ResearchRiverflow 2.0

Today we are introducing Riverflow 2.0 from Sourceful, a production-grade image generation and editing model focused on marketing and creative teams. Riverflow 2.0 is built to solve the three pain points that keep frontier diffusion models out of professional workflows: reliability, font control, and detail preservation.

Why Riverflow 2.0 exists

As we interviewed designers and marketers, the same challenges showed up over and over again. Riverflow 2.0 is a direct response to these real, costly workflow blockers.

1) Reliability

Frontier models can be brilliant, but they are often unreliable. In production workflows, that unreliability turns into a real cost: extra candidate generations, extra review time, and a lingering uncertainty that the final asset is truly correct.

To measure this, we took learnings from other AI fields and created Hype-Edit-1, our first public open-source benchmark for reliability. Hype-Edit-1 repeatedly runs the same task and measures not just whether a model can do it once, but whether it can do it again and again.

From this, we defined the Effective Cost of a Successful Edit -- a metric that accounts for:

• The per-image request cost
• The number of retries needed
• The human cost of reviewing candidates

Riverflow 2.0 uses existing open, closed, and our own diffusion models, but improves on them by adding a reasoning model that automatically reviews candidates and makes corrections without human intervention.

2) Font control

Text-on-image has been demoed extensively, but in real marketing assets it breaks down quickly: models don't reliably reproduce exact fonts and they rarely support custom brand fonts.

Riverflow 2.0 introduces font control. The model understands public and custom fonts you provide, and validates that the rendered creative matches your typography -- a tedious and error-prone task for humans. It checks aperture, ligatures, thickness, and descenders while still allowing for creative typography.

Riverflow 2.0 supports up to 2 fonts at a time and up to 300 characters per image.

3) Detail preservation at high resolution

Even at 2K (2048x2048) and 4K (4096x4096), frontier models frequently distort fine text and product details. Fixing those errors by hand in tools like Photoshop is costly and requires expert time.

Image super-resolution can help, but it often fails on fine text or artwork details where the low-resolution input doesn't contain enough information. Reference-Based Super Resolution improves on this by providing high-resolution references to guide the details -- but until now, it hasn't been reliable enough for production use.

Riverflow 2.0 supports reference-based super resolution for fixing product detail issues. Provide high-quality reference artwork, and Riverflow 2.0 identifies and patches problem areas automatically. It supports up to 4 replacement instances per request, whether that's multiple fixes on one product or several products in one scene.

Benchmark results

Riverflow 2.0 was benchmarked independently with Artificial Analysis in their Image Editing and Text To Image Arenas.

What Riverflow 2.0 enables

With reliability, font control, and detail preservation in place, teams can finally build human and agentic creative flows with confidence. Riverflow 2.0 turns AI-generated visuals into assets that are ready for production, not just experimentation.

At launch, we are excited to have Runware, OpenRouter and Replicate as our exclusive API distribution partners.

For more examples and to see how you can use Riverflow 2.0 in the API or in our platform, visit Riverflow 2.0.