.275416	null	0.65718	null	0.060509	null	0.082378	null	-0.420571	null	18.589079	null	null	null	null	null	sparse
sdxl	1,411	0.133092	null	0.311928	null	0.167883	null	0.850497	null	0.072999	null	0.859848	null	0.178534	null	0.281633	null	13.922042	null	0.149167	null	0.374363	null	0.185955	null	0.817706	null	0.01942	null	0.049282	null	-0.420465	null	13.922042	null	null	null	null	null	moderate
sdxl	487	0.116834	null	0.269608	null	0.148098	null	0.87967	null	0.075679	null	0.862387	null	0.159787	null	0.266032	null	10.182752	null	0.118556	null	0.333518	null	0.153146	null	0.86311	null	0.023729	null	0.022037	null	-0.36295	null	10.182752	null	null	null	null	null	dense
sd35	3,600	0.234019	0.385261	0.452997	0.323791	0.234165	0.202957	0.742557	0.228826	0.120365	0.128933	0.815794	0.090007	0.308579	0.185823	0.122375	0.150747	14.655546	8.376049	0.247814	0.377786	0.530666	0.40772	0.25439	0.204003	0.711539	0.237942	0.063995	0.150013	0.142061	0.219345	-0.483688	0.329001	14.655546	8.376049	70.79463	0.627245	0.084067	null	null
sd35	1,200	0.192122	null	0.393878	null	0.187801	null	0.817761	null	0.125345	null	0.852716	null	0.272117	null	0.074423	null	15.960338	null	0.204837	null	0.514895	null	0.215917	null	0.773374	null	0.069439	null	0.116835	null	-0.420299	null	15.960338	null	null	null	null	L0_empty	null
sd35	1,200	0.23909	null	0.47592	null	0.240359	null	0.721888	null	0.099574	null	0.82998	null	0.355132	null	0.135928	null	14.553011	null	0.258958	null	0.548912	null	0.262595	null	0.690434	null	0.041462	null	0.192561	null	-0.494125	null	14.553011	null	null	null	null	L1_basic	null
sd35	1,200	0.27097	null	0.489316	null	0.274471	null	0.687838	null	0.136273	null	0.764107	null	0.298749	null	0.156773	null	13.440466	null	0.279681	null	0.528204	null	0.28469	null	0.670757	null	0.081216	null	0.116634	null	-0.536639	null	13.440466	null	null	null	null	L2_full	null
sd35	1,702	0.261889	null	0.495293	null	0.250224	null	0.713483	null	0.132979	null	0.820938	null	0.346585	null	0.073836	null	16.075758	null	0.289586	null	0.609254	null	0.281685	null	0.66816	null	0.072988	null	0.13992	null	-0.446053	null	16.075758	null	null	null	null	null	sparse
sd35	1,411	0.217156	null	0.425297	null	0.224152	null	0.763465	null	0.111983	null	0.804667	null	0.288398	null	0.171365	null	14.339712	null	0.220723	null	0.468554	null	0.236965	null	0.742054	null	0.058403	null	0.159146	null	-0.530733	null	14.339712	null	null	null	null	null	moderate
sd35	487	0.185707	null	0.385783	null	0.207185	null	0.78335	null	0.101415	null	0.830364	null	0.238012	null	0.150071	null	10.790897	null	0.180486	null	0.436294	null	0.209594	null	0.77455	null	0.049466	null	0.100263	null	-0.478911	null	10.790897	null	null	null	null	null	dense
flux1	3,600	0.198558	0.283583	0.415716	0.329267	0.21544	0.183831	0.760005	0.240304	0.115332	0.154429	0.819044	0.10386	0.198117	0.159895	0.155603	0.169787	14.845962	8.369252	0.230573	0.31983	0.519819	0.451246	0.246352	0.192948	0.713466	0.251627	0.058939	0.155188	0.03186	0.190568	-0.45046	0.326947	14.845962	8.369252	56.283754	0.588274	0.115737	null	null
