FAL Flux 2 Turbo Edit

Flux 2 Turbo is the distilled, few-step version of Flux 2 - the "draft fast, refine later" tier. FAL Flux 2 Turbo Edit is this pack's wrapper around fal-ai/flux-2/turbo/edit, and it exists for one job: rapid iteration. You're trying out edit directions, you don't care about perfect fidelity yet, and you want an answer in seconds instead of a 9B Klein run at full quality.

The README is upfront about the trade: "a faster, streamlined version of Flux 2 editing optimized for quick iterations. Fewer parameters for simpler workflows when speed matters more than fine-tuning." Turbo is the distilled model - think of it as the edit-model equivalent of Schnell or a turbo LoRA: 8-ish steps instead of 28+, way faster per image, and the KB's distillation coverage explains the catch - distilled models trade the fine control that full multi-step guidance buys you. You get direction, not precision. That's the deal.

The inputs reflect the "fewer parameters" promise. Required: api_key, prompt (default "Make this image more vibrant and detailed"), image1, and width/height (512–2048, step 8). Optional: image2/image3 for multi-reference, guidance_scale (default 2.5 - low, because turbos don't want much guidance), seed, output_format (jpeg/png/webp), and enable_safety_checker. Notice what's missing next to the Klein and LoRA edit nodes: no num_inference_steps, no negative prompt, no LoRA inputs, no acceleration tier. Deliberately stripped down, which is the whole point.

Outputs are the standard FAL set: image, image_path, generation_info, generation_time. Saved to output/API/FAL/flux-2-turbo/ with a metadata JSON.

The honest framing: you reach for this node when iteration speed outranks quality. The classic workflow is turbo first - throw three prompts at it, see which direction survives - then take the winner to Klein 9B or Qwen-Image-Edit for the real pass. The corpus has real signal that turbo-tier Flux 2 got taken seriously fast; fal.ai themselves open-sourced the FLUX.2-dev-Turbo weights, and the community's first question was about how much quality the distillation cost. The answer, broadly: it's a draftsman, not a finisher.

A few practical notes. It's a paid API - every queue run spends FAL credits, so the "fast" framing is about wall-clock, not wallet. And because it's an API call there's an inherent round-trip: upload image, queue, poll, download. Even "fast" means several seconds plus network. If you're doing hundreds of edits, consider whether the latency-per-image fits your flow, or whether you want the local models this pack deliberately lets you skip.

Install is the same pack story as the rest of these: ComfyUI Manager search "Node Fun", or clone https://github.com/nofunstudio/Node_Fun_ComfyUI into custom_nodes, pip install -r requirements.txt (the fal-client dependency), restart. Paste your FAL key into the node or set FAL_KEY, and you're running. The default prompt is a decent starting point - "more vibrant and detailed" is a safe, reversible first edit to sanity-check that your key, upload, and output path all work before you spend on something ambitious.