WaveSpeedAI Flux Kontext Pro

FLUX.1 Kontext Pro is the day-one API tier of BFL's instruction editor - the one people actually used for iterative image editing before the Max tier came along to be flashier. It's API-only, so "local" was never an option. This node from razvanmatei-sf/razv-wavespeed wraps the Pro endpoint in a ComfyUI node: one image in, one sentence of instruction, one edited IMAGE tensor out. If Kontext Max is the experimental sports car, Pro is the sedan that does the same job all day without drama.

What you're actually getting is the famous Kontext behavior: your reference image is consumed in-context with the prompt (not through an adapter embedding), which is the mechanism behind its strong character consistency across successive edits. That's why it became the backbone of turnaround-sheet and character-variation workflows. Restage a character, swap the outfit, drop in a new background - the identity holds in a way that IP-Adapter-style edits struggle with.

Same caveats as the family, though: it's heavily censored, BFL's Kontext-era license is aggressively anti-NSFW-fine-tune, and on a hosted tier you get the guardrails whether you want them or not. This is the wrong tool if that's your workflow.

How it works

Standard pack pattern: a WaveSpeedAI Client node supplies the connection, you give it a prompt and an image_url, and the node calls the Kontext Pro endpoint on WaveSpeed AI. Sync mode defaults to on, so the node blocks until the edit completes and returns an IMAGE tensor for preview or save. The parameter list is even leaner than Max's: just prompt, image_url, guidance_scale, and enable_sync_mode.

• guidance_scale (1–20, default 3.5) - prompt adherence, 1.0 loose to 20.0 strict. Kontext Pro is guidance-distilled and runs low; the default is a good starting point.
• The node ships with "A woman is brewing tea" as its default prompt - a demo instruction, not a directive. Replace it.

The inputs that matter

• prompt - the edit instruction. One clear sentence ("change the shirt to red," "move her to a beach at sunset") beats a paragraph.
• image_url - the image to transform. Hosted URL; the "Upload Image node" the tooltips mention isn't in the current build.
• guidance_scale - the only real dial, and you'll rarely need to move it far.

Install

Shared pack:

cd ComfyUI/custom_nodes
git clone https://github.com/razvanmatei-sf/razv-wavespeed
# restart ComfyUI

Or ComfyUI Manager → search "razv-wavespeed" → Install. Add a WaveSpeed AI key via the WaveSpeedAI Client node, config.ini (copy from config.ini.tmp, [API] section), or a WAVESPEED_API_KEY env var.

Common issues

• Paid per edit, and iterative edits multiply. Kontext's whole value is chaining edits, and chaining edits is chaining bills. Know your cost before you start a 30-edit restyle run.
• Full-body anatomy degrades - the family's documented weakness is heads and hair artifact once a full body fits the frame. Crop tighter or use a different tool for small-in-frame subjects.
• URL input, not tensor. Paste publicly reachable URLs; the pack registers no upload node despite the tooltips.
• Censorship is real. If your edits keep getting blocked, that's the filter doing its job, not a broken node.