PaperlayerView exact revision on arXiv

Source-linked AI summary

Mine, Sample, Render: Test-Time Diversity Alignment Of Generative Models

Kostas Triaridis, Morris E. Sharp, Nebojsa Jojic

arXiv:2610.21185v1cs.AIcs.LG

TL;DR

Generative models can produce plausible outputs yet repeatedly resolve unspecified choices in similar ways, limiting coverage of real-data variation. MSR mines and samples corpus-derived attribute configurations to guide generation, and evaluations across five domains report improved diversity and alignment with real distributions.

  • Problem

    Generative models often produce a narrower range of variation than real data, and prompt changes or requests for diversity do not specify which choices should vary or how often.

    Source: §1 · p. 1

  • Method

    MSR mines an interpretable attribute schema and joint distribution from a reference corpus, samples attribute configurations, and uses a generative model as a conditional renderer.

    Source: §1 · p. 1, §2 · p. 2

  • Results

    Across weather reports, user simulation, web design, image generation, and agent-created 3D assets, MSR improves output diversity and alignment with real distributions.

    Source: §1 · p. 2, §4 · p. 10

  • Takeaways & Limitations

    The paper concludes that MSR provides an interpretable mechanism for capturing real-world variability in generated populations.

    Source: §4 · p. 10

Abstract

from arXiv · show

Modern generative models produce high-quality individual outputs but struggle to sample diversely, even when explicitly instructed to do so. We introduce MINE-SAMPLE-RENDER (MSR), a test-time approach that externalizes the diversity into an explicit, corpus-derived sampling procedure. Our pipeline (i) mines an interpretable categorical schema from a corpus of real artifacts and supplies statistics for sampling; (ii) samples attribute configurations from the corpus-derived statistics; (iii) uses a generative model purely as a conditional renderer of a sampled configuration. Across weather report generation, user simulation, webpage design, text-to-image generation, and agent-generated 3D assets, we evaluate the diversity and output quality of MSR-guided generation and show that MSR improves alignment with corpus distributions and diversity of generated outputs.

1 INTRODUCTION

Generative models can produce plausible outputs while repeatedly resolving unspecified choices in familiar ways, leaving generated populations narrower than real data. MSR mines corpus attributes and their joint distribution, then samples configurations to guide generation; evaluations span five domains and report improved diversity and distributional alignment.

1 INTRODUCTION

  • Five baseline outputs across webpages, text-to-image, and 3D chair generation illustrate how models can repeatedly make familiar choices despite prompt changes.The paper frames this as a gap between convincing individual outputs and coverage of real-data variation.Source: §1 · p. 1, Figure 1 · p. 2
  • MSR mines attributes from a reference corpus, estimates their joint distribution, and samples configurations for a generative model or agent to render.The approach separates what can be generated from how often it should be generated.Source: §1 · p. 1
  • MSR’s distinguishing contribution is discovering compositional, artifact-level attributes through direct cohort comparisons and sampling their corpus-estimated joint distribution.The paper contrasts this with related approaches including fLSA (Xu et al., 2025a).Source: §1 · p. 2
  • Across weather reports, user simulation, webpages, text-to-image, and agent-generated 3D assets, MSR improves diversity and alignment with real distributions.The reported 3D results also include increased asset quality.Source: §1 · p. 2

2 MINE-SAMPLE-RENDER

MSR mines an interpretable attribute schema from source artifacts, estimates attribute statistics, and uses sampled configurations to guide a renderer. Its sampling methods use corpus-derived associations, with conditioning features linking ordinary-language requests to attributes.

2 MINE-SAMPLE-RENDER

  • The pipeline mines a corpus schema, estimates attribute statistics, samples a configuration, and conditions a renderer on it together with the prompt.The schema records artifact properties as attributes, whose granularity depends on the mining instructions and corpus.Source: §2 · p. 2, Figure 2 · p. 3

2.1 MINING ATTRIBUTES BY COMPARISON

  • LLMs compare small artifact cohorts to find shared or varying attributes, merge synonymous attributes, and label the corpus to estimate frequencies and dependencies.The labels form a sparse occurrence matrix over artifacts and attributes.Source: §2.1 · p. 3

2.2 SAMPLING SPARSE CONFIGURATIONS

  • Ordinary-language requests condition sampling through required or compatible attributes and lexical features, while the sampled configuration contains mined attributes only.Explicit requirements may be fixed; lexical matches provide statistical evidence when a request does not match a complete attribute description.Source: §2.2 · p. 3

