PVL Qwen Image Edit Multi (fal.ai)

The basic Qwen edit node takes one image and a sentence. This one takes up to six images and turns the sentence into a whole batch of attempts. PVL Qwen Image Edit Multi (fal.ai) wraps fal.ai's fal-ai/qwen-image-edit-plus endpoint and does what its name promises: multi-image editing context, plus parallel generation - set num_images to 12 and it fires 12 separate Qwen requests at once, each with its own incremented seed, and stacks the results into one IMAGE batch. If you've ever wanted to "show the model the front, side, and back of the character, then generate 8 outfit variants," this is the node that does it in one pass.

Qwen-Image-Edit is the open instruction editor that absorbed the jobs masks and adapters used to do (per the KB it's now the default open editor since late 2025). Everything it does well - garment swaps, relighting, re-posing - it does here with extra context images to pin down who's who.

How it works

image1 through image6 are all sent to every request as combined context. The node splits your prompt on a delimiter (default [++]), so you can pack multiple edit instructions into one prompt field and have them spread across the parallel calls. Each call retries up to max_retries (3) with linear backoff, respects timeout_sec (180), and seeds auto-increment (seed, seed+1, ...) so your batch isn't a wall of identical frames. The outputs land stacked in a single images tensor, and prompt echoes back what was used.

Inputs that matter

• prompt (required) and image1..image6 (optional sockets). Feed as many context images as your edit needs.
• num_images (1-12) = number of parallel generations. This is both your variety dial and your credit-card dial.
• delimiter ([++]) splits one prompt field into multiple per-image prompts.
• aspect_ratio (custom default, or 1:1 / 3:4 / 9:16 / 4:3 / 16:9) with width/height when custom.
• num_inference_steps (50), guidance_scale (4), negative_prompt, seed, enable_safety_checker (off), output_format, acceleration, sync_mode, debug.

Install

Same pack, same steps:

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

Restart, set FAL_KEY (fal.ai/dashboard), done. Like the sibling fal nodes, the pack's full requirements.txt installs regardless - you only actually need requests for this one.

Where people get burned

• Parallel means parallel cost. num_images 12 is 12 paid API calls the moment you hit run. Use 3-4 for first passes.
• Qwen edits still drift off-mask. Multi-image context fixes identity, not the model's habit of re-emitting the whole frame; if only one region should change, pair this with the mask-based PVL Qwen Inpaint node instead.
• The prompt-splitting delimiter is easy to break. If your prompt contains [++] naturally, you'll get unexpected splits - change the delimiter to something you control.