Panoramas can serve as shared scene backgrounds for multi-shot video generation, but most existing panorama generators take text alone or a single input image. We present a single model that generates a full equirectangular panorama from a text caption and K in {0,...,4} perspective conditioning views taken by a four-camera rig at one viewpoint. We fine-tune a pretrained text-to-image rectified-flow transformer in two stages. In the first stage, each conditioning view is appended to the token sequence together with the ID of its camera slot, and the number of views varies during training, so one set of weights serves every input setting. In the second stage, we improve the region not covered by any conditioning view. Supervised training computes its loss on noisy latents of the ground truth and never scores a panorama the model has generated and decoded, so we propose Uncovered-Region GRPO (UR-GRPO), a reinforcement-learning recipe that rewards the candidates a vision-language model prefers in pairwise comparisons on crops around this region. Because the slot ID alone does not tell the model which way the rig points, we also offer an optional variant that writes each view's camera direction into its position encoding, which lets the user decide where each view appears in the output. With one conditioning view, our model outperforms the previous state of the art, 360Anything, trained on the same data and evaluated with the same code, on Laval Indoor and gives mixed results on an indoor subset of SUN360; the second stage improves the results for every number of conditioning views.
Figures & tables
Figure 1: Overview of training. Stage 1 fine-tunes one model for any number K of conditioning views, with a flow-matching loss on the noisy ground-truth panorama; each view token carries its camera slot ID (Ours) and, optionally, its direction (Ours + pose). Stage 2 trains a LoRA adapter on the frozen model: a VLM compares G candidates per input in pairs on crops around the uncovered region H , and the winners are reinforced.
Laval Indoor
SUN360 (indoor subset)
Method
crop-FID ↓
KID ↓
CLIP-FID ↓
FAED ↓
CS ↑
DS
crop-FID ↓
KID ↓
CLIP-FID ↓
FAED ↓
CS ↑
DS
Guidance embedding only
360Anything (re-impl.)
30.66
1.92
8.16
6.25
28.38
0.310
34.06
1.69
10.97
8.07
27.98
0.475
Ours, Stage 1
34.25
2.16
10.85
5.71
27.74
0.565
43.31
2.70
16.12
7.04
27.27
0.670
Ours, Stage 1+2
13.44
0.64
3.78
4.71
30.17
0.695
18.45
0.58
6.49
5.30
28.25
0.796
Ours + pose, Stage 1
45.41
3.06
13.12
6.12
27.53
0.554
42.91
2.65
15.03
5.78
27.52
0.668
Table 1: Comparison with 360Anything at K=1 (crops at azimuths 90∘ – 270∘ ). All rows are produced by the same evaluation code on the same test panoramas. 360Anything is our re-implementation, since the original code and weights are not released. Bold: best in each group; DS is not marked because the ground truth itself has a non-zero DS. KID is multiplied by 100 . The SUN360 subset differs from the split used in earlier papers ( Sec. 4 ).
Figure 2: Held-out Laval Indoor scenes at K=1 with CFG (same noise seed for all models). Left: the conditioning view of our models and the caption (shortened); the 360Anything re-implementation receives a 90∘ square view of the same direction. Each other column shows the panorama, two perspective crops at azimuths 135∘ and 225∘ , which face away from the input when the output is aligned, and a strip around the left/right seam. The scenes are taken from those where the output of Ours is aligned, so Ours keeps the input near the panorama centre here; the re-implementation places it at an arbitrary longitude, so some of its crops show its input region. Selection rule: Sec. 14 .