2.2.1 SAMPLING WITH A TOPIC MODEL

  • Topic-based sampling draws a topic from request-matched features, then samples mined attributes according to that topic’s attribute loadings.The topic model uses a positive-NPMI association matrix factorized into topics.Source: §2.2.1 · pp. 3–4

2.2.2 ANCESTRAL SAMPLING

  • Ancestral sampling draws attributes sequentially from empirical conditionals, dropping lower-ranked dependencies until at least m source samples remain.The method also permits a mixture of ancestral and topic-based sampling, controlled by λ.Source: §2.2.2 · p. 4

3 EXPERIMENTS

Across five settings, MSR uses corpus-derived attributes to improve diversity or alignment with real-data variation. The experiments test attribute recovery, user simulation, webpages, images, and functional 3D designs.

3.1 ATTRIBUTE RECOVERY FROM WEATHER REPORTS

  • 0.570 to 0.181 Forward KL, 0.484 to 0.203 JS, and 21% to 68% coverage: MSR sampling better matches held-out weather configurations.In this controlled setting, the mined schema was almost identical to the data-generating schema.Source: §3.1 · p. 4

3.2 USER SIMULATION

  • MSR improves the simulated-user distribution across tested LLMs, while PersonaHub personas yield limited behavioral variation.The comparison concerns simulated coding-task users and real user behavior.Source: §3.2 · p. 5, Table 1 · p. 5

3.3 WEBPAGE GENERATION

  • MSR increases visual and structural webpage diversity over prompt-only generation across both schemas and all tested models.Schema-conditioned pages account for 95.1–97.7% of odd-one-out removals for GPT-5.5 and 75.6–80.7% for Kimi K2.6.Source: Table 2 · p. 6, §3.3 · p. 7

3.4 TEXT-TO-IMAGE GENERATION

  • 6.68–6.77 versus 4.32–4.55: MSR–Resolved reaches higher DINOv2 Vendi scores than coordinated rewriting across three image generators.It also has the lowest mean odd-one-out removal rank for every generator; all 500 resolved images are removed for GPT Image 2 and FLUX.2-flex, and 499 of 500 for MAI-Image-2.5.Source: §3.4 · p. 8, Table 3 · p. 8

3.5 AGENTIC 3D ASSET GENERATION

  • MSR produces the most diverse folding-chair designs without reducing quality, while novelty prompting increases variety at a quality cost.The experiment tests a tool-using agent creating functional 3D chairs.Source: §3.5 · pp. 8–9, Table 4 · p. 9

4 CONCLUSION

MSR uses corpus-derived attribute configurations to separate population-level sampling from conditional rendering. The paper positions this approach against related work and describes a sampler designed for sparse configurations.

4 CONCLUSION

  • MSR is a training-free framework that mines interpretable categorical schemas, samples attribute configurations from real-world corpora, and uses a generative model as a conditional renderer.Across weather, user simulation, web design, image generation, and agent-created 3D assets, the authors report improved diversity and alignment with real distributions.Source: §4 · pp. 9–10

A RELATED WORK

  • Unlike fLSA’s segment tags and GenZ’s features for predicting y, MSR models joint distributions over compositional attributes and renders artifacts from sampled configurations.It proposes semantic coordinates through direct comparisons within small cohorts, then uses global labeling to estimate their frequencies and dependencies.Source: A RELATED WORK · p. 16

B.1 ANCESTRAL SAMPLING

  • For sparse configurations, the sampler uses a fresh random attribute ordering, retains supported high-mutual-information conditions, smooths conditionals, and masks incompatible values.The authors describe this as a pragmatic combination of established techniques, not a new claim about ancestral sampling, mutual information, or backoff.Source: A RELATED WORK · p. 17, §B.1 · p. 17

A RELATED WORK

  • MSR induces its semantic control space and corpus-estimated configuration distribution, unlike prior structured-generation approaches that separate structure from realization.The related-work discussion distinguishes generating varied pages for the same request from optimizing the fidelity of one rendered page.Source: A RELATED WORK · p. 17
  • For open-ended outputs without a unique target distribution, the paper identifies reference-free effective-diversity measurement and quality-filtered semantic diversity as distinct evaluation tools.It motivates assessing distributional fidelity, set diversity, condition adherence, and per-output acceptability separately.Source: A RELATED WORK · p. 17

B.2.1 DATASET CONSTRUCTION

The weather experiment uses city-conditioned reports based on discretized observations to test schema recovery and sampling. It also specifies how configurations are generated and how valid outputs are retained.