flux1	1,200	0.171801	null	0.385014	null	0.184979	null	0.798648	null	0.113946	null	0.841935	null	0.191893	null	0.094974	null	15.845429	null	0.213958	null	0.539929	null	0.230608	null	0.732051	null	0.05803	null	0.037229	null	-0.399747	null	15.845429	null	null	null	null	L0_empty	null
flux1	1,200	0.197661	null	0.41708	null	0.214565	null	0.75694	null	0.110777	null	0.83101	null	0.200667	null	0.172836	null	14.503252	null	0.234167	null	0.521435	null	0.246331	null	0.709728	null	0.052431	null	0.038215	null	-0.457217	null	14.503252	null	null	null	null	L1_basic	null
flux1	1,200	0.226305	null	0.445153	null	0.246882	null	0.724307	null	0.121324	null	0.783804	null	0.201921	null	0.198998	null	14.180733	null	0.243608	null	0.498074	null	0.262131	null	0.698605	null	0.066429	null	0.019878	null	-0.494415	null	14.180733	null	null	null	null	L2_full	null
flux1	1,702	0.235109	null	0.478868	null	0.242669	null	0.711755	null	0.144155	null	0.798346	null	0.239659	null	0.100406	null	15.733069	null	0.28683	null	0.625346	null	0.289682	null	0.641685	null	0.084164	null	0.033037	null	-0.419483	null	15.733069	null	null	null	null	null	sparse
flux1	1,411	0.170034	null	0.366354	null	0.19556	null	0.799548	null	0.092591	null	0.831654	null	0.164542	null	0.208623	null	14.544649	null	0.188086	null	0.431244	null	0.214396	null	0.771455	null	0.039011	null	0.035289	null	-0.493475	null	14.544649	null	null	null	null	null	moderate
flux1	487	0.153759	null	0.338546	null	0.178104	null	0.813664	null	0.082405	null	0.85345	null	0.154022	null	0.194891	null	12.733402	null	0.157292	null	0.40808	null	0.187683	null	0.796024	null	0.030455	null	0.018028	null	-0.434091	null	12.733402	null	null	null	null	null	dense
hunyuan	3,600	0.222927	0.332443	0.439987	0.303885	0.232516	0.192728	0.739549	0.232783	0.136343	0.143147	0.78671	0.094559	0.254856	0.162453	0.178078	0.184784	15.84059	9.233164	0.248327	0.335997	0.536169	0.424711	0.260027	0.194469	0.691624	0.241985	0.079776	0.166784	0.088084	0.203995	-0.427984	0.322243	15.84059	9.233164	82.174385	0.611203	0.093487	null	null
hunyuan	1,200	0.202722	null	0.428916	null	0.205715	null	0.769546	null	0.144184	null	0.821997	null	0.252817	null	0.098038	null	18.344163	null	0.231758	null	0.568437	null	0.245062	null	0.702554	null	0.087683	null	0.097366	null	-0.396684	null	18.344163	null	null	null	null	L0_empty	null
hunyuan	1,200	0.232448	null	0.467101	null	0.244372	null	0.715703	null	0.128143	null	0.78722	null	0.246804	null	0.186009	null	15.129064	null	0.26632	null	0.562216	null	0.274378	null	0.666375	null	0.070044	null	0.084049	null	-0.444044	null	15.129064	null	null	null	null	L1_basic	null
hunyuan	1,200	0.233648	null	0.423891	null	0.247511	null	0.733377	null	0.136665	null	0.75046	null	0.265114	null	0.250188	null	14.02657	null	0.246915	null	0.477827	null	0.260653	null	0.705935	null	0.081578	null	0.082623	null	-0.443225	null	14.02657	null	null	null	null	L2_full	null