Laval Indoor
SUN360 (indoor subset)
Input
Model
crop-FID ↓
KID ↓
CLIP-FID ↓
FAED ↓
CS ↑
crop-FID ↓
KID ↓
CLIP-FID ↓
FAED ↓
CS ↑
K=0 + CFG
Stage 1
24.33
1.19
7.49
16.94
30.19
40.41
3.08
16.10
14.50
29.12
Stage 1+2
16.57
0.62
8.15
16.13
31.39
35.63
2.63
11.54
14.80
29.61
K=2
Stage 1
4.88
0.16
1.65
4.20
30.11
9.97
0.32
3.69
2.31
27.71
Stage 1+2
3.62
0.07
1.03
3.77
30.57
9.07
0.24
3.15
2.01
27.81
K=3
Stage 1
3.81
0.11
1.22
3.98
30.38
9.27
0.28
3.25
1.92
27.70
Table 2: Ours across view counts, before and after Stage 2. K=0 uses CFG; K=2 – 4 do not; K=1 is in Tab. 1 . Bold: the better of the two stages. With two or more views, the inputs already cover most crop directions, so these rows are closer to reconstruction than to generation. All metrics and both model variants are in the supplementary material, with both guidance settings at K=0 and K=1 ; at K=2 – 4 we did not run CFG.
Figure 3: Ours (Stage 1+2) on a held-out Laval Indoor scene with zero, two, three and four conditioning views (same caption and noise seed in all rows). K=0 uses CFG; K=2 – 4 do not and condition on the front and back views, plus the top view at K=3 and all four views at K=4 , which together cover 47% , 67% and 87% of the sphere. Red lines mark the boundary of the area covered by the conditioning views, drawn identically on the output and on the ground truth. The scene is taken from those whose outputs are aligned with the ground truth, so that the outline is meaningful ( Sec. 14 ). The black band at the bottom comes from the Laval images.
Figure 4: The uncovered region of Ours before (left) and after (right) Stage 2, on held-out Laval scenes at K=2 and 3 with the same conditioning views and noise seed, without CFG ( K=1 is compared in Fig. 2 , and at K=4 little of the panorama is uncovered). Each row is a perspective crop (horizontal field of view 85∘ ) of a side that the conditioning views leave open; the red line marks the boundary of the uncovered region. We picked these rows because the Stage-1 result is visibly smooth there ( Sec. 14 ); they illustrate the effect and do not show how often it occurs.
Judge
tie option
agreement
pairs/group
Qwen3-VL-32B
discouraged
0.622[.567,.674]
–
Qwen3-VL-32B
neutral
0.757[.701,.805]
11.1
InternVL3-38B
discouraged
0.653[.585,.716]
–
InternVL3-38B
neutral
0.789[.714,.848]
6.4
Both judges agree
neutral
0.871[.773,.931]
3.0
Gemma4-12B
discouraged
0.393[.329,.461] ∗
–
Table 3: Agreement with human rankings ( 40 groups) on pairs the human did not tie, counting only verdicts that survive swapping the two candidates. Pairs/group: how many of the 28 pairs remain. ∗ Below chance; excluded.
Figure 5: Rotated layout ( K=4 , no CFG, Stage 1+2, held-out Laval scenes). The camera layout is rotated by 135∘ about the vertical axis, so the front view should appear at azimuth 135∘ (red outline), where it also lies in the ground truth. Ours + pose places it there, which puts the back view near the centre. Ours ignores the rotation and places the front view at its reference orientation, the panorama centre (blue dashed outline). Scene selection: Sec. 14 .
Input
Model
crop-FID ↓
KID ↓
CLIP-FID ↓
FAED ↓
CS ↑
DS
raw-ERP FID ↓
K=0
Ours, Stage 1
86.03
5.69
17.84
29.50
26.18
0.783
87.11
Ours, Stage 1+2
33.27
1.90
10.00
16.20
30.25
0.996
41.58
Ours + pose, Stage 1
91.31
6.26
18.60
17.06
26.51
0.825
93.50
Ours + pose, Stage 1+2
41.38
2.55
11.31
15.69
29.59
0.867
52.93
K=0 + CFG
Ours, Stage 1
24.33
1.19
7.49
16.94
30.19
0.840
42.80
Ours, Stage 1+2
16.57
0.62
8.15
16.13
31.39
0.967
34.39
Table 4: Laval Indoor, all inputs, models and metrics. Crops at azimuths 90∘ – 270∘ . KID is multiplied by 100 . Bold: best within each input setting, except DS (ground-truth DS is 0.366 , so a lower value is not necessarily better). Raw-ERP FID is computed directly on the equirectangular image and is given only as a cross-check.
