HD scene comparison previews — October 4, 2026

Scene comparisons published October 4, 2026

The preview gallery now contains 31 pairs, retaining all 21 historical rows and their image hashes. Ten added pairs cover decorated reading rooms and night markets across GPT 6 Astra, GPT 6.1 Sol, Opus 5.5 and Sonnet 5.5, plus coastal cafés from Astra and Sol.

Eight pairs have two counterbalanced Astra reviews: 4 favor the baseline, 1 favor the plugin, and 3 have split votes or ties. Two market pairs lack valid reviews; no winner is assigned. These are one generation per condition and directional examples, not a general success rate.

Controls and retained evidence

• Same task, generator model, medium effort and 30-minute generation budget within each pair; pinned plugin skills/scripts versus an isolated baseline. MCP is absent in both.
• Blender 5.2.2 LTS. Frozen plugin fingerprint: 6c2f75fb8f57c83625f23e2a254f146bcd2b45569b5a920e0d5aefc05a114891.
• Six pairs completed before the CPU-resource halt; four completed afterward under the shared 20% CPU cap, Below Normal priority and one generation at a time. Pair metadata retain the resource policy and shared guidance hashes. Durations are descriptive and cannot rank latency across those policies.
• Technical scores use common scorer 9 against unchanged artifact hashes. Raw scorer 7 results remain in the downloadable data. The calibration permits finite render-only scene bounds and recognizes ordinary food names; geometry and export gates remain.
• The initial publication used existing authored camera evidence at its recorded resolution, without another Blender render pass. The HD update below replaces the display previews. Camera positions and lighting can differ between conditions. Saved source and models are unchanged. Astra market evaluator images were repaired from the same hashed source after an output-directory collision; those images have no new visual votes. Sol's original plugin market review-camera evidence remains unavailable; new HD display previews are recorded separately.
• All ten pairs with completed evaluator summaries in both conditions were included, regardless of scores, failed checks or visual preferences. Fourteen remaining planned pairs have no complete paired evaluator results: the Claude cafés and all game-character fixtures. Their original launch/evaluator failures remain archived; they are not counted as completed quality comparisons.

Added rows

Scores and technical gates read no plugin / with plugin. Votes read no plugin / with plugin / tie.

Task	Model	Technical score and gate	Votes
Complete decorated reading room	GPT 6 Astra	96 fail / 100 pass	1 / 1 / 0
Complete decorated reading room	GPT 6.1 Sol	96 fail / 93.5 fail	2 / 0 / 0
Complete decorated reading room	Opus 5.5	96 fail / 96 fail	1 / 1 / 0
Complete decorated reading room	Sonnet 5.5	100 pass / 100 pass	0 / 2 / 0
Complete low-poly night-market courtyard	GPT 6 Astra	96 fail / 100 pass	Unreviewed
Complete low-poly night-market courtyard	GPT 6.1 Sol	96 fail / 100 fail	Unreviewed
Complete low-poly night-market courtyard	Opus 5.5	100 pass / 100 pass	2 / 0 / 0
Complete low-poly night-market courtyard	Sonnet 5.5	100 pass / 100 pass	2 / 0 / 0
Decorated coastal cafe holdout	GPT 6 Astra	96 fail / 93.5 fail	2 / 0 / 0
Decorated coastal cafe holdout	GPT 6.1 Sol	96 fail / 100 pass	1 / 1 / 0

Workflow improvements

HD preview update

The initial publication reused the low-resolution review evidence. The updated
viewer displays only HD previews: all 60 authored camera views for the ten new
pairs render at a 1536-pixel longest edge. These are new renders from the same
hashed saved models, preserving camera aspect ratios, lighting, materials and
compositing. Cycles renders prefer OptiX and GPU OpenImageDenoise; authored Eevee
scenes retain their engine. Jobs run sequentially under the 20% CPU cap.

The new previews include the previously missing Sol room/market and Astra café
camera views. Original scores, votes and original evidence are unchanged. The
Images dropdown is removed. A missing view disables the slider and displays the
available image in full instead of stretching a one-pixel placeholder across it.

Observed topology failures in repeated lathed props led to a pole-sharing lathe helper and prototype validation before duplication. The native helper probe removed zero-length pole edges on two profiles while preserving bounds and face counts. This is a targeted topology correction; the unchanged gallery outputs do not demonstrate a visual improvement from the revised helper. Rendering guidance now distinguishes GPU rendering from GPU denoising, and benchmark jobs enforce the resource limits above.

Scene comparison gallery — October 4, 2026

Scene comparisons published October 4, 2026

The preview gallery now contains 31 pairs, retaining all 21 historical rows and their image hashes. Ten added pairs cover decorated reading rooms and night markets across GPT 6 Astra, GPT 6.1 Sol, Opus 5.5 and Sonnet 5.5, plus coastal cafés from Astra and Sol.

Eight pairs have two counterbalanced Astra reviews: 4 favor the baseline, 1 favor the plugin, and 3 have split votes or ties. Two market pairs lack valid reviews; no winner is assigned. These are one generation per condition and directional examples, not a general success rate.

