Gemini 2.5 Flash Image sparks speed hunt

Hot take: this is more an inference-routing problem than a model-selection problem. For img2img, warm capacity, queue bypass, and smaller fast variants matter as much as raw model quality.

• –Google’s official docs place Gemini 2.5 Flash Image in the Gemini API and AI Studio, but they do not promise a sub-7-second SLA: https://ai.google.dev/gemini-api/docs/image-generation
• –Google’s changelog now points to Nano Banana 2 / Gemini 3.1 Flash Image Preview as a speed-optimized, high-volume image model, which is the most relevant Google path for latency-sensitive apps: https://ai.google.dev/gemini-api/docs/changelog
• –fal.ai’s realtime docs describe persistent WebSocket inference and say supported real-time models can generate images in under 100ms once connected, which is the right architectural pattern for interactive image editing: https://docs.fal.ai/model-apis/real-time
• –Fireworks’ FLUX docs explicitly say FLUX serverless does not support image-to-image generation, so it is not a fit for this exact use case: https://fireworks.ai/docs/faq-new/models-inference/does-flux-support-image-to-image-generation