B.2.1 DATASET CONSTRUCTION

  • The controlled dataset uses 2016 training and 2017 held-out weather observations from ten cities, discretized into four categorical weather axes and represented by 16 binary attributes.City indicators are conditioning variables, not scored weather axes; the source record has 26 binary attributes in total.Source: §B.2.1 · p. 18

B.2.2 RECOVERED SCHEMA AND ANNOTATION

  • The recovered schema distinguishes all four temperature values and three values each for sky, wind, and precipitation, but does not recover windy as a separate value despite its source support.Foggy and storm have zero support in source configurations; the comparison cannot determine whether windy was lost during source rendering or mining.Source: Table 5 · p. 18, §B.2.2 · p. 19

B.2.3 CITY-CONDITIONAL ATTRIBUTE SAMPLING

  • City remains a conditioning variable while the sampler draws the four free weather axes in random order, retaining mutual-information-ranked conditioners only when at least 10 source rows remain.The sampler uses additive smoothing with β = 1; model-judged compatibility constraints reject snow with hot or warm temperatures but are not physical-impossibility guarantees.Source: §B.2.3 · p. 19

B.2.4 GENERATION CONDITIONS AND VALID OUTPUTS

  • The prompt-only, append, and MSR conditions each generate 300 reports per city, while generated reports lacking a complete extracted four-axis tuple are excluded from distributional scoring.MSR supplies sampled weather values and asks the renderer to preserve unusual combinations; the append baseline receives the source reports and is asked not to copy them.Source: §B.2.4 · p. 19

B.2.5 HELD-OUT DISTRIBUTIONAL EVALUATION

The held-out evaluation compares generated weather configurations against each city’s 2017 joint distribution. It reports divergence and coverage using extracted labels, with results averaged across cities.

B.2.5 HELD-OUT DISTRIBUTIONAL EVALUATION

  • Generated reports are scored against each city’s empirical 2017 joint distribution of discretized temperature, sky, wind, and precipitation.The candidate distribution uses tuples extracted from generated text, so scoring reflects expressed conditions rather than only requested configurations.Source: §B.2.5 · pp. 19–20
  • Forward KL uses a half-uniform mixture over the union of occupied joint cells, whereas Jensen–Shannon divergence uses unsmoothed distributions and natural logarithms.Zero-probability terms are omitted for Jensen–Shannon divergence.Source: §B.2.5 · p. 20
  • Coverage is the unweighted fraction of reference-occupied joint cells also occupied by generated outputs, and each metric is averaged unweighted across ten cities.Coverage is not probability-mass-weighted and is not the fraction of reports successfully labeled.Source: §B.2.5 · p. 20

B.3.1 EVALUATION AND SCHEMA-CONSTRUCTION DATA

The user-simulation evaluation separates schema construction from its reference conversations, mines behavioral attributes, and samples profiles for matched simulation conditions. Its setup includes two schema recipes and a fixed conversation harness.

B.3.1 EVALUATION AND SCHEMA-CONSTRUCTION DATA

  • The evaluation uses 5,000 real conversations, while schema-training candidates are held out from that reference set and filtered to remove duplicates and very long conversations.A frozen sample of 600 English-eligible conversations is used to learn both schemas, and reconstructed goals are reused across methods.Source: §B.3.1 · p. 20

B.3.2 SCHEMA MINING AND CONSOLIDATION

  • The miner proposes topic-agnostic binary behavioral attributes from real-user conversations, consolidates synonymous proposals, and back-applies the vocabulary to all 600 training conversations.The compact recipe yields 40 axes and 84 attributes; the fine-grained recipe yields 292 axes and 465 attributes.Source: §B.3.2 · pp. 20–21

B.3.3 ATTRIBUTE FREQUENCIES AND CO-OCCURRENCE

  • The attribute and cross-axis pair tables characterize source support across 600 conversations; common pairs can have lift near one when both attributes are individually frequent.The tables describe source behavior, not whether generated conversations realize requested attributes.Source: §B.3.3 · p. 21, Table 7 · p. 21, Table 8 · p. 21

