Gemus-AI/DesignAnything-Results

Design Anything Results

The final outputs of Design Anything on all 228 tasks of the Design Anything benchmark (corpus v0.6), with two language models:

systemId	Language model	Runs
da-gpt	GPT-5.6 Sol High	228
da-deepseek	DeepSeek V4.1 Flash High	228

For every run this dataset holds the deliverables the agent produced, the canvas (the node graph the agent built), the export manifest and the official evaluation (evaluator v1.18). Results of the baseline systems in the paper are not included.

Related material:

• Task briefs, scoring specifications, the evaluation protocol and the evaluator source: https://github.com/Gemus-AI/DesignAnything
• The input images the tasks provide: https://huggingface.co/datasets/Gemus-AI/DesignAnything-Bench

Contents

Content	Files	Size
Deliverables (blobs/)	6,805	18.94 GB
Canvas preview images (blobs/)	5,418	0.25 GB
Canvas snapshots (workflow.json)	455	0.13 GB
Export manifests (manifest.json)	456	0.03 GB
Evaluations (evaluation.json)	456	0.16 GB
Sanitization records (sanitization.json)	456	0.01 GB

metadata/runs/da-gpt.jsonl                 one row per run (the dataset viewer shows these)
metadata/runs/da-deepseek.jsonl
runs/<systemId>/<taskId>/manifest.json     export manifest (the copy the evaluator scored), sanitized
runs/<systemId>/<taskId>/workflow.json     canvas snapshot, sanitized
runs/<systemId>/<taskId>/evaluation.json   official evaluation, sanitized
runs/<systemId>/<taskId>/sanitization.json what was changed, by JSON path
blobs/sha256/<first two hex digits>/<sha256>.<ext>   deliverables and canvas preview images, byte-identical
CHECKSUMS.sha256                           SHA-256 of every other file

There is a single split, results. The rows are not a train/test split; filter on systemId to compare the two models.

Row fields (metadata/runs/*.jsonl)

• Identity: systemId, model, taskId, runKey, runId, attemptId, runVerdict.
• Files of the run: manifestPath, workflowPath (null when the run has no canvas, with workflowUnavailableReason), evaluationPath, sanitizationPath, nodeCount, edgeCount.
• Digests: manifestSha256Original / manifestSha256, workflowSha256Original / workflowSha256, evaluationSha256Original / evaluationSha256 (before and after sanitization), and productManifestSha256Original (the manifest inside the run's own export, which the canvas belongs to).
• Evaluation: evaluatorVersion, rubricVersion, judge, taskSetSha, specSha, evaluatedAt, delivered, expected, constraintsPassed, constraintsApplicable, designQuality, scoredManifestEquivalence.
• How the run was produced: postRunPatched, replacedEarlierAttempt, operatorRepair, endpoint.
• slots: per deliverable slot, the deliverables the evaluator assigned to it and the representative one (paths into blobs/).
• outputs: every non-input file the export manifest lists (deliverables, canvas previews, page renders) with its role, size and SHA-256; page renders and previews not referenced by a final canvas have no sha256 / path because they are not included.
• inputs: the input images the run received, as ../DesignAnything-Bench/blobs/... paths; inputLicenses lists their licenses.

Paths such as blobs/sha256/... inside the JSON files are relative to the Results dataset root. Input image paths are relative links into ../DesignAnything-Bench/blobs/...; download both datasets to adjacent directories with these names. These links do not depend on a mutable Hub branch. The file names in outputs, in manifest.json and in evaluation.json are the original export names; outputs[].path maps included files to blobs/. Canvas assets[].file is also rewritten to its dataset path. This preserves references for audit; it is not an importable Gemus workflow or a standalone HTML site.

Verifying files

Every file under blobs/ is named after the SHA-256 of its bytes. CHECKSUMS.sha256 lists all files except itself; after downloading a fixed revision, run sha256sum -c CHECKSUMS.sha256 in the dataset root.

