Experiment · complete

Local OSS round: counting and spatial relations

Which common local OSS checkpoint most accurately renders the exact object counts, ordering, and behind relation?

Runs
7
Variable
local OSS checkpoint
Prompt
oss-round-spatial-prompt v1.0

Protocol

One unseeded local generation per checkpoint using the identical frozen hosted-round prompt, common 1344x768 output, and each checkpoint's existing calibrated defaults.

Frozen prompt template

CONTROL PROMPT AB-R1-SPATIAL-001. A wide 16:9 museum-catalog photograph of a matte charcoal tabletop containing exactly seven objects and nothing else: three red wooden cubes in one straight row on the left; two transparent glass spheres in the center; one small aged-brass pyramid directly behind the two spheres; and one folded cream paper crane on the far right. Camera elevated thirty degrees, all seven objects fully visible and separated, physically realistic materials, neutral softbox lighting, dark seamless background. No labels, no text, no watermark, no duplicate objects, no containers.

Review rubric

  • Prompt Adherence
  • Object Counting
  • Spatial Relations
  • Material Realism

Notes

Seven common checkpoints through SwarmUI on the authorized local 3090 Ti. Requests were sequential to avoid model-load contention. Random seeds were requested but the realized seed was not exposed, so run records use null.

Evidence

Recorded runs

Generated specimen oss-r2-spatial-chroma1-base oss-r2-spatial-chroma1-base
successChroma1 Base

Oss R2 Spatial Chroma1 Base

Vision adherence score 6.0/10. Nine objects: 3 red cubes, 2 glass spheres, 2 pyramid forms, and 2 crane-like paper forms. It gets the cube and sphere counts but duplicates both the pyramid and crane, and the pyramid is not directly behind the spheres.

Size
1344×768
Seed
not exposed
Method
local inference
View experiment →
Generated specimen oss-r2-spatial-dreamshaper-xl-v21 oss-r2-spatial-dreamshaper-xl-v21
successDreamShaper XL v2.1

Oss R2 Spatial Dreamshaper Xl V21

Vision adherence score 2.0/10. Eight objects: 3 red spheres, 2 pyramids, 1 white sphere, 1 wireframe polyhedron, and 1 beige sphere. There are no cubes, no paper crane, and no required left/center/right relation.

Size
1344×768
Seed
not exposed
Method
local inference
View experiment →
Generated specimen oss-r2-spatial-flux1-dev oss-r2-spatial-flux1-dev
successFLUX.1 Dev FP8

Oss R2 Spatial Flux1 Dev

Vision adherence score 4.0/10. Ten objects: 4 red cubes, 4 glass spheres, 1 brass pyramid, and 1 paper crane. The object classes are recognizable but the cube/sphere counts are both high and the pyramid sits between rather than directly behind the spheres.

Size
1344×768
Seed
not exposed
Method
local inference
View experiment →
Generated specimen oss-r2-spatial-flux1-schnell oss-r2-spatial-flux1-schnell
successFLUX.1 Schnell FP8

Oss R2 Spatial Flux1 Schnell

Vision adherence score 2.5/10. Four principal objects: 1 red cube, 1 glass sphere, 1 brass pyramid, and 1 paper crane. The classes are present but three cubes and one sphere are missing and the requested spatial arrangement is not preserved.

Size
1344×768
Seed
not exposed
Method
local inference
View experiment →
Generated specimen oss-r2-spatial-juggernaut-xl-v9 oss-r2-spatial-juggernaut-xl-v9
successJuggernaut XL v9

Oss R2 Spatial Juggernaut Xl V9

Vision adherence score 0.5/10. Four red spheres on a framed tray. No cubes, transparent glass spheres, brass pyramid, paper crane, or required arrangement are present.

Size
1344×768
Seed
not exposed
Method
local inference
View experiment →
Generated specimen oss-r2-spatial-realvis-xl-v40 oss-r2-spatial-realvis-xl-v40
successRealVisXL v4.0

Oss R2 Spatial Realvis Xl V40

Vision adherence score 0.5/10. A crowded still life of well over seven objects including red/white/brass spheres, black cubes, pyramids, and small polyhedra. The required counts, materials, crane, and left/center/right arrangement all fail.

Size
1344×768
Seed
not exposed
Method
local inference
View experiment →
Generated specimen oss-r2-spatial-sdxl-base-1 oss-r2-spatial-sdxl-base-1
successStable Diffusion XL Base 1.0

Oss R2 Spatial Sdxl Base 1

Vision adherence score 1.0/10. Five main sculptural objects plus stands: a tall brass wedge, a black pyramid, a small brass sphere, a red sphere, and a red globe on a pedestal. No required cubes or paper crane appear and the exact inventory fails.

Size
1344×768
Seed
not exposed
Method
local inference
View experiment →