Spaces:
Running
Running
Commit History
Yesterday
-
Text-to-Scene: Gemma 4 31B (thinking on) cost estimated from its logged tokens
Verified signatureGemma 4 31B ran on self-hosted weights, so nothing was billed. Its cost is now the token counts its agent logged on every model call (scripts/t2s_token_cost.py) at Gemma 4 31B's OpenRouter list price ($0.20 input, $0.40 output per million): median $0.056 per case, shown as $0.06*, with the basis in the table, tooltips, Pareto caption and gallery. It now has a Pareto point, on the frontier. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>391a498 -
Text-to-Scene: four more finished runs (19 configurations); Xiaomi mark; open-weights tags
Verified signatureSnapshot 2026-10-16 16:01 UTC of the internal leaderboard: GPT-6 Luna (max), Muse Spark Open (max), MiMo V2.6 Pro (max) and Gemma 4 31B (thinking on) have all 20 cases scored and join the bars, Pareto, ranking, tooltips and gallery (19 configurations); Inkling has not finished and stays out. Xiaomi gets its own mark (Simple Icons) and the internal leaderboard's amber; Muse Spark Open and MiMo V2.6 Pro are tagged open weights (OpenRouter lists their Hugging Face weights). Gemma 4 31B ran on self-hosted weights, so it has no billed cost and no Pareto point, and the pages say so. Scores never rounded up; median cost per case. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>0c7052b -
Gallery links name gallery/index.html: the static host does not serve a folder's index
Verified signature"gallery/" redirected to huggingface.co/gallery, so the menu item, the Open-the-gallery buttons and the gallery's own Gallery / By agent links went to the wrong site. Every link now names the page, as the rest of the site does with index.html#.... Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>b53e86d -
Text-to-Scene: Muse Wonder Full 1.0 cost from its CLI's own token logs
Verified signatureThe harness stream of the muse CLI has no usage, so the internal table left Muse Wonder Full without a cost. The CLI's own session logs record every model call's provider token counts (they match its cost ledger); scripts/t2s_muse_cost.py prices them at Muse Wonder's OpenRouter list rate ($4 input, $0.20 cached input, $20 output per million; Wonder Full is sold by subscription), with the CLI's background model Muse Spark 1.3 at its own rate: median $3.66 per case. The skill-reminder side calls are not logged readably and are left out. Ranking, Pareto, tooltips and gallery updated; the pages say how this cost is derived. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>b85bdcb -
Text-to-Scene: the October 2026 re-run replaces the paper's Research Track numbers; latest cases; gallery
Verified signatureLeaderboard (Text-to-Scene tab): bars, Pareto, ranking and tooltips from the internal leaderboard's sandboxed re-run, only configurations with all 20 cases scored; scores never rounded up, median cost per case. Cases: the four-agent comparison videos with a scene picker; the Industrial Harbor showcase stays. Gallery: every shown configuration on all 20 scenes, by agent and by scene. Image-to-Scene, the Task Format example, the Takeaways and the case explorer keep the paper's evaluation. Rebuild: scripts/t2s_sandbox_round.py and scripts/t2s_gallery.py (see README). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>988be21
12 days ago
-
Industrial Harbor card on the same live three.js viewer as the other cases
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>e998ba1 -
Four more Image-to-Scene cases: Suburban Umbrella Pair, Rosie's Service Bin, Rosie's Roadside Billboard, Suburban Driveway Mailbox
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>cca5a6d -
Feature GPT-6 Astra's Industrial Harbor in the Text-to-Scene cases
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>c3746b4 -
Text-to-Scene / Image-to-Scene tabs in every section; restore the Image-to-Scene task, views, cases, takeaways and models
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>5bf5905
16 days ago
-
Name the evaluation tracks: Research Track (20 scenes) and Full Benchmark (160 scenes)
Verified signatureThe leaderboard's Text-to-Scene results are the Research Track's; the case-count strip states the two tracks, and the leaderboard section says why the smaller track exists. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>7399765 -
Draw the charts in the paper figures' kit style
Verified signatureTakeaway: the requirement-family dot plot becomes the kit's grouped bars (Spatial Composition in the agent's company colour, the other families grey, logos and rotated names), with hover details; the mismatch chart and the judge's checks take the kit's neutrals; the evidence scene gets a company-colour chip. Leaderboard: 'Text-to-Scene' qualifier and the Qwen glyph. Pareto: y axis is the Text-to-Scene score; effort in title case. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>86e4992 -
Text-to-Scene only: drop the Image-to-Scene task, views, cases, takeaways and models; v31 teaser
Verified signatureThe page now shows the construction setting alone (user, 2026-10-01). scripts/t2s_only.py makes the page edit; the Image-to-Scene models, their progressive geometry and texture bundles, media/i2s and the old teaser are removed.96718cc -
Render geometry before progressive textures and self-host Three.js
c77f6d9