The evaluation records the digest of the manifest it scored (requirements.manifestSha). It equals manifestSha256Original in the run's row, not the digest of the sanitized manifest.json.

Sanitization

The three JSON files of each run were produced on a local benchmark machine. Before publication they were rewritten with rule set da-hf-sanitize-1; deliverables, preview images and input images were not modified.

Rule	What it replaces	Replacement
object-url	addresses of the local object store used during the runs	the object's path in this dataset (blobs/...), or a relative path into the adjacent Bench directory for input images; unmapped objects use <unavailable-object>
object-key	bare object-store keys	the same paths
local-path	local file paths in evaluator messages	<local-path>
asset-file	canvas asset file names	the corresponding dataset path
session-id	sessionId in the export manifest (an internal database id)	<redacted>

Rule	Replacements
object-url	47,342
object-key	24,583
local-path	1,089
session-id	456
asset-file	13,113

sanitization.json lists every changed JSON path, per file, with the rule and the number of replacements; it does not record the original text. The sanitized files are re-serialized with two-space indentation, so number formatting and escapes may differ from the originals while the content is the same apart from the listed changes.

Known gaps and caveats

• Runs without a canvas: da-deepseek/bench-da-story-base-a-n24. The run was resumed after an interruption, and the resumed export did not include a canvas. Its deliverables and evaluation are included; the canvas taken before the resume is not, because it does not contain the deliverables added by the resume.
• 36 da-gpt runs are manifest-equivalent: the manifest the evaluator scored and the manifest in the run's export differ only in exportedAt. The dataset uses the scored manifest, while the canvas belongs to the product export. Their exportedAt values therefore differ on these 36 runs. They are not marked identical.
• Design quality is available for six evaluations. The corrected Book A evaluations include M3 results; the other 450 evaluations remain not-run. Detailed M3 results and judge identity are retained in evaluation.json. This dataset does not fill missing values from the paper's aggregate design-quality columns.
• Book A evaluations were corrected. The six runs for bench-da-book-base-a-n06/n14/n24 use the 25 September 2026 rescoring against the astronomical-prints brief received by both Design Anything configurations. Other evaluations retain their original corpus identity. See the GitHub repository's corpus-version information and benchmark/KNOWN_ISSUES.md.
• Page renders are not included. For slide decks the evaluator looked at page images rendered at export time; evaluations still name those files. The fonts and support files used for that rendering are not included either, so re-rendering the HTML deliverables will not necessarily reproduce what the evaluator saw.
• One deck has a dead local link: an HTML deliverable of da-deepseek/bench-da-fashion-material-01 contains a link to the local object store used during the run. The deliverable is kept byte-identical; the link does not resolve.
• Relative links inside HTML deliverables point at the layout of the original export (../inputs/..., ../support/...) and do not resolve inside this dataset.
• The two systems were run and scored under different conditions: some runs had missing deliverables re-generated after the run (postRunPatched), and some runs replace an earlier failed attempt (replacedEarlierAttempt); da-gpt made videos with Seedance 2.0 and da-deepseek with MiniMax-H3; scoring ran on two machines whose judge caches were not shared.

systemId	postRunPatched	replacedEarlierAttempt	operatorRepair
da-deepseek	1	5	0
da-gpt	50	59	2

License

CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), The Design Anything Authors, to the extent the authors hold rights in the material. Many deliverables were generated by AI models. Input images are not part of this dataset; each keeps its own license in the Bench dataset. Where a run received a CC BY-SA or Open Government Licence – Toronto input image (inputLicenses contains CC-BY-SA or OGL-Toronto-1.0), a deliverable that adapts that image may also be subject to that license and its attribution requirement. The license grants no trademark rights; some briefs name real products or brands.

To report a problem with any file, or to request its removal, open a discussion on this dataset page.

Citation

Please cite the Design Anything paper. Citation details are on the project page: https://designanything.gemus.ai