Remix.Camera data · September 2026

What people generate
and keep.

Name: Remix.Camera data: What people generate and keep, September 2026
Creator: Remix.Camera
Published: 2026-09-07
License: https://aiphoto.ai/research#research-terms-title

Data by Remix.Camera. Analysis and publication by AI Photo.

Real usage reveals demand. Downloads show which outputs people choose to keep. Arena votes show preferences between specific images. Explore the three signals separately.

Usage: August 5 to September 3, 2026, Pacific. Arena: cumulative snapshot on September 7. Source: Remix.Camera usage data and the shared Remix.Camera / AI Photo Arena.

66,533completed image & video outputs
from 1,455 accounts

10,475pack remixes, included in outputs
across 3,100 packs

10,223outputs downloaded at least once
15.4% of completed outputs

3,100Arena ballots, not unique voters
2,690 mapped image-quality ballots

Three findings to use

01 / DEMAND

The biggest audience is not the busiest audience.

Baddie-tagged packs attracted 446 accounts and 1,207 remixes. NSFW-tagged packs attracted 153 accounts and 5,412 remixes. Reach and intensity point to different content opportunities. Accounts and tags overlap.

02 / DOWNLOADS

A model’s overall rate can hide its workflow fit.

Seedream V5 Lite had a 42.4% overall output download rate, but 12.0% for pack remixes (49/408). Its overall mean account rate was 10.9%. Check the workflow and user concentration before treating an aggregate as quality.

03 / PREFERENCE

Keep the test set attached to the vote result.

V5 Pro received 79 of 130 decisive votes against Lite in the reviewed exact-input set (60.8%). In the adult comparison set it received 157 of 350 (44.9%). These are different prompts and samples, not proof that content type caused the change.

01 / What people create

Compare audience size with repeat use.

Use account count to find styles with broader observed reach. Use remix volume to find repeated activity. Of 10,475 pack remixes, 10,467 link to people-only pack metadata. This dataset cannot support product or pet model recommendations.

Compare styles by

Pack tags · 10,475 completed remixes across 904 accounts
Content tag	Accounts	Remixes	Downloaded	Download rate	Packs used
Sexy	772	3,603	643	17.8%	1,364
Baddie Aesthetic	446	1,207	164	13.6%	441
Luxury lifestyle	439	1,563	308	19.7%	745
Coquette Aesthetic	408	1,215	228	18.8%	401
Mermaidcore Aesthetic	320	884	176	19.9%	372
Normcore Aesthetic	202	451	67	14.9%	192
Ethereal Aesthetic	187	469	106	22.6%	265
NSFW	153	5,412	1,616	29.9%	790
2000S	124	326	80	24.5%	170
Creative	88	235	34	14.5%	150
Retro90S	82	173	36	20.8%	100
Fashion Week	74	265	42	15.8%	198
Old Money	73	180	22	12.2%	121
Professional	60	154	24	15.6%	117
Cowgirl Aesthetic	57	205	36	17.6%	152
Gothcore Aesthetic	54	150	35	23.3%	97
Hair Salon	51	119	25	21.0%	83
Barbiecore Aesthetic	51	117	19	16.2%	73
Balletcore Aesthetic	51	104	33	31.7%	60
Luxurymen	39	126	18	14.3%	88

Tags overlap: do not sum rows. These are current pack labels, not classifications of private prompts. Availability and placement affect volume. Sparse tags are omitted.

02 / What people download

Compare models within a workflow.

In pack remixes, V5 Pro recorded 1,484 downloads from 5,007 outputs (29.6%), Seedream 4.5 recorded 442 from 2,010 (22.0%), and Lite recorded 49 from 408 (12.0%). These observed rates include different users, packs, settings and routing decisions. They are not a controlled model ranking.

MediaWorkflowSort by

Image outputs · Pack remix · August 5 to September 3, 2026
Model	Accounts	Outputs	Downloaded	Output download rate	Mean account rate	Largest account share
Seedream 4.5	455	2,010	442	22.0%	15.0%	4.2%
Nano Banana	376	2,110	321	15.2%	8.3%	26.3%
Nano Banana 2	261	540	95	17.6%	10.8%	8.5%
Seedream V5 Pro	256	5,007	1,484	29.6%	14.7%	15.6%
Seedream V5 Lite	246	408	49	12.0%	10.0%	4.2%
FLUX LoRA	63	274	82	29.9%	18.4%	18.6%

Rows need at least 100 outputs and 20 accounts. Accounts can use several models. The mean account rate gives each account equal weight; the output rate gives each output equal weight. Neither controls for task difficulty or who saw each model.

Why show user concentration? One account produced 55.6% of Seedream 4.5 outputs in this window. Its overall output download rate was 7.7%, while the equally weighted mean account rate was 16.1%. A high-volume account can change the story.

03 / Which image people prefer

Preference depends on the tested images.

The shared Arena snapshot contains 3,100 ballots, including 2,690 mapped image-quality ballots. The other 410 cover workflow, intent or other aggregate-only records and are excluded below. Each row compares the same two model labels within one test set.