B.3.4 ANCESTRAL PROFILE SAMPLING

  • Profiles are sampled with a fresh axis ordering, retaining a previous assignment as a conditioner only when at least 10 source conversations remain.The sampling manifold includes an explicit absent value and removes realizations with support below eight; each schema produces 5,000 frozen profiles reused across models.Source: §B.3.4 · p. 22

B.3.5 SIMULATION ARMS AND CONVERSATION HARNESS

  • Matched simulation conditions compare prompt-only, PersonaHub, and compact and fine-grained MSR profiles while keeping the assistant fixed to GPT-5.4.MSR profiles include non-absent attributes, grounded where available by real-user quotations; conversations stop at termination, ten simulated user turns, or a 30,000-character limit.Source: §B.3.5 · p. 22

B.3.6 MATCHED EVALUATION SETS

The evaluation uses matched identifiers across five arms within each user model, while allowing provider-specific policy exclusions.

B.3.6 MATCHED EVALUATION SETS

  • For each user model, examples missing from any arm, empty, policy-blocked, or unannotatable are removed from all five arms to preserve paired comparisons.Different providers can therefore have different policy exclusions.Source: §B.3.6 · p. 22
  • Table 9 reports the final effective sample sizes for the user simulators.Source: Table 9 · p. 23

B.3.7 EVALUATION

Evaluation follows Mehri et al. (2026), using GPT-5.4 to annotate six facets and text-embedding-3-large for embeddings.

B.3.7 EVALUATION

  • The evaluation follows Mehri et al. (2026), with GPT-5.4 at low reasoning effort annotating six facets and text-embedding-3-large supplying the first 1,024 embedding dimensions.Source: §B.3.7 · p. 23

B.4.1 TASKS AND EXPERIMENTAL CONDITIONS

This section describes webpage and image-generation experiments using corpus-derived attribute schemas and sampling to guide conditional rendering, including webpage validation and evaluation. It also covers Direct and Sol-assisted rendering conditions, including how Sol translates sampled properties into visual instructions and realizes controls.

B.4.1 TASKS AND EXPERIMENTAL CONDITIONS

  • The webpage experiment compares prompt-only generation with WebSight- and Design2Code-conditioned MSR across 100 tasks, two generators, and 6,000 accepted pages.Each task uses the same authoring request, grounding text, and available images across conditions; ten pages are generated per condition. The prompt-only template omits the sampled design-system block.Source: §B.4.1 · p. 23

B.4.2 SCHEMA MINING AND CONSOLIDATION

  • WebSight consolidation yields 485 attributes from 1,000 pages, while Design2Code mining and consolidation yields 5,484 canonical attributes from 250 mining pages.WebSight uses cohort-mined properties, entailment checks, and complete-link grouping, then annotates 9,000 additional pages. Design2Code mines consequential page decisions and consolidates equivalent names while preserving component identity, atomicity, and transferable meaning; its fixed vocabulary is annotated across 480 benchmark pages.Source: §B.4.2 · pp. 23–24

B.4.3 SCHEMA SAMPLING AND REQUEST-SPECIFIED ATTRIBUTES

  • Webpage designs are sampled from frozen empirical matrices, then request-specified attributes are enforced after a full-matrix draw to avoid collapsing support for rare attributes.The WebSight matrix covers 10,000 pages and 485 attributes; Design2Code uses 480 pages and 5,484 attributes. The sampler uses additive smoothing β = 1.0 for WebSight and β = 0.01 for Design2Code, with minimum support m = 10; Kimi reuses the same saved designs and GPT-5.5-derived specified attributes.Source: §B.4.3 · p. 24

B.4.4 GENERATION, VALIDITY CHECKS, AND RECOVERY

  • Generated pages must be self-contained HTML/CSS and pass static resource and safety checks before Chromium rendering; pages with broken images or under 100 rendered body-text characters are rejected.Rendering disables JavaScript and external requests and suppresses animations; the checks establish basic renderability, not content fidelity or usability.Source: §B.4.4 · p. 24, §B.4.5 · p. 25

