Riverflow 2.5 Pro

Riverflow 2.5 Pro is the most powerful model in Sourceful’s Riverflow 2.5 family. It treats generation as a production workflow: an integrated reasoning model plans multi-step edits, and an internal judge scores candidates before accepting a result.

Use it for top-quality, control-rich generation and editing where output reliability matters more than latency.

Recommended timeout: up to 10 minutes per request (higher thinking_level = better outputs but slower)

What’s new in 2.5

• Adjustable thinking effort via thinking_level (low, medium, high, xhigh)
• Candidate scoring with optional scoring_prompt and scoring_rubric to steer the judge
• Background control (original, transparent, solid with hex color)
• More output formats: webp, png, jpg

For speed-sensitive workflows, use sourceful/riverflow-v2.5-fast instead.

Capabilities

Text-to-image

Generate images from a text instruction. Supports:

• Resolutions: 1K, 2K, 4K
• 11 aspect ratios including auto
• Transparent or solid-color backgrounds
• Optional prompt enhancement
• Multi-iteration agentic refinement, controlled by thinking_level

Image-to-image

Edit or transform up to 10 input images with a text instruction:

• Style transfer, re-rendering
• Background changes, object edits
• Layout and scene changes
• Brand-safe variations

Scoring

Pass a scoring_prompt (free text) and/or scoring_rubric (JSON array of dimensions with weights and optional passing scores) to tell the internal judge what to optimize for. Useful when you have explicit acceptance criteria, e.g. legibility, brand consistency, color accuracy.

Thinking levels

Higher thinking_level means more editing passes and a stricter judge.

• low — quick exploration
• medium — balanced (default)
• high — production runs
• xhigh — batch runs that need high repeatability

Tips

• For image inputs, prefer URLs over base64. Sourceful imposes a 4.5 MB request size limit.
• For runs needing accurate text rendering, write the exact text in instruction verbatim and include casing and placement guidance.
• Lower thinking levels are dramatically faster; reserve xhigh for batch or high-stakes runs.