hunyuan	1,702	0.262471	null	0.50296	null	0.258226	null	0.692368	null	0.176603	null	0.769029	null	0.285316	null	0.094779	null	18.23439	null	0.306362	null	0.642665	null	0.301147	null	0.620408	null	0.116273	null	0.078102	null	-0.42511	null	18.23439	null	null	null	null	null	sparse
hunyuan	1,411	0.196677	null	0.393736	null	0.215933	null	0.772484	null	0.104581	null	0.795646	null	0.232342	null	0.255528	null	14.702929	null	0.207787	null	0.44983	null	0.231841	null	0.744217	null	0.051001	null	0.10309	null	-0.44657	null	14.702929	null	null	null	null	null	moderate
hunyuan	487	0.161108	null	0.354428	null	0.19092	null	0.808627	null	0.090346	null	0.821415	null	0.216426	null	0.244801	null	11.080424	null	0.163195	null	0.414568	null	0.198155	null	0.787845	null	0.038397	null	0.078676	null	-0.384181	null	11.080424	null	null	null	null	null	dense
kolors	3,600	0.205786	0.319111	0.417862	0.307459	0.216785	0.184859	0.771369	0.220382	0.135506	0.131294	0.765123	0.100166	0.261027	0.170395	0.171631	0.184635	14.56809	7.398215	0.223397	0.337879	0.49297	0.395766	0.236371	0.189132	0.736931	0.23329	0.079186	0.160608	0.094029	0.20216	-0.434431	0.313747	14.56809	7.398215	103.068567	0.621643	0.106081	null	null
kolors	1,200	0.209055	null	0.450817	null	0.215916	null	0.75735	null	0.160294	null	0.732387	null	0.28818	null	0.053145	null	16.65668	null	0.228262	null	0.555455	null	0.241974	null	0.711112	null	0.104411	null	0.132816	null	-0.441577	null	16.65668	null	null	null	null	L0_empty	null
kolors	1,200	0.199113	null	0.406529	null	0.208982	null	0.781356	null	0.130159	null	0.77559	null	0.261272	null	0.194347	null	14.48756	null	0.218201	null	0.483075	null	0.229645	null	0.746931	null	0.072086	null	0.097903	null	-0.435706	null	14.48756	null	null	null	null	L1_basic	null
kolors	1,200	0.2092	null	0.396169	null	0.225485	null	0.775414	null	0.115789	null	0.787728	null	0.232877	null	0.267402	null	12.539074	null	0.223725	null	0.440327	null	0.23749	null	0.752772	null	0.060764	null	0.050126	null	-0.426011	null	12.539074	null	null	null	null	L2_full	null
kolors	1,702	0.258551	null	0.509765	null	0.258253	null	0.69163	null	0.179104	null	0.714846	null	0.328596	null	0.065863	null	16.526134	null	0.291068	null	0.613456	null	0.288676	null	0.643831	null	0.119315	null	0.121056	null	-0.454025	null	16.526134	null	null	null	null	null	sparse
kolors	1,411	0.16539	null	0.34311	null	0.183457	null	0.836927	null	0.099432	null	0.803792	null	0.206569	null	0.269761	null	13.715332	null	0.172612	null	0.394363	null	0.196901	null	0.812022	null	0.045852	null	0.077317	null	-0.432337	null	13.715332	null	null	null	null	null	moderate
kolors	487	0.138848	null	0.314011	null	0.168761	null	0.859449	null	0.090548	null	0.825445	null	0.188591	null	0.256961	null	10.449008	null	0.134317	null	0.358076	null	0.168148	null	0.844356	null	0.038598	null	0.050841	null	-0.372022	null	10.449008	null	null	null	null	null	dense