Rows with at least 100 decisive votes across 10 prompts. Percentages refer to model A.
Test set / Model A vs B	A / B votes	A vote share	Prompts	Equal-prompt A share	Prompt sensitivity range
Explicit adult comparison Seedream 4.5 / Seedream V5 Lite	131 / 250	34.4%	20	34.3%	26.9% to 42.4%
Explicit adult comparison Seedream V5 Lite / Seedream V5 Pro	193 / 157	55.1%	19	54.3%	45.2% to 62.4%
Explicit adult comparison Seedream 4.5 / Seedream V5 Pro	119 / 211	36.1%	19	35.8%	27.6% to 44.1%
Seedream portrait archive Seedream 4.5 / Seedream V5 Pro	46 / 90	33.8%	29	39.2%	28.4% to 50.6%
Seedream portrait archive Seedream V5 Lite / Seedream V5 Pro	78 / 57	57.8%	29	55.9%	46.2% to 66.4%
Exact-input reviewed comparison Seedream V5 Lite / Seedream V5 Pro	51 / 79	39.2%	39	36.7%	26.8% to 46.2%
Seedream portrait archive Seedream 4.5 / Seedream V5 Lite	38 / 84	31.1%	28	29.5%	20.8% to 38.6%

Ties and “both bad” remain in the downloadable data but are excluded from decisive-vote shares. The range resamples prompts, not people, and cannot correct repeat voting across matchups. It is a sensitivity check, not a population winner declaration. Test-set names describe their source protocols; historical sets do not all have identical controls.

Inspect the current main Arena results →

Use the signals together without merging them.

Content tags identify where Remix users spend effort. Download rates describe observed output retention within a model and workflow. Arena votes describe comparative image preference. The datasets do not link the same customers, prompts or outputs, so they cannot establish that a high-demand style has a particular model-quality gap.

For a creator, start with the style audience and inspect relevant comparison images. For a model team, use the workflow rates to choose the next controlled evaluation. For researchers, use the CSV files to test alternative weighting and compare the samples.

Download, inspect and cite.

Model & workflow CSV Content-demand CSV Arena-pair CSV Per-matchup vote CSV Full aggregate JSON Download chart

Remix.Camera data. “What people generate and keep.” Analysis and publication by AI Photo. September 7, 2026. Usage window: August 5 to September 3, 2026 (Pacific); Arena snapshot: September 7, 2026. https://aiphoto.ai/research/ai-generation-usage

Credit Remix.Camera as the data source and link to this AI Photo report when reusing the data or charts. Preserve the relevant limitations. No customer identifiers, prompts or media are included. AI Photo and Remix.Camera are operated by the same team.

Methods and coverage

Archived outputs matter. The active-only download rate was 19.0%. Including 31,288 archived outputs brings the full-cohort rate to 15.4%. Archives include deleted or expired records; deletion does not itself imply dissatisfaction.

Population, definitions and exclusions

population: Completed outputs from confirmed non-admin Remix.Camera accounts, including archived outputs. Created August 5 to September 3 Pacific and completed before September 4; downloads observed at snapshot.
download: At least one recorded download: positive downloadCount or lastDownloadedAt. Repeated downloads count once. Download is a retention signal, not a visual-quality judgment.
maturity: Newest three full calendar days excluded. Follow-up length differs by creation day.
accountMean: Compute each account download fraction within the selected row, then average those fractions with equal account weight.
model: modelVariantId first, then explicit seedreamV5Quality, then resolvedModelType, then modelType. Stored routes are used; WAN route is not promoted to a different marketed version.
demand: Observed activity on one platform. Defaults, availability, pricing, routing, visibility and repeat users influence volumes. No exposure-normalized demand or causal model effect is estimated.
content: Current public pack tags joined by remixedFrom. Tags overlap and can change. Only pack-linked outputs have this classification; no private prompt text was read.
threshold: Usage rows require at least 100 outputs and 20 accounts. All nine qualifying model totals and their qualifying workflow rows are exported; sparse rows are omitted.
arena: Main and adult lanes deduplicated by battle ID, read from named aggregate parents. The 3,100 ballot total includes 410 non-image-quality or unmapped ballots. Pair analysis uses only mapped image-quality ballots. Votes are not unique people.
uncertainty: Arena equal-prompt share weights each prompt equally. The 95% prompt-resampling range uses 2,000 bootstrap samples with seed 20260907. It describes sensitivity to tested prompts, not a model-wide confidence guarantee; repeat voters cannot be controlled from aggregate data.
pairThreshold: Highlighted Arena rows require at least 100 decisive votes across 10 prompts. No population winner is claimed.
join: Arena preferences and production usage are separate cohorts. A model name does not connect a vote to a production generation, and taxonomy labels are not shared.
privacy: Public files contain aggregate data only. No account identifiers, private prompts or customer media.
relationship: AI Photo and Remix.Camera are operated by the same team.
attribution: Credit Remix.Camera as the data source. AI Photo provides the analysis and publication. Arena ballots come from the shared Remix.Camera / AI Photo Arena.

Read completed at 2026-09-07T23:57:31.311Z. Exclusions: 1,973 internal/admin records, 3,552 uploads, 3,104 non-completed records, and 4 outputs completed after the cutoff. Source record hashes and read counts are in the JSON.

What this release does not measure

Market share, paid demand, unique Arena voters, exposure-adjusted preferences, satisfaction without a download, or a causal quality ranking. All-time votes and a 30-day usage cohort have different time coverage. Outputs from accounts no longer present in authentication cannot be classified as confirmed non-admin; none were encountered in this window. We did not read private prompts or voter subcollections.