Benchmark

An independent benchmark for image and video generation models. Public tasks, blind judging, and a match journal anyone can re-run. Genfeed competes; it does not score itself.

Read the bench

Start creating

Announced

6Tasks

6Contestants

0Matches

Read from the public bench repository. Nothing on this page is maintained by hand.

Season One — Image

Standings.

Elo from recorded pairwise matches only. A voided match is counted and shown but never moves a rating.

No matches yet

The ladder is empty on purpose.

Season One — Image is announced with 6 contestants entered and no match recorded. Ranking models that have never met would be inventing a result, so there is nothing here until the first verdict lands in the journal.

The suite

What we ask models to make.

Chosen for where models differ in work people get paid for, not where they demo well. Every prompt is public — there is no hidden set.

Brand kit fidelity

image

Real content work is constrained work. A model that ignores a supplied palette and typeface is unusable for a brand however pretty its output. Uses a public synthetic brand kit so anyone can re-run this without our customers' assets.The prompt, in fullA launch announcement image for the brand in the supplied kit. Follow the kit palette exactly, use the kit wordmark once at the size the kit specifies, and leave the lower third clear for caption text. Subject: a single ceramic mug on a concrete surface, hard side light.1 output · 1:1 · references: brand-kit

Character consistency

image

Serialised content needs the same face across a week of posts. This is where reference-conditioned models and edit models pull decisively ahead of pure text-to-image, and where a compile step can either help or destroy likeness.The prompt, in fullUsing the supplied character sheet, produce three frames of the same person: (1) waiting at a bus stop in the rain, (2) laughing at a kitchen table, (3) walking away from camera down a corridor. Same person, same age, same hair, three genuinely different situations.3 outputs · 4:5 · references: character-sheet

Product consistency

image

The commerce case. The product must survive relighting and relocation without the model redesigning it. Hallucinated logos, changed proportions, or invented buttons are disqualifying in a way that a slightly worse composition is not.The prompt, in fullUsing the supplied product shot, place the exact same product in three scenes: on a bathroom shelf with morning light, in a gym bag on a bench, held in one hand outdoors against a blurred street. Do not alter the product's shape, label, colour or proportions.3 outputs · 1:1 · references: product-shot

Prompt adherence

image

Counting, spatial relations and negation are the three instructions image models most reliably ignore. This task is deliberately unglamorous: a beautiful image that gets the brief wrong loses to a plainer one that gets it right.The prompt, in fullExactly five green apples and one lemon on a wooden table. The lemon is to the left of every apple. No leaves, no basket, no knife, no text. Overhead view, soft diffused light.1 output · 3:2 · references: none

Legible text

image

The single most commercially decisive image capability and the one that moves most between releases. A model that cannot set six words correctly cannot make an ad, a thumbnail, or a menu.The prompt, in fullA hand-painted sandwich board outside a coffee shop. The board reads exactly: CLOSED MONDAYS OPEN 7-4 OTHERWISE Chalk on black, slightly uneven letterforms, wet pavement, early morning.1 output · 4:5 · references: none

Unbranded craft

image

Strips away every constraint except making a good picture. Separates models whose defaults are tasteful from models that need heavy prompting to stop looking synthetic. This is the task a raw model is most likely to win.The prompt, in fullAn editorial photograph of a woman in her fifties closing up a neighbourhood bakery at dusk. Shot on 35mm, available light only, the shutters half-down behind her. She is tired and pleased. Natural skin texture, no retouching.1 output · 3:2 · references: none

Season Two

Video, not yet running.

Published now so the tasks are fixed in public before anyone knows which model they will favour. The harness refuses to draw a draft into a match.

Camera instruction

videoSeason Two draft

Whether the model can be directed. A generator that produces a lovely shot other than the one you asked for cannot be used in an edit.The prompt, in fullStart on a close-up of a kettle beginning to steam. Push in slowly for two seconds, then whip-pan left to a person looking up from a laptop. No cuts.1 output · 16:9 · 6s · references: none

First-frame fidelity

videoSeason Two draft

The production path that matters most: art-direct a still, then animate it. Measures how much of the approved frame survives contact with the video model.The prompt, in fullAnimate the supplied style frame. The subject breathes and shifts weight; a curtain moves in the background. Nothing else changes. The first frame of the output must match the supplied frame.1 output · 9:16 · 5s · references: style-frame

Motion coherence

videoSeason Two draft

Whether a subject survives its own movement. Limb count, ground contact and object permanence are where video models still fail in ways a still image never reveals.The prompt, in fullA courier carries a stack of four flat boxes across a busy pavement, turns side-on to pass a lamp post, and continues. Single continuous shot, handheld, daylight.1 output · 16:9 · 5s · references: none

Contestants

Genfeed is entered as a contestant, not as the scoreboard.

One route runs through Genfeed's brief compiler; the rest are raw provider routes, including the same model the compiled route wraps. If the compile step makes a task worse, the ladder says so.

Genfeed compile → Seedream 5 ProGenfeedSeedream 5 ProNano Banana 2FLUX.2 ProGPT Image 2Recraft V4 Pro