IGF-Bench: Indoor Geometric Fidelity Benchmark

Anonymous mirror for NeurIPS 2026 Evaluations and Datasets Track double-blind review. The de-anonymised author/maintainer information will replace this header at camera-ready.

IGF-Bench is the first benchmark for evaluating structural-level geometric fidelity of conditionally generated indoor scene images, going beyond perceptual metrics like FID and LPIPS. It pairs 3,600 calibrated synthetic ground-truth views with 21,600 generated images from six state-of-the-art ControlNet models, plus 25,200 monocular-depth estimates, all evaluated with four complementary geometric metrics: planarity (L_plane), orthogonality (L_ortho), edge alignment (L_edge), and vanishing-point consistency (L_vp).

Quick stats	Value
Calibrated GT views	3,600 (300 rooms × 3 complexity levels × 4 viewpoints)
Paired generated images	21,600 (6 ControlNet models, all conditioned on identical Canny maps)
Paired depth estimates	25,200 (DepthPro on all GT + generated; DAv2 / ZoeDepth subsets)
Camera FOV	90°
Render resolution	1024×1024
Total size	≈ 219 GB
License	CC BY-NC-SA 4.0 (data) + Apache 2.0 (code)
Code repo	https://anonymous.4open.science/r/IGF-Bench-Code
Paper	NeurIPS 2026 E&D Track (under review)

What's in this Repository

This HuggingFace dataset repository ships everything needed to evaluate or extend IGF-Bench:

✅ Ground-truth 3D-FRONT renders with calibrated cameras (RGB + EXR depth + 7-class semantic mask + Canny edge map) at 3 complexity levels (L0_empty / L1_basic / L2_full).
✅ Paired generated images from six ControlNet pipelines (SD 1.5, SDXL 1.0, SD 3.5 Large, Flux.1 Dev, Hunyuan-DiT, Kolors) under a uniform protocol (Canny conditioning, seed=42, no negative prompts, cn_scale=1.0).
✅ Per-view depth estimates from DepthPro on all 25,200 images, plus DAv2 and ZoeDepth on the 200-view ablation subset (App C.1).
✅ Pre-computed evaluation JSONs that already populate every table and figure in the paper (evaluation/).
✅ A trained LoRA adapter (experiments/finetune/dav2_lora_adapter/) that demonstrates IGF-Bench is usable as supervision for fine-tuning a pretrained MDE (App C.8).
✅ A 200-view MDE consistency subset with DAv2 + ZoeDepth depth pre-computed for paper App C.1 (tab:mde_ablation).
✅ A formal Datasheet for Datasets (DATASHEET.md) and Croissant 1.0 metadata (croissant.json).

Headline Findings

Each finding here is reproducible from the JSONs shipped under evaluation/ — no recompute required.

Perceptual quality ≠ geometric fidelity. Models with comparable FID (e.g., SDXL FID = 64.2 vs. Flux.1 FID = 56.3) can still differ by ≈2× in ΔL_ortho (Flux.1 0.032 vs. SDXL 0.061). Perceptual metrics alone miss structural failures.
Conditioning architecture matters. Flux.1's channel-concatenation conditioning achieves the lowest ΔL_ortho (0.032), substantially better than the residual-injection ControlNet variants used by the other five models.
Edge alignment is the universal bottleneck. All six models show 68%–89% relative L_edge degradation versus GT — even the best model fails on fine geometric structure.
Cleanliness of synthetic GT. DepthPro AbsRel on IGF-Bench GT is 0.056, vs. 0.084 on NYU-v2 (real Kinect) — the lower MDE-noise floor makes IGF-Bench's relative-degradation (ΔL_*) design more sensitive to generation-induced artefacts.
L_ortho works as a label-free quality signal. Picking the per-view best of 5 random Flux.1 seeds by minimum L_ortho cuts L_ortho by 35.1% relative to the best fixed seed (App C.7.1).
The synthetic supervision transfers (in-domain). Fine-tuning DepthAnything V2-Small with a 1.75%-parameter LoRA adapter improves in-domain AbsRel from 1.044 to 0.081 (−92%) on the 3D-FRONT held-out test set (App C.8).

Quick Start

Option A — Verify the paper without downloading bulk data

The full evaluation already shipped in evaluation/ is < 50 MB. You can verify every paper number by reading the JSONs directly:

import json
from huggingface_hub import hf_hub_download

# Example: verify the GT L_plane = 0.056 number
fp = hf_hub_download(
    repo_id="igfbench-neurips2026/IGF-Bench",
    filename="evaluation/igf_summary.json",
    repo_type="dataset",
)
data = json.load(open(fp))
gt = next(d for d in data if d["model"] == "GT_baseline")
print(round(gt["l_plane_residual_mean"], 3))   # → 0.056

Option B — Download everything and re-run

pip install huggingface_hub
huggingface-cli download igfbench-neurips2026/IGF-Bench --repo-type dataset \
    --local-dir ./igf-bench-data
export IGF_BENCH_ROOT=$(pwd)/igf-bench-data

Then clone the code at https://anonymous.4open.science/r/IGF-Bench-Code and run:

python scripts/evaluate_igf.py \
    --renders_root  $IGF_BENCH_ROOT/renders_textured \
    --generated_root $IGF_BENCH_ROOT/generated \
    --depth_root    $IGF_BENCH_ROOT/depth_results \
    --output_summary igf_summary.json

This reproduces the full Table 3 from the paper. Runtime ≈ 6 hours on a single RTX 4090; 99% of the time is spent in I/O reading per-view depth NPYs.