Input
Model
crop-FID ↓
KID ↓
CLIP-FID ↓
FAED ↓
CS ↑
DS
raw-ERP FID ↓
K=0
Ours, Stage 1
92.10
8.11
36.97
23.78
27.23
0.717
125.99
Ours, Stage 1+2
49.18
4.21
17.54
20.30
29.30
0.921
63.52
Ours + pose, Stage 1
87.50
8.00
32.60
28.28
27.52
0.804
126.09
Ours + pose, Stage 1+2
50.26
4.48
18.67
24.98
29.16
0.847
75.28
K=0 + CFG
Ours, Stage 1
40.41
3.08
16.10
14.50
29.12
0.762
82.38
Ours, Stage 1+2
35.63
2.63
11.54
14.80
29.61
0.876
63.79
Table 5: SUN360 indoor subset, all inputs, models and metrics (internal comparison only). Crops at azimuths 90∘ – 270∘ . KID is multiplied by 100 . Bold: best within each input setting, except DS (ground-truth DS is 0.839 , so a lower value is not necessarily better). Raw-ERP FID is computed directly on the equirectangular image and is given only as a cross-check.
Laval (all)
Laval (held-out)
SUN360
Method
wide
away
wide
away
wide
away
Guidance embedding only
360Anything (re-impl.)
30.66
31.82
30.79
31.93
34.06
36.11
Ours, Stage 1
34.25
37.98
34.78
38.51
43.31
48.50
Ours, Stage 1+2
13.44
14.67
14.24
15.55
18.45
20.29
Ours + pose, Stage 1
45.41
52.75
45.95
53.29
42.91
48.60
Table 6: Crop FID at K=1 under two crop ranges and on held-out scenes. wide : crops at azimuths 90∘ – 270∘ , as in [ 37 ] ; away : azimuths 110∘ – 250∘ , so that no crop overlaps our conditioning view when the output is aligned with it ( Sec. 4 ). Laval (held-out) excludes the 473 Laval scenes used during development. Bold: best in each group.
Method
FID ↓
KID ↓
CLIP-FID ↓
FAED ↓
CS ↑
OmniDreamer [ 1 ]
71.0
5.17
23.9
19.2
–
PanoDiffusion [ 36 ]
58.6
4.08
26.6
106.8
–
Diffusion360 [ 8 ]
33.1
2.07
16.9
23.7
26.38
CubeDiff [ 17 ]
9.5
0.32
3.2
18.4
27.02
360Anything [ 37 ]
8.0
0.22
4.6
9.8
29.21
Table 7: Published numbers on Laval Indoor at K=1 , copied from Table 1 of [ 37 ] (which in turn copies the first four rows from [ 17 ] ). They come from different evaluation code and preprocessing and are listed for reference only; they should not be compared directly with Tab. 1 .
Figure 6: Ours (Stage 1+2) at every number of conditioning views K , with the same caption and noise seed within each panorama. Neither panorama is used in training; the indoor one is from the SUN360 test set. Left: “a cozy, well-lit bedroom with … a private balcony … a scenic view of the sea”. Right: “a sunny, open urban park … a central square with a modern fountain” (captions shortened). At K=0 the model sees only the caption, so its panorama need not resemble the ground truth.
Figure 7: One panorama from our model, generated from a caption alone ( K=0 , with CFG; “a cozy, rustic interior with exposed wooden log walls and a white-beamed ceiling …”), used as the shared background of four shot images. The views of the character and the shot images ( 512×512 ) are generated with GPT-image [ 24 ] ; each shot image is generated from the views of the character, the panorama and a text prompt, and the four shots show the same room from different camera angles.