milvey

Run Omni, Kling, Wan or your own ComfyUI graph from the same credits.

A detailed wooden toy ship with carved masts and sails glides over a plush blue carpet that mimics the waves of the sea.

CogVideoX 2BRun~5 cr

Published sample · CogVideoX 2B · Apache-2.0 · published prompt · open models run as a ComfyUI workflow on a GPU pod · price is a sample

The price is on the Generate button before you press it. Credits are held while the run works and charged when the file arrives. If the run can’t be delivered, you pay nothing.

No card needed. Credits don't expire.

free credits when you sign up
150
free credits when you sign upsample
video and image models
27
video and image models
GPU types, 24 to 141 GB
6
GPU types, 24 to 141 GB
cheapest GPU, metered per second
4 cr/min
cheapest GPU, metered per secondsample

Three things you can do with the same credits.

Create

Pick a model, or let Auto pick one under your cap.

Each model lists its time and price before you run it. Auto chooses the strongest model that fits your cap and tells you which one it used.

See the models
Workflows

Drop in a ComfyUI JSON. Get a form anyone can use.

Missing custom nodes and model files are listed by name and installed for you. Choose which inputs go on the form. Everything else in the graph stays out of sight.

How import works
Deploy

Call your workflow over HTTP.

A saved workflow becomes an endpoint on the GPU you choose. It sleeps when nobody calls it, so an idle day costs nothing, and each request is logged with its cost.

See endpoints

What happens when a provider fails

This is the ledger for one run. Kling returned an error on every route, so the run moved to the next model under your cap. You pay for the file you get, and the difference goes back to your balance. Open models on Milvey GPUs queue instead of moving to a paid API.

r_4823 · “Neon market in the rain” · 5s · 720pdelivered
14:02:11Hold for Kling V3 (under your 120 cr cap)−110
14:02:40Kling returned 503 on every route. Looking for the next model under your cap.0
14:02:41Rerouted to Wan 2.2 (next model under your cap)
14:03:58File delivered by Wan 2.2. Charged its price.−26
14:03:58Unused part of the hold returned+84
Net cost 26 cr. If the run could not be delivered: 0 cr. sample run

Sample outputs from open models

Sources and licences

Candle flame with smoke and bokeh (description; prompt not published)

Mochi 1· Apache-2.0

Woman in a red dress among red lanterns (description; prompt not published)

Mochi 1· Apache-2.0

Hand picking lemons from a bowl, studio product shot (description; prompt not published)

Mochi 1· Apache-2.0

A detailed wooden toy ship with intricately carved masts and sails is seen gliding smoothly over a plush, blue carpet that mimics the waves of the sea. The ship's hull is painted a rich brown, with ti

CogVideoX 2B· Apache-2.0

The camera follows behind a white vintage SUV with a black roof rack as it speeds up a steep dirt road surrounded by pine trees on a steep mountain slope, dust kicks up from it’s tires, the sunlight s

CogVideoX 2B· Apache-2.0

Cowboy riding in the desert (Commons description; generated with Runway Gen-3 Alpha)

Runway Gen-3· Public domain (Wikimedia Commons)

These come from the model makers’ own published samples and Wikimedia Commons, credited on the sources page. Outputs made on Milvey will replace them as the beta runs.

The inputs on your form are the fields in the API

Pick three widgets in your graph. They show up on the form your client fills in, and as the JSON body below. The rest of the graph stays private.

Brand film · v7

product_image

bottle.png

prompt

slow orbit, soft studio light

duration

5 s

Run · 12 crsample

request
curl https://api.milvey.example/v1/endpoints/brand-film/runs \
  -H "Authorization: Bearer $MILVEY_KEY" \
  -d '{
    "product_image": "https://…/bottle.png",
    "prompt": "slow orbit, soft studio light",
    "duration": 5
  }'
response · 200
{
  "id": "run_7f3a",
  "status": "completed",
  "output": { "video_url": "https://…/run_7f3a.mp4" },
  "credits": 12,
  "gpu": "RTX PRO 6000",
  "seconds": 89.4
}

What it costs

ModelGood atTypical timeSample price
Veo 3.1Native audio, Camera language~1m 50s500 crper 5 s clip at 720p
Kling V3Human motion, Start/end frames~1m 10s110 crper 5 s clip at 720p
Wan 2.2Open weights, LoRA support~2m 30s26 crper 5 s clip at 720p
Seedance 1.0 ProMulti-shot, Speed~55s65 crper 5 s clip at 720p
LTX-2Speed, Open weights~25s6 crper 5 s clip at 720p
HunyuanVideoPrompt adherence, Open weights~2m 10s23 crper 5 s clip at 720p
Gemini Omni FlashNative audio and dialogue by default, Multi-turn video editing with character consistencyUnknown130 crper 5 s at 720p
Kling 4.03-30 s clips, Native stereo audioUnknown200 crper 5 s at 720p
Runway Gen-4.5Camera and motion control, Aleph 2 video-to-video editingUnknown150 crper 5 s
Luma Ray 3.2HDR / EXR output, Video reframe and editUnknown75 crper 5 s at 720p
MiniMax Hailuo H3Very cheap per second, Native audioUnknown65 crper 5 s
Seedance 2.54-30 s multi-shot clips, Native audio~1 min (unverified)290 crper 5 s at 720p
LTX-2.5Native synchronized audio + video, Fast 8-step distilled variantFast (8-step distilled)11 crper 5 s at 1080p
Kandinsky 5Fast text-to-video, Low cost per clipUnknown (Pro distilled to 10 steps)20 crper 5 s
Mochi 1Strong motion fidelity, Clean Apache-2.0Unknown170 crper 5 s clip at 720p

Sample prices. You see the exact price on the Generate button before you run anything.

Saving and versioning workflows is free. You pay when something runs.

2026

I’m building Milvey on my own.

The first job is getting real ComfyUI workflows to run without a day of fixing custom nodes. If yours doesn’t import, send it to me. Each one that fails shows me what the importer still gets wrong.

Below is what’s done and what’s next. The changelog has the details.

Faiz

Status

In demo

  • Workflow importWeek 01

    Drop a ComfyUI JSON. Custom nodes and models are detected and listed.

  • Node and model resolverWeek 02

    Node packs pinned to a commit. Weights matched by hash and cached.

Building now

  • Form from a graphWeek 03

    Click a widget to put it on the form.

Next

  • Templates and versionsWeek 05

    Save, fork, roll back, and share a pinned version.

  • Endpoints and client appsLater

    Call a saved workflow over HTTP, or hand a client a form.

Questions people ask first

Will my ComfyUI workflow run as it is?
Most do. The importer reads the JSON, installs the custom nodes and fetches the models. If a node can’t be resolved, it’s named before you spend anything.
What do I pay when a run fails?
Nothing. The hold goes back to your balance. If a closed model fails on every route, Auto moves to the next model under your cap and you pay the price of the one that delivered. Open models run on Milvey GPUs, so they wait in a queue until capacity returns; if the run can’t be delivered, the hold is released.
Which files can I import?
UI or API workflow JSON, a PNG with the workflow embedded, or a link to a shared workflow file.
Who can see my workflows?
Only you, until you publish one. You can share a pinned version without showing the graph.
Can I use the outputs commercially?
It depends on the model. Each model page shows its licence, and non-commercial models are flagged before you publish a template that uses them.
Do credits expire?
No. 150 come free at signup (sample number while pricing is set) and packs never expire.

Send me a workflow that won’t run.

Join the list, then reply with a ComfyUI JSON. I’ll get it running on Milvey and send you the cost per run.

One email when signups open · Unsubscribe any time