Option C — Download only what you need for one row of the main table

For example, to recompute only the Flux.1 row of Table 3 (≈ 70 GB):

huggingface-cli download igfbench-neurips2026/IGF-Bench --repo-type dataset \
    --local-dir ./igf-bench-data \
    --include "renders_textured/*" \
              "generated/flux1_canny/*" \
              "depth_results/gt/depthpro/*" \
              "depth_results/gen/flux1/depthpro/*"

Directory Structure

igfbench-neurips2026/IGF-Bench/
├── README.md                          ← this file
├── DATASHEET.md                       ← formal Datasheets-for-Datasets record
├── croissant.json                     ← MLCommons Croissant 1.0 metadata
├── dataset_card.md                    ← short HF dataset card (4.4 KB summary)
├── LICENSE                            ← CC BY-NC-SA 4.0
├── selected_rooms.json                ← canonical 300-room list
├── room_statistics.json               ← summary statistics
│
├── renders_textured/                  ≈ 36 GB — Ground-truth rendered views
│   └── {scene_id}_{room_type}-{room_id}/
│       └── {L0_empty,L1_basic,L2_full}/
│           └── view_{0,1,2,3}/
│               ├── rgb_textured.png        RGB render
│               ├── depth.exr               metric depth, OpenEXR float32
│               ├── depth.png               8-bit depth visualisation (QA only)
│               ├── depth_gt.npy            NumPy float32 cache of depth.exr (LoRA training I/O)
│               ├── canny.png               Canny edges (thresholds 100/200)
│               ├── semantic_id.png         7-class IDs (0..6)
│               ├── semantic_mask.png       colour-mapped semantic (QA only)
│               ├── wireframe_3d.png        3D wireframe overlay (QA only)
│               └── camera.json             intrinsics + extrinsics
│
├── generated/                         ≈ 36 GB — Generated images per model
│   │
│   │   --- Main protocol: 6 models × 3,600 views = 21,600 PNG (Table 3) ---
│   ├── sd15_canny/                    SD 1.5 native 512² (300 rooms × 3 levels × 4 views)
│   ├── sdxl_canny/                    SDXL 1.0
│   ├── sd35_canny/                    SD 3.5 Large
│   ├── flux1_canny/                   Flux.1 Dev
│   ├── hunyuan_canny/                 Hunyuan-DiT
│   ├── kolors_canny/                  Kolors
│   │
│   │   --- Ablation: negative prompt, App C.3 (3,552 paired views) ---
│   ├── sd15_canny_with_neg/sd15_canny/   SD 1.5 with the legacy negative prompt
│   ├── sdxl_canny_with_neg/sdxl_canny/   SDXL with the legacy negative prompt
│   │
│   │   --- Ablation: resolution, App C.4 (3,552 paired views) ---
│   ├── sd15_upsampled/sd15_canny/     SD 1.5 outputs bicubic-upsampled to 1024²
│   │
│   │   --- Ablation: seed sensitivity (App C.6) + quality gating (App C.7), 200 views each ---
│   └── ablation/
│       ├── flux1_seed123/             App C.6 / C.7 — Flux.1 with seed=123
│       ├── flux1_seed456/             App C.6 / C.7 — Flux.1 with seed=456
│       ├── flux1_seed789/             App C.7 only — Flux.1 with seed=789
│       ├── flux1_seed1024/            App C.7 only — Flux.1 with seed=1024
│       ├── sdxl_w050/                 App C.7 — SDXL with cn_scale=0.50