Controls and retained evidence

• Same task, generator model, medium effort and 30-minute generation budget within each pair; pinned plugin skills/scripts versus an isolated baseline. MCP is absent in both.
• Blender 5.2.2 LTS. Frozen plugin fingerprint: 6c2f75fb8f57c83625f23e2a254f146bcd2b45569b5a920e0d5aefc05a114891.
• Six pairs completed before the CPU-resource halt; four completed afterward under the shared 20% CPU cap, Below Normal priority and one generation at a time. Pair metadata retain the resource policy and shared guidance hashes. Durations are descriptive and cannot rank latency across those policies.
• Technical scores use common scorer 9 against unchanged artifact hashes. Raw scorer 7 results remain in the downloadable data. The calibration permits finite render-only scene bounds and recognizes ordinary food names; geometry and export gates remain.
• Preview images use existing authored camera evidence at its recorded resolution, without another Blender render pass. Camera positions and lighting can differ between conditions. Saved source and models are unchanged. Astra market evaluator images were repaired from the same hashed source after an output-directory collision; those images have no new visual votes. Sol's plugin market camera evidence remains unavailable and is clearly marked.
• All ten pairs with completed evaluator summaries in both conditions were included, regardless of scores, failed checks or visual preferences. Fourteen remaining planned pairs have no complete paired evaluator results: the Claude cafés and all game-character fixtures. Their original launch/evaluator failures remain archived; they are not counted as completed quality comparisons.

Added rows

Scores and technical gates read no plugin / with plugin. Votes read no plugin / with plugin / tie.

Task	Model	Technical score and gate	Votes
Complete decorated reading room	GPT 6 Astra	96 fail / 100 pass	1 / 1 / 0
Complete decorated reading room	GPT 6.1 Sol	96 fail / 93.5 fail	2 / 0 / 0
Complete decorated reading room	Opus 5.5	96 fail / 96 fail	1 / 1 / 0
Complete decorated reading room	Sonnet 5.5	100 pass / 100 pass	0 / 2 / 0
Complete low-poly night-market courtyard	GPT 6 Astra	96 fail / 100 pass	Unreviewed
Complete low-poly night-market courtyard	GPT 6.1 Sol	96 fail / 100 fail	Unreviewed
Complete low-poly night-market courtyard	Opus 5.5	100 pass / 100 pass	2 / 0 / 0
Complete low-poly night-market courtyard	Sonnet 5.5	100 pass / 100 pass	2 / 0 / 0
Decorated coastal cafe holdout	GPT 6 Astra	96 fail / 93.5 fail	2 / 0 / 0
Decorated coastal cafe holdout	GPT 6.1 Sol	96 fail / 100 pass	1 / 1 / 0

Workflow improvements

Observed topology failures in repeated lathed props led to a pole-sharing lathe helper and prototype validation before duplication. The native helper probe removed zero-length pole edges on two profiles while preserving bounds and face counts. This is a targeted topology correction; the unchanged gallery outputs do not demonstrate a visual improvement from the revised helper. Rendering guidance now distinguishes GPU rendering from GPU denoising, and benchmark jobs enforce the resource limits above.

Vanilla vs plugin modeling pilot — October 2, 2026

Curated preview evidence for the no-plugin versus Blender Agent Studio skill-workflow pilot: eight pairs, four models, two tasks, seven views per condition. Includes original image hashes, run provenance, technical checks, and counterbalanced visual judgments.

No-plugin baselines were collected later and paired with existing frozen plugin runs. One generation per condition; no general model ranking or MCP-effect claim. Raw runs, agent traces, models and exports remain outside this public archive.

This archive is separate from the earlier current-skills versus added-guidance pilot.

Spatial modeling pilot — October 2, 2026

Curated gallery evidence for the October 2 spatial modeling pilot: 16 generations across GPT 6 Astra, GPT 6.1 Sol, Opus 5.5 and Sonnet 5.5, comparing current skills with added spatial checks.

The archive contains eight paired comparisons, 112 standardized PNG views and the comparison JSON with raw/corrected structural scores, image and artifact hashes, and counterbalanced visual reviews. It excludes Blender models, raw agent traces, downloaded reference images and local paths.

The results do not establish a general improvement: six pairs favored the current workflow, one favored added checks, and one split. The whole-scene benchmark is implemented and calibrated, but its model campaign has not been run.

SHA-256 (spatial-pilot-gallery.tar.gz): 70b2e0ef45a9c48869d50a0a0d334cbe5782dc125aa0e4ff5cbe229926587df1

This is an evidence-data release, not a new plugin version. The docs deployment restores this exact archive through a pinned checksum.

Complete benchmark model matrix — October 2, 2026

The gallery now covers five tasks with all four requested models: GPT 6 Astra, GPT 6.1 Sol, Opus 5.5 and Sonnet 5.5. Ten new paired entries supplement the existing results.

The historical GPT 6 Sol lantern shows the owner-selected run 3 only. Earlier attempts remain in the previous gallery release. This curated example is not an average or a random sample.

This archive contains 21 pairs and 656 images: denoised HD previews, reviewed views, and original submission views for the new matrix entries. New review framing hides named studio meshes; saved models and technical failures are unchanged. These separate historical cohorts do not establish an aggregate model ranking.

Expanded HD benchmark gallery — October 2, 2026

Expanded no-plugin / Blender Agent Studio comparison gallery: thirteen pairs, including three signal lantern runs, a lever press and a drawbridge.

Contains 172 new 1536-pixel Cycles previews with 64 samples and OpenImageDenoise, plus 172 original review images. HD staging exclusions and source hashes are recorded per condition. Historical scores and votes refer to the original evidence; models were not regenerated or repaired.

Models and study cohorts remain separate. One lantern repeat favored the no-plugin result. These small samples are not a general success rate.