State of Stream Overlays 2026: What 235 AI Overlay Packs Reveal

What this is

This is an original-data report from AlertForge production usage, measured on July 13, 2026. It extends our [April–June stream-alert statistics](/blog/stream-alert-statistics-2026) with a new dataset: complete AI-generated overlay packs (scenes + panels + alerts sharing one look), which launched as a product surface on May 16, 2026.

Everything below is aggregate and anonymised — counts, rates, and dates only. Methodology and limitations are at the end; the headline numbers first.

Headline numbers

| Metric | Value (Jul 13, 2026) | |---|---| | AI overlay packs created | 235 (since May 16 launch — 59 days) | | Distinct creators who built a pack | 156 | | Alert templates created (all-time) | 563 | | Distinct alert-template creators | 318 | | Settled video renders analysed | 310 (227 succeeded / 83 failed) | | Overlay pipeline jobs completed | 316 of 320 (98.8%) | | Registered users | 535 |

Overlay packs: adoption since launch

Complete-pack generation is new — the feature shipped mid-May 2026 — so this is the first public look at how it's being used:

| Month | Packs created | |---|---| | May 2026 (from the 16th) | 18 | | June 2026 | 127 | | July 2026 (through the 13th) | 90 |

July is tracking toward roughly double June's volume if the daily pace holds (90 packs in 13 days ≈ 215/month pace). The average pack creator makes ~1.5 packs (235 packs / 156 creators), which matches what we see qualitatively: most streamers generate one pack for their current brand, and a minority iterate across multiple themes before settling.

What's inside the pipeline: images, design passes, video

Each overlay pack is assembled from individual pipeline jobs. Of the 320 jobs recorded:

| Job type | Count | Share | |---|---|---| | Image generation (scene art, cutouts) | 183 | 57% | | Design passes (style sheets, layout) | 112 | 35% | | Video generation (animated scenes) | 25 | 8% |

The pipeline completed 316 of 320 jobs (98.8%). That number is higher than raw text-to-video success rates (below) because pack assembly is mostly image work with a compositing step — video generation is the smallest and hardest slice.

Render success rates by model

For standalone video renders (alert clips and animated scenes), we count only settled renders — those that reached a terminal succeeded/failed state. 310 renders settled in the analysis window:

| Model | Succeeded | Failed | Success rate | |---|---|---|---| | Wan-Alpha (fal.ai) | 95 | 16 | 85.6% | | Veo 3.1 Lite | 90 | 16 | 84.9% | | Veo 3.1 Fast | 42 | 8 | 84.0% |

Two observations worth making:

1. The three models have converged. In our [June report](/blog/stream-alert-statistics-2026), Veo 3.1 Lite led at 90.3% with Veo Fast at 81.0% — a 9-point spread. That spread is now under 2 points. As prompt templates and retry handling improved on our side, model choice stopped being the dominant success factor. 2. ~15% of renders still fail, and the leading causes are unchanged: content-filter refusals on prompts that mention weapons/brands, and motion requests too complex for a 5-second clip. This is why credit-based iteration (rather than pay-per-final-video) remains the honest pricing model for AI generation — some fraction of attempts will always need a reroll.

Alerts vs packs: how the mix is shifting

Alert templates grew from 518 (June 10) to 563 (+45 in ~5 weeks), while overlay packs added 217 in the same window (18 existed on June 10). In other words: pack creation is now outpacing standalone alert creation by roughly 4–5× among new projects. Streamers increasingly arrive wanting the whole look — scenes, panels, cam frame, alerts — rather than a single alert clip, which matches the broader shift we wrote about in [our overlay-generator comparison](/blog/best-ai-stream-overlay-generators-2026).

Methodology

Source: AlertForge production database, read-only aggregate queries executed July 13, 2026. No individual user data, prompts, or identifying information was extracted.

"Settled renders" = renders with a terminal succeeded or failed status. Queued/abandoned rows (e.g. a user closing the tab before render completion) are excluded rather than counted as failures.

Model table includes only models with ≥10 settled renders in the window. Composite pipeline rows whose outcomes are tracked in the job system (not the render queue) are excluded from the render table and reported separately as pipeline jobs.

Overlay pack counts = distinct pack projects created; a project edited multiple times counts once.

Limitations — read these before citing

Single-platform data. This is AlertForge usage, not the streaming industry. Streamers who choose an AI generator are early adopters by definition.

Small absolute numbers. 235 packs and 310 settled renders are honest but modest samples; percentages carry meaningful variance (±5pp on the model table would not be surprising).

Short window for packs. 59 days of a new feature reflects launch dynamics — growth curves this early say more about discovery than steady-state demand.

Success rate ≠ satisfaction. A render can succeed technically and still not match what the streamer imagined; reroll behaviour (not measured here) would be the better proxy.

Use this data

Cite freely with attribution to AlertForge (alertforge.ai) and a link to this page. We'll re-run these queries quarterly; the next update lands October 2026. Questions about methodology: [email protected].