│       ├── sdxl_w075/                 App C.7 — SDXL with cn_scale=0.75
│       ├── sdxl_w125/                 App C.7 — SDXL with cn_scale=1.25
│       └── sdxl_w150/                 App C.7 — SDXL with cn_scale=1.50
│   (each leaf directory contains the same {room}/{level}/view_{0..3}.png structure)
│
├── depth_results/                     ≈ 150 GB — Per-MDE depth estimates (NPY)
│   ├── gt/                            Ground-truth-render depth (3 MDE backbones)
│   │   ├── depthpro/                  3,600 NPY (full set)
│   │   ├── dav2/                      ablation 200-view subset (App C.1)
│   │   └── zoedepth/                  ablation 200-view subset (App C.1)
│   │
│   └── gen/                           Generated-image depth
│       │
│       │   --- Main protocol: 6 models × 3,600 = 21,600 NPY ---
│       ├── sd15/depthpro/             3,600 NPY
│       ├── sdxl/depthpro/             3,600 NPY  (also dav2/, zoedepth/ for App C.1 200-view subset)
│       ├── sd35/depthpro/             3,600 NPY
│       ├── flux1/depthpro/            3,600 NPY  (also dav2/, zoedepth/ for App C.1 200-view subset)
│       ├── hunyuan/depthpro/          3,600 NPY
│       ├── kolors/depthpro/           3,600 NPY
│       │
│       │   --- Ablation depth maps ---
│       ├── sd15_neg/depthpro/         3,552 NPY  (App C.3 negative prompt)
│       ├── sdxl_neg/depthpro/         3,552 NPY  (App C.3 negative prompt)
│       ├── sd15_upsampled/depthpro/   3,552 NPY  (App C.4 resolution)
│       ├── flux1_seed123/depthpro/    200 NPY    (App C.6 / C.7 seed=123)
│       ├── flux1_seed456/depthpro/    200 NPY    (App C.6 / C.7 seed=456)
│       ├── flux1_seed789/depthpro/    200 NPY    (App C.7 seed=789)
│       ├── flux1_seed1024/depthpro/   200 NPY    (App C.7 seed=1024)
│       ├── sdxl_w050/depthpro/        200 NPY    (App C.7 cn_scale=0.50)
│       ├── sdxl_w075/depthpro/        200 NPY    (App C.7 cn_scale=0.75)
│       ├── sdxl_w125/depthpro/        200 NPY    (App C.7 cn_scale=1.25)
│       └── sdxl_w150/depthpro/        200 NPY    (App C.7 cn_scale=1.50)
│       (each: {room}/{level}/{view}.npy)
│
├── evaluation/                        ≈ 41 MB — Pre-computed metric outputs
│   ├── igf_summary.json               per-model aggregated (Table 3 source)
│   ├── igf_results.json               per-view detailed (≈ 25 MB)
│   ├── error_decomposition.json       App C.2
│   ├── mde_ablation_*.json            App C.1
│   ├── neg_prompt_ablation.json    App C.3
│   ├── n1_resolution_ablation.json    App C.4
│   ├── n3_lvp_improved.json           App C.5
│   ├── seed_ablation.json             App C.6
│   ├── wilcoxon.json               per-pair Wilcoxon + Bonferroni (paper Table 3)
│   ├── anova.json                  two-way Type II ANOVA (paper §4.3)
│   └── fid_lpips.json                 perceptual baselines
│
└── experiments/finetune/
    └── dav2_lora_adapter/             ★ 1.8 MB — Trained LoRA (App C.8)
        ├── adapter_config.json        peft config: r=16, α=32, Q/K/V, dropout=0.05
        ├── adapter_model.safetensors  442,368 trainable params (1.75%)
        └── README.md                  load instructions

