VideoRouter Image Edit

VideoRouter Image Edit does the thing Kontext made normal, except the model lives on someone else's GPU. You feed it an image and a sentence - swap the jacket, drop the background, fix the sign text - and it POSTs to videorouter.sh's /v1/images with your picture attached as a reference, then hands the result back as a normal IMAGE. It's the image-to-image sibling of VideoRouter Image Generate in the same pack: identical auth, identical live model dropdown, one extra input and one extra network step.

When it's the right tool

If you own an edit-capable local model and a GPU, this node is probably not what you want. Qwen-Image-Edit is Apache 2.0 and has become the default instruction editor precisely because its licence lets people build LoRAs and tools around it; Flux Kontext started the category but its anti-circumvention clause turned NSFW derivatives into licence violations and pushed the ecosystem away. Local editing costs you electricity instead of per-call money, and nothing leaves the machine.

The exception is the same one that justifies every API node: the interesting edit models are the closed ones. Kontext Pro and Max are API-only, Nano Banana Pro and GPT Image 1.5 have no weights at all, and ideogram-style text rendering in an existing image is a hosted trick. If you want one of those inside your graph next to your local upscaler, this is the door. Accept that your source image goes to a third party, that the provider's content filter applies and can't be patched out, and that you pay per call.

How it works, and the extra step

There's no polling - one HTTP round trip plus a staging upload. The node takes the incoming IMAGE tensor, encodes frame 0 as PNG bytes, POSTs it to /v1/uploads, gets back a presigned URL good for roughly half an hour, and then sends that URL inside an input_references array on the /v1/images call. The generated entries that come back carry either a url or b64_json, which the node fetches or decodes, converts to RGB and batches.

That two-step shape explains two real behaviours. Big source images are slower, because they get uploaded before anything generates. And only the first frame of an image batch is uploaded - the conversion is hard-coded to index 0, so wire in a batch of five and four are silently ignored. If you need five edits, use n.

The inputs and output

• image - the source. A single IMAGE, first frame only.
• prompt - multiline, the edit instruction. Write it as an instruction ("change the shirt to a red flannel"), not as a fresh scene description; that's how these editors behave.
• model - the same live combo as the Generate node, defaulting to auto, refreshed from the real /v1/images/models catalog. The node's own description is blunt about the catch: it requires an edit-capable model. That dropdown is not filtered to edit models, so picking a text-only model will fail. If you only ever see auto in the list, your key isn't resolving and the catalog fetch is falling back.
• n - 1 to 4. More than one result arrives as a single IMAGE batch.
• aspect_ratio and resolution - optional free-text strings, empty by default and simply not sent when empty. Handy here, because the output size is not inherited from your input image; if a specific shape matters, this is how you ask for it. Nothing validates the format, so use whatever videorouter.sh documents.

The only output is one IMAGE - a batch if n > 1 - which wires into Save Image, an upscaler, or a second edit pass if you're doing a chain of instructions.

Installing it

Same pack, same install. Via Manager, search "VideoRouter", or:

cd ComfyUI/custom_nodes
git clone https://github.com/synvia-corp/comfyui-videorouter
# restart ComfyUI

Then ComfyUI Settings → VideoRouter → Auth → API Key (keys at videorouter.sh/keys). There's no per-node key field on purpose, so a shared workflow can't leak a live key. Dependencies are aiohttp, pillow and numpy, all standard ComfyUI fare - no models to download. Be aware that installing this pack also switches on a Partner Node relay that re-bills a long list of ComfyUI's official nodes through videorouter.sh, plus a price-badge patch on three of them.

Common problems

No VideoRouter API key configured is the keyless error, and as the message says, the settings panel is the only place to fix it. A 402 means the credits ran out. A model that generates fine but refuses to edit is almost always the unfiltered dropdown picking a text-only model, not a bug.

Finally: the pack is new and the corpus has zero threads about it, so you're ahead of the crowd and on your own. It's MIT and small enough to skim before you trust it with a key and a client's photo.