v0.4.0 — Sentinel Evaluation Engine, New Models, Workflow Pipelines

v0.4.0 is our biggest release yet. The headline feature is Sentinel, our automated model evaluation engine, but there is a lot more packed into this one. Here is the full breakdown.

Sentinel Evaluation Engine

Every output generated through Runflow is now automatically scored across three dimensions: FID (distributional similarity), CLIP (prompt alignment), and human eval calibration. Scores are weighted per niche, so a corporate headshot is evaluated differently than a creative portrait or a product photo.

This powers our benchmark tables and gives customers real-time quality metrics in the dashboard. Read the full technical deep-dive in Building Sentinel: Our Automated Model Evaluation System.

New Models

• Flux.2 Klein — A lightweight variant of Flux.2 optimized for speed. 2x faster than Flux.2 [schnell] with only a 3-point quality tradeoff. Ideal for real-time preview generation and interactive applications.
• SDXL Lightning v2 — Updated 4-step distilled model with improved face coherence. Scores 3 points higher than v1 on our portrait benchmark while maintaining the same latency profile.

Workflow Pipelines

You can now chain multiple operations into a single API call. Define a pipeline that generates an image, scores it, and enhances it, all in one request with a single webhook callback.

const result = await runflow.workflow({
  steps: [
    { action: "generate", model: "flux.2-dev", prompt },
    { action: "score", niche: "corporate-headshot" },
    { action: "enhance", model: "real-esrgan-x4" }
  ]
});

New Benchmark Categories

Sentinel now tracks three additional niche categories:

• E-commerce product shots — Scoring optimized for white-background product photography, edge clarity, and color accuracy
• Creative portraits — Artistic style generation with emphasis on aesthetic quality and prompt creativity adherence
• Virtual try-on — Clothing overlay accuracy, body proportion preservation, and garment texture realism

Bug Fixes

• Fixed an issue where webhook callbacks would occasionally fire before the image was fully uploaded to CDN, resulting in 404s on the image URL
• Resolved a race condition in the async job queue that could cause duplicate processing of the same request under high concurrency
• Fixed EXIF orientation handling for uploaded reference images that were being rotated incorrectly on iOS Safari uploads
• Corrected timeout handling for long-running SDXL inpainting jobs that exceeded the default 30-second window

Update to v0.4.0 by running npm update @runflow/sdk. The Sentinel scoring API is available immediately on all plans. Workflow pipelines are in beta and available on Pro and Enterprise.