Zhenzhen Seedream V5 Lite Edit Fal

The display name calls it "Seedream V5 Lite Edit Fal," and that's a fair summary: it's ByteDance's Seedream V5 Lite edit model routed through the fal queue, taking up to ten reference images and producing edited images that respect them. Where the seedream-v5-pro nodes in this pack are about generating (or editing) at scale, this one is the lighter, faster edit path - think "here's my product shot and ten context images, keep it consistent, change this thing."

Category-wise the pack files it under zhenzhen/Google, which is a leftover from the pack's history and definitely doesn't mean this is a Google model - it's ByteDance. Don't let the menu placement confuse you.

It's a FAL node, so the token lives in the node's own api_key field and traffic runs through the zhenzhen proxy (https://ai.t8star.org/fal in front of queue.fal.run). The endpoint is fal-ai/bytedance/seedream/v5/lite/edit.

The inputs that matter

• prompt - required. For an edit, describe the change; reference images are matched positionally.
• image1–image10 - the reference images. Up to ten, wired in as IMAGE tensors. More isn't always better - consistency degrades once you overload the context.
• num_images - how many output images (1–4).
• max_images - per-generation cap; the tooltip notes total output = num_images × max_images. Mind that multiplication if you're metered per image.
• image_size - presets from auto_2K (default) through auto_4K, plus square and portrait/landscape 4:3/16:9 options.
• enable_safety_checker - on by default.
• image_way - image_url (default) or base64 for how references are shipped to the API. Keep the default unless you're chasing a specific failure.

Outputs: images (tensor list), image_urls, and response. Poll defaults here are a bit longer than the pack's standard - 3s interval with a 1-hour cap by default, which reflects how slow Seedream edits can be.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen

or ComfyUI Manager (search "Comfyui-zhenzhen"), restart. No models to download; paste a zhenzhen token (overseas https://ai.t8star.org) into api_key. FAL routes are prepaid - credits are held up front and the difference refunded after the job.

Common issues

• Edits ignore the references - double-check image_way. If references are being sent as URLs and your images aren't uploaded, the model sees nothing. Also re-verify your image wiring order.
• Cost creep from num_images × max_images - this is metered per output image, and the combination multiplies fast. Keep both at 1 until you need batches.
• Slow or timed-out edits - the default timeout is already an hour, but multi-reference edits at 4K can still push it. The job may finish on fal's side after the node gives up.

For "same character, ten angles, one consistent edit," this is the node - the multi-reference count is its whole reason to exist, and it's cheaper than the pro route for jobs that don't need pro's headroom.