Reproducing Each Paper Experiment

Paper Section	Required HF subsets	Expected runtime	Code entrypoint
§4 / Table 3 main IGF metrics	`renders_textured/`, `generated/{all}/`, `depth_results/{gt,gen}/depthpro/`	~6 h	`scripts/evaluate_igf.py`
App C.1 MDE robustness	+ `depth_results/{gt,gen}/{dav2,zoedepth}/` (200 views)	~30 min	`scripts/evaluate_mde_ablation.py`
App C.2 Error decomposition	already in `evaluation/error_decomposition.json`	<1 min	`scripts/error_decomposition.py`
App C.3 Negative-prompt	needs the legacy generations (archived under `generated/sd15_canny_with_neg/sd15_canny/` and `generated/ablation/sdxl_with_neg/sdxl_canny/` on HF)	~3 h	`scripts/evaluate_neg_prompt.py`
App C.4 Resolution confound	`generated/sd15_canny/` + local bicubic upsample to 1024²	~1 h	`scripts/run_n1_resolution_ablation.py`
App C.5 VP 2D	`renders_textured/` + `depth_results/{gt,gen}/depthpro/`	~30 min	`scripts/run_n3_lvp_improved.py`
App C.6 Seed sensitivity	`generated/ablation/flux1_seed{123,456}/` (already on HF)	~30 min	`scripts/evaluate_seed_ablation.py`
App C.7 Quality gating	`generated/ablation/flux1_seed{123,456,789,1024}/` (seed=42 reuses `generated/flux1_canny/`) + `generated/ablation/sdxl_w{050,075,125,150}/` (cn=1.0 reuses `generated/sdxl_canny/`)	~30 min	`experiments/analyze_solutions.py`
App C.8 LoRA fine-tune	`renders_textured/` + base DAv2-S model	~1 h training + ~5 min eval	`experiments/finetune/{create_split,train_single_gpu,eval_nyu}.py` (training wrapper); pre-trained adapter at `experiments/finetune/dav2_lora_adapter/`
App F.1 Cross-dataset	NYU-v2 + iBims-1 from official sources (NOT redistributed)	~30 min	`scripts/evaluate_cross_dataset.py`
App F.2 Complexity-MDE	already in main 219 GB	~5 min	included in `evaluate_igf.py`
App F.3 3D reconstruction	`renders_textured/` (30 rooms L2_full) + 3D-FRONT meshes	~1 h	`experiments/run_appendix_F_experiments.py --section F3_reconstruction`
App F.4 MDE domain gap	+ NYU-v2 from official source	~30 min	`experiments/run_appendix_F_experiments.py --section F4_mde_domain`

Pre-Computed Result JSONs

If you only need to verify the paper's numbers (no recompute), the JSONs below are sufficient.

Paper item	JSON path on HF	Key
Table 3 (main)	`evaluation/igf_summary.json`	per-model aggregated means
Tab error_decomp	`evaluation/error_decomposition.json`	per-model AbsRel decomposition
Tab mde_ablation	`evaluation/mde_ablation_summary.json`	SDXL/Flux × 3 MDE backbones
Tab neg_prompt	`evaluation/neg_prompt_ablation.json`	per-model w/ vs w/o
Tab n1_resolution	`evaluation/n1_resolution_ablation.json`	SD 1.5 native vs upsampled
Tab n3_lvp_2d	`evaluation/n3_lvp_improved.json`	per-model 2D L_vp
Tab seed_ablation	`evaluation/seed_ablation.json`	Flux.1 across 3 seeds
Wilcoxon p-values	`evaluation/wilcoxon.json`	per-pair Bonferroni-corrected
FID + LPIPS	`evaluation/fid_lpips.json`	per-model perceptual baselines

App F downstream JSONs are released with the paper supplement (https://anonymous.4open.science/r/IGF-Bench-Code) under experiments/results/.

Schema Reference

`selected_rooms.json`

{
  "rooms": [
    {
      "scene_id": "00110bde-f580-40be-b8bb-88715b338a2a",
      "room_id":  "Bedroom-43072",
      "room_type": "Bedroom",
      "space_type": "Bedroom",
      "area_m2": 12.3
    },
    ...
  ]
}

The on-disk directory name for each room is {scene_id}_{room_id}.

`evaluation/igf_summary.json`

A list of dicts, one per model and one for GT_baseline:

[
  {
    "model": "GT_baseline",
    "n_samples": 3600,
    "l_plane_residual_mean": 0.056,
    "l_ortho_mean": 0.166,
    "l_edge_f1_mean": 0.606,
    "l_vp_deg_mean": 1.79,
    "blender_abs_rel_mean": 0.0,
    "depthpro_abs_rel_mean": 0.056,
    ...
  },
  {
    "model": "sdxl",
    "n_samples": 3600,
    "delta_l_plane_mean": 0.052,
    "delta_l_ortho_mean": 0.061,
    "delta_l_edge_mean": -0.413,
    ...
  },
  ...
]

