Flux Kontext Max (Channel Edition)

This is the node people actually look for when they hear "Kontext." Feed it one or more images and an instruction - "change the coat to leather," "swap the background to a rainy street," "remove the watermark" - and it returns the edited image. Kontext does this by consuming your reference images in-context alongside the prompt, rather than through adapter embeddings like IP-Adapter or PuLID. That's the architectural trick behind its famous character consistency: identity is preserved structurally, so a subject holds across successive edits with minimal drift.

This specific channel is Kontext Max I2I, hosted by SiliconFlow. Two things make it worth your attention. First, it uses a COMFY_AUTOGROW_V3 image input - an auto-growing multi-image slot, so the node can accept several reference images in one call rather than a single fixed socket. That's genuinely useful for multi-reference character work. Second, it's the hosted tier: BFL's Kontext Max quality without the 12B download or the license reading. The trade is cloud credits and latency.

Inputs and outputs

• images - required, auto-growing. One or more reference images. This is what gets edited and what the model uses for context, so give it the actual picture you want changed.
• prompt - required. The instruction. Full sentences: "turn the sky to storm clouds" beats "sky storm clouds." This is an instruction model, not a tag model.
• aspect_ratio - dropdown, default 1:1.

Outputs: image, request_id, urls (JSON string of output file URLs). Standard triple.

How it works

Same pack machinery as every API node: serialize your images and prompt, submit to flux-kontext-max-base/image-to-image, poll while it renders, download the result. The key comes from BizyAirPlus settings or BIZYAIR_API_KEY. Because the images ride up to the cloud and come back down, an edit run takes longer than local inference on a fast GPU - that's the honest cost of hosted editing.

Where people get burned

• Short, vague prompts. Kontext is a wizard at following precise instructions and mediocre at reading minds. "make it better" gets you nothing; "replace the background with a desert at dusk" gets you a desert.
• Iterative edits accumulate drift. Each edit is a new generation. The KB notes anatomy degrades at full-body scale, and successive edits compound it - if a character needs to hold identity across a chain, do as few edits as possible and reuse the earliest output as the next input.
• Red placeholder output on failure is the pack's skip_error mode - check the console and urls for the real story.

Install

With the pack:

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/siliconflow/BizyAirPlus.git
cd BizyAirPlus
python -m pip install -r requirements.txt

Restart ComfyUI, confirm BizyAirPlus is ON and keyed. Then do what everyone does first: try the watermark. It's the demo that made this model famous, and it's exactly the job this hosted channel was built for.