flux-2-klein

Install the RunComfy CLI with npm i -g @runcomfy/cli, authenticate via runcomfy login, then run runcomfy run blackforestlabs/flux-2-klein/9b/text-to-image with a JSON input containing your prompt. The CLI polls until completion and downloads the output to your specified directory.

The 4B variant is latency-first, delivering sub-second 4-step inference for live iteration and concepting. The 9B variant is fidelity-first, using around 25 steps for polished final output with stronger texture and microdetail.

Route to Flux 2 Pro for maximum prompt adherence and extreme detail, Seedream 5 for 2K–4K hero images, and GPT Image 2 for embedded text, logos, or multilingual signage. Flux 2 Klein is best when iteration speed matters more than ceiling resolution.

Yes, both variants accept up to four simultaneous reference images on the same endpoint for style transfer and guided composition. Keep reference aesthetics aligned, since mixing watercolor, photoreal, and 3D styles in one call degrades results.

The resolution ceiling is approximately 2K, aspect ratio is capped at 16:9, and prompts degrade beyond roughly 512 tokens. Requests exceeding these limits are cropped or produce lower-quality output rather than failing cleanly.

Exit code 77 means you are not signed in or your token was rejected. Run runcomfy login to complete the browser device-code flow, or set the RUNCOMFY_TOKEN environment variable in CI and container environments.