`evaluation/igf_results.json`

A dict keyed by model (gt_baseline, sd15, …, kolors), each holding a list of 3,600 per-view records:

{
  "gt_baseline": [
    {
      "room": "00110bde-..._Bedroom-43072",
      "level": "L0_empty",
      "view": "view_0",
      "l_plane_residual": 0.045,
      "l_plane_inlier_ratio": 0.84,
      "l_ortho": 0.130,
      "n_normals": 5,
      "l_edge_f1": 0.612,
      "l_edge_precision": 0.59,
      "l_edge_recall": 0.64,
      "l_vp_deg": 0.41,
      "depthpro_abs_rel": 0.052,
      "depthpro_rmse": 0.118,
      "depthpro_si_rmse": 0.041,
      "depthpro_delta125": 0.984,
      ...
    },
    ...
  ],
  "sdxl": [...],
  ...
}

`experiments/finetune/dav2_lora_adapter/adapter_config.json`

{
  "base_model_name_or_path": "depth-anything/Depth-Anything-V2-Metric-Indoor-Small-hf",
  "peft_type": "LORA",
  "r": 16,
  "lora_alpha": 32,
  "lora_dropout": 0.05,
  "target_modules": ["query", "key", "value"],
  "bias": "none"
}

Loadable via peft.PeftModel.from_pretrained(base, adapter_dir).

License Chain

IGF-Bench dataset
├── License: CC BY-NC-SA 4.0
│   • Inherits NC clause from upstream 3D-FRONT (Alibaba Tianchi NC license).
│   • Attribution required, share-alike, non-commercial use only.
│
├── Code (separate): Apache 2.0
│   • Located in the supplementary code repo, not on this dataset HF repo.
│
├── Per-asset upstream model licenses (binding for downstream redistribution):
│   ├── SD 1.5             → CreativeML Open RAIL-M
│   ├── SDXL 1.0           → CreativeML Open RAIL++-M
│   ├── SD 3.5 Large       → Stability AI Community License (NC if revenue ≤ $1M/yr)
│   ├── Flux.1 Dev         → FLUX.1 [dev] Non-Commercial License
│   ├── Hunyuan-DiT        → Tencent Hunyuan Community License
│   ├── Kolors             → Apache 2.0 + Kwai commercial-registration requirement
│   ├── DepthPro           → Apple Sample Code License (apple-amlr; output redistribution
│   │                         in a "gray area" — see DATASHEET §6 for full disclosure)
│   ├── Depth-Anything V2-Small (metric-indoor) → Apache 2.0
│   └── ZoeDepth           → MIT
│
└── Eval-only datasets (NOT redistributed by us; obtain from the official source):
    ├── NYU-v2  → https://cs.nyu.edu/~silberman/datasets/nyu_depth_v2.html
    └── iBims-1 → https://www.cvg.cit.tum.de/data/datasets/ibims1

For the full per-asset table see DATASHEET.md §6 (Distribution).

The dav2_lora_adapter/ weights inherit CC BY-NC-SA 4.0 from the 3D-FRONT supervision data, even though the base DAv2-S model is Apache 2.0.

Maintenance

The authors commit to maintaining IGF-Bench for at least 5 years post-publication, including:

Hosting on HuggingFace with versioned releases (v1.0.0, v1.1.0, …);
Bug fixes via the GitHub Issues tracker (URL pending de-anonymisation);
Adding new generation models as they become available;
Periodic re-evaluation when major MDE backbones are released.

Older versions remain accessible on HuggingFace forever (see the refs/convert/parquet/<commit> history).

Citation

@inproceedings{igfbench2026,
  title     = {IGF-Bench: Evaluating Geometric Fidelity of Conditional Image
               Generation Beyond Perceptual Metrics},
  author    = {Anonymous},
  booktitle = {Advances in Neural Information Processing Systems
               (Datasets and Benchmarks Track)},
  year      = {2026}
}

When citing the upstream 3D-FRONT scenes, please also cite Fu et al., 3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics (ICCV 2021).

Downloads last month: 420

Number of rows:

Total file size:

234 GB