PVL Kontext Dev LoRA (fal.ai)

The plain Kontext Dev node edits your image by instruction. This one does that and layers a LoRA into the mix - so you can edit a photo into your trained character, your studio style, or your brand's look in a single instruction-based call. It wraps fal's flux-kontext-lora endpoint: one image, one edit instruction, one fal-hosted LoRA path, and the result comes back as a regular IMAGE tensor. For anyone who's trained a Kontext-flavored LoRA and wants to apply it to new inputs without local VRAM, this is the node.

It's part of the ComfyUI Assistant Node pack from pvlprk, and it sits one step up from the pack's plain Kontext Dev node in both capability and setup.

How it works

Your image is base64-encoded and sent with prompt to the Kontext LoRA endpoint, alongside an optional LoRA applied with lora_path + lora_strength. The prompt is read as an edit instruction on the input image - "make her wear this jacket," "convert this to watercolor" - and the LoRA steers the edit's style or identity. lora_strength (default 1.0, range −2 to 2) controls how hard the LoRA pulls.

Everything else mirrors the Kontext Dev node: num_inference_steps (default 30), guidance_scale (2.5), seed, num_images, output_format, sync_mode, acceleration, and the same resolution_mode menu (match_input by default, or an aspect ratio to reframe the edit). enable_safety_checker is on by default here.

The inputs that matter

• image - the input to edit.
• prompt - the edit instruction.
• lora_path - the fal-side LoRA identifier. As with the pack's Flux Dev LoRA node, this is not a local file path - the LoRA must exist on fal's side (uploaded via fal's storage).
• lora_strength - how much the LoRA influences the edit.
• resolution_mode - keep input dimensions or reframe to an aspect ratio.

Output: one IMAGE.

Installing it

Part of pvlprk/comfyui-pvl-api-nodes. ComfyUI Manager (search "ComfyUI Assistant Node") or:

cd ComfyUI/custom_nodes
git clone https://github.com/pvlprk/comfyui-pvl-api-nodes

Set FAL_KEY (or FAL_API_KEY / FAL_CLIENT_KEY), restart, add credits.

Where people get burned

The local-path trap is the big one - a LoRA you trained and keep in models/loras won't resolve here; it has to be uploaded to fal and referenced by its fal-side identifier. Beyond that, LoRA + Kontext editing is two models pulling on the same image, and they can fight: if the edit drifts from your input, drop lora_strength before you blame the instruction prompt. And the Kontext family's known full-body distortion still applies, so keep edits mostly local.