B.4.6 ODD-ONE-OUT EVALUATION

  • Odd-one-out evaluation compares schema-conditioned and prompt-only top-view screenshots, while diversity embeddings represent webpage structure or combine top, middle, and bottom visual views.The odd-one-out judge sees neutral-labeled screenshots without HTML, schema identifiers, or condition labels; schema removal share and mean removal rank measure stylistic distinctiveness, not quality preference or task success. Structural representations omit literal copy and asset details, while visual representations use DINOv2-G.Source: §B.4.6 · p. 25, §B.4.7 · p. 25

B.4.8 ATTRIBUTE FREQUENCIES AND CO-OCCURRENCE

  • In the saved design draws, type grotesque sans occurs in 71.90% of WebSight designs and body typography sans serif in 80.90% of Design2Code designs.These draws are shared across the two generators. The source-page tables describe prevalence and co-occurrence rather than generated-page realization or strong dependence; highest-lift pairs are ranked separately subject to at least 1% joint support.Source: §B.4.8 · pp. 26–27, Table 10 · p. 26, Table 11 · p. 26

B.5.1 MINING

  • For image generation, MSR mines 28,875 attributes from 7,879 COCO images and samples 20 compatible controls by mixing empirical and prompt-conditioned topic distributions.The sampler selects from a request-specific pool of at most 250 attributes, and controls guide underspecified choices without exhaustively specifying the scene. The empirical and topic components are mixed at each selection step; the λ = 0 endpoint uses the same implementation.Source: §B.5.1 · p. 27, §B.5.2 · p. 28, §B.5.3 · p. 29

B.5.3 RENDERING

  • 80.2% of originally sampled attributes were realized with Sol-assisted rendering, versus 61.6% with direct rendering; Sol reconciled 99.2% of sampled attributes.Both variants were reported as producing more diverse images than prompt-only and prompt-rewriting baselines while maintaining high visual plausibility. In the Sol-assisted procedure, GPT-5.6 Sol assigns entities to role placeholders and produces coherent instructions for GPT Image 2; realization rates come from a post-hoc model-based audit.Source: §B.5.3 · p. 29
  • X and Y are local to each property, allowing different visible entities to be bound without giving the image model globally inconsistent symbolic variables.Source: §B.5.3 · p. 30
  • For each sampled property, Sol returns an auditable JSON row with a use decision, local X/Y bindings, a visual instruction, and a rationale.Source: §B.5.3 · pp. 29–30
  • Sol renders the concrete scene plan and retained instructions under the same primary-prompt-priority contract as Direct.Source: §B.5.3 · p. 30
  • At λ = 0.5, Sol retained 9,920/10,000 controls (99.2%); 32 configurations required documented moderation repairs.The repairs removed minor-specific optional content and rebound safe properties to adults or objects.Source: §B.5.3 · p. 30

B.5.4 EXPERIMENTAL DESIGN

The image evaluation compares four generation arms across 50 prompts and assesses diversity alongside prompt satisfaction, visual acceptability, and attribute realization.

B.5.4 EXPERIMENTAL DESIGN

  • The experiment uses one frozen prompt from each of 50 distinct categories and generates 1024 × 1024 images under the stated generator settings.Independent and coordinated rewriting vary composition, lighting, focus, viewpoint, and treatment while preserving stated facts.Source: §B.5.4 · p. 30, §B.5.5 · p. 30

B.5.6 FOUR-ARM REPEATED ODD-ONE-OUT REMOVAL

  • The blinded odd-one-out evaluation pools ten images from each of four arms, repeatedly removes the image judged most unlike the others, and leaves ten survivors after 30 rounds.Lower mean exit rank indicates greater diversity; images surviving all removals receive rank 40.Source: §B.5.6 · p. 32

B.5.7 VALIDITY AND ATTRIBUTE REALIZATION

  • GPT-5.6 Sol judges prompt satisfaction, visual acceptability, and attribute realization for candidates 1 and 6 across 100 images per variant and generator.Attribute-realization scores estimate realization of originally sampled attributes, not fidelity to resolved instructions.Source: §B.5.7 · p. 32
  • 95–99% of audited Direct images and 90–97% of Resolved images meet both validity criteria, though resolution lowers validity most for FLUX.2-flex, from 95/100 to 90/100.At an assumed 90% per-attempt success rate, screening and resampling imply about 11% additional generation cost, plus screening.Source: §B.5.7 · pp. 32–33, Table 14 · p. 33

B.5.8 LAMBDA ABLATION

The λ ablation finds that the midpoint retains empirical guidance while matching topic-endpoint diversity under CLIP and exceeding it under DINOv2. The folding-chair corpus and its mined properties reveal coherent design patterns alongside room for recombination.

