Where models compete on real creative work. Commissioned battles and long-form research, drawn from the network for creative judgement.
The Human Craft Benchmark The first eval that scores models the way working creatives do.Best performing models
Bradley–Terry leaderboard pooled across 9 image-generation studies and 240 completed tournaments.
Higher Elo means stronger aggregate head-to-head performance across the studies in the pool.
Methods & standards
See how we score these models → Methodology
Latest research
28 studiesEvery battle, model profile, and field note, sorted by newest first, tagged by domain.
10 designers, 4 models, 240 images. Spelling is solved. Typographic craft and client-readiness are where Lumen v4 pulls away.
ReadTypography win-rate, pooled
Share of head-to-head typography prompts · 4 image models · 20 prompts × 12 iterations · pooled 1 – 5
11 production-style sessions. Prism made the edit 75% of the time, kept the rest of the frame intact 58% of the time.
ReadPrism controllability checks · 11 sessions, four tasks per session, both graders in agreement
5 designers, 5 openings, 1 luxury brief. The first prompt set what each session could reach.
ReadSpecificity score (0–1)
First-prompt specificity by participant, blind-scored · 5 designers, 1 brief
5 sessions, 5 designers, 1 real-world client brief. Strong as a starting structure, breaks under precision edits.
ReadAtlas Design designer evaluation · 5 designers, 1 real-world client brief
12 models, 5 creative domains. One repeated complaint from working evaluators: the work all looks the same.
ReadConvergence themes coded across 5 domains and clustered by two independent graders
Survey of 340 high-earning independent creatives. What they actually do with AI on real client work.
Readof AI output makes it to final deliverables
The rest is stripped, redrawn, or replaced
Self-reported survey · 340 independent creatives · fielded over six weeks
Vellum connects working creative minds with the teams training models to understand taste. This is expert input, not crowd labour — the creative layer under the next generation of tools.