B.5.8 LAMBDA ABLATION

  • With semantic resolution, the midpoint and topic endpoint tie under CLIP, while the midpoint is higher under DINOv2 and retains empirical guidance without giving up pure-topic diversity.Source: §B.5.8 · p. 33

B.6.1 DATA COLLECTION

  • The curated corpus contains 160 design/appearance records and 305 evidence images, including 159 folding chairs and one demountable chair.Collection expanded beyond a seven-image ABO pilot through targeted searches for additional chair families and appearances; records came from ABO, Wikimedia Commons, and three museum sources.Source: §B.6.1 · p. 34

B.6.2 MINING RECOMBINABLE PROPERTIES AND DISCOVERING CORRELATIONS

  • Observed chair-property associations are recurrent design regularities, not hard constraints; for example, metal-looking feet co-occur with tubular frames in 95.3% versus 6.3% when absent.The association table uses records where both properties are observed and family weighting; alternatives are not ruled out.Source: §B.6.2 · p. 35
  • Mining retained sparse properties without recoding them as absent, then fit six interpretable topics using observed entries and family weighting.The six-topic inspection model treated unknown cells as omitted from the likelihood; six topics was an inspection choice, not a validated optimum.Source: §B.6.2 · p. 35
  • The fitted topics combine construction, attachment, and surface treatment into coherent design vocabularies, while allowing chairs to mix topics and form new combinations.Examples include textile camping frames, timber slats, carved timber, striped slings, laced lounge chairs, and molded seating.Source: §B.6.2 · p. 35

B.6.3 MODEL OUTPUT

The model receives a short, open-ended folding-chair assignment, while MSR supplies part-level attributes without specifying dimensions or mechanism. The illustrated agent builds and animates an articulated chair, checks selected motion constraints, and records departures from sampled properties.

B.6.3 MODEL OUTPUT

  • The baseline asks for a functional 3D folding chair, an opening-and-closing demonstration, and a report of checks and limitations, while leaving design and construction open.Agents were also required to construct geometry rather than import chair models and save an editable Blender scene, GLB, and demonstration outputs.Source: §B.6.3 · pp. 35–36
  • MSR briefs specify 63–111 positive part-level attributes, but leave unlisted properties unspecified and do not prescribe dimensions, joint locations, or a folding mechanism.Attributes cover parts including the frame, seat, back, arms, feet, and hardware, along with materials, colors, textures, construction, and spatial relationships.Source: §B.6.3 · p. 36
  • The example agent built a Blender chair with linked articulated parts, programmed opening and folding, checked closure and selected clearances, and documented changes where its realization conflicted with sampled properties.The programmed fold includes releasing retaining hooks and lifting the chair; Figure 8 depicts open, intermediate, and folded states, not physically verified performance.Source: §B.6.3 · pp. 36–37, Figure 8 · p. 37

B.6.4 QUALITY EVALUATION

The section describes a blinded, judge-scored chair-quality evaluation and provides visual comparisons for webpage and image generation. It also illustrates MSR use beyond the COCO categories from which its image schema was mined.

B.6.4 QUALITY EVALUATION

  • A fresh judge froze five criteria, weights, and scoring anchors before reviewing anonymous chair outcomes, then scored each criterion from 0 to 5 in half-point increments.The judge evaluated the chair rather than presentation polish or generation process; Table 15 reports cohort means by criterion.Source: §B.6.4 · p. 37, Table 15 · p. 37

C.1 WEBPAGE GENERATION

  • Figures 9 and 10 compare prompt-only webpages with MINE–SAMPLE–RENDER outputs using Websight and Design2Code schemas: Figure 9 uses a manufacturing-training-hub prompt, and Figure 10 an Axium-hackathon-recap prompt.Source: Figure 9 · p. 38, Figure 10 · p. 39

C.2 TEXT-TO-IMAGE

C.2.1 OUT-OF-DISTRIBUTION EXAMPLES

  • Although mined from COCO images, the schema includes general composition and co-occurrence information that the authors say is useful for images outside the aligned COCO categories.Figure 14 compares prompt-only and MSR images for a dragon swinging a bat with a dog watching.Source: §C.2.1 · p. 42, Figure 14 · p. 42
Loading 2610.21185v1…