DEV Community: Reecho1977

Is Your AI Image API Upscaling? Check the C2PA Credentials

Reecho1977 — Fri, 30 Oct 2026 14:32:42 +0000

Originally published on the Yir blog. I work on Yir, an image & video API router, so the last section describes our own product.

Anyone buying an image API through a reseller eventually asks: is this 4K image a native 4K render, or a 1K image that was upscaled? Is a result labeled GPT Image 2 API actually something ChatGPT made in the app?

Your eyes won't settle it. AI-upscaled images often look cleaner than the originals: noise smoothed away, edges very crisp. What does settle it is the C2PA signature inside the file.

What a C2PA signature is

Think of it as a certificate of origin for an AI image. When OpenAI, Google or ByteDance generate an image or video, they write a record into the file: who made it, which model, when, and whether it is AI-generated. The record is signed with the vendor's certificate, so nobody else can edit or forge it.

The signature is also bound to the file's content. If the file is upscaled, re-compressed or edited, the content no longer matches. So if the signature is there but the content doesn't match, someone touched the file after the vendor produced it.

C2PA is an open standard started by Adobe, Microsoft, Intel, the BBC and others, also known as Content Credentials. It is not EXIF: anyone can edit EXIF, while C2PA detects changes to either the content or the record.

Which AI vendors sign their output

Vendor	Format	What you can learn
OpenAI GPT Image	PNG	API or ChatGPT app, time of generation; some files name the exact model
Google (Nano Banana, Veo and more)	Image / MP4	Generated by Google and unchanged; does not distinguish API from the Gemini app
ByteDance Seedance	MP4	Volcengine Ark or BytePlus, model name, time of generation
ByteDance Seedream	Image	Volcengine Ark or BytePlus, model name
Black Forest Labs FLUX	Image	Generated by BFL and unchanged; does not say API or the exact model
Recraft	Image	Generated by Recraft and unchanged (tested on V3)
Adobe	PNG	Adobe signs on a partner model's behalf and names it, such as GPT Image 2; not signed by OpenAI itself

The table is based on original files we received in October 2026 and covers only the models we tested. Platforms such as fal and ElevenLabs also sign files, but that does not prove the underlying model; Grok signatures cannot be verified yet.

For checking channels, OpenAI's signature is the most useful because it records whether the image came from the API or from ChatGPT. Those are different goods: the API renders the size you requested, while the ChatGPT app decides size and quality itself, around 1.5 megapixels.

Google's signature only proves the pixels are Google's untouched output. It can't tell the API from the Gemini app, or Nano Banana 2 from Pro.

Check it yourself: Yir's C2PA Viewer

We built a free tool: C2PA Viewer. Drop in an image or video; the file is parsed in your browser and never uploaded.

Watch the 20-second demo

Three fields are enough:

Signer: OpenAI, Google, ByteDance, or a third party such as Adobe.
Generated by: the API or an app like ChatGPT.
Changed since: whether the content still matches the signature.

General credential viewers, including the official Content Credentials Verify, will tell you "issued by OpenAI, generated with AI", but they don't separate the API from ChatGPT, and an upscaled file just shows up as invalid. Those two points are exactly what you need when checking a channel.

There is one case where most tools just say "no credential" and ours looks one step further: the credential was wiped. A blank block of the same length sits where the credential used to be, and its length hints at who signed it. That is a useful clue, but with the signature gone the tool will not call the file an original.

Before you start:

Use the original file. Download it from the API's result URL as-is. Don't re-save, screenshot or convert it. Chat apps and social platforms usually strip credentials.
No credential doesn't mean fake. Many models never write one; it only means there's no evidence.
One clean file doesn't make a clean channel. Channels often mix several sources, so sample more than once, over time.

Tested: APIMart and KIE

On October 7, 2026 we checked the original files APIMart and KIE delivered for each GPT Image tier, and spot-checked their Nano Banana and Seedream output; APIMart's GPT Image 2K/4K routes were retested on October 13. This is a snapshot. Sources change, and next month may look different.

APIMart

Model · route · tier	What you get
GPT Image 2 · standard route · 2K/4K	API original or Adobe re-sign, native output (verified)
GPT Image 2 · official route · 2K/4K	API original
GPT Image 2.5 Flare · standard route · 4K	API original
GPT Image 2.5 · Ext route · 2K/4K	API original
Nano Banana 2	Google's untouched output
Seedream 5.0 Pro	Vendor original

Every APIMart file we checked was native output. The official route for GPT Image 2 returns API originals. The standard route mixes API originals with Adobe Firefly re-signs at the requested size; we grade those re-signs as native output (verified), not as vendor originals. Every Nano Banana and Seedream file was the vendor's untouched output.

KIE

Model · tier	What you get
GPT Image 2 / 2.5 · 1K	Mostly untouched ChatGPT app output
GPT Image 2 · 2K/4K	API original; some files had the credential wiped, pixels untouched
GPT Image 2.5 · 2K/4K	Mostly credential wiped, pixels untouched
Nano Banana 2	Google's untouched output

None of the KIE files we tested were upscaled. 1K came from the ChatGPT app; 2K/4K were API originals, many with the credential wiped; the blank length and encoding show they started as API renders with untouched pixels. We have few GPT Image 2 text-to-image samples so far and will keep checking.

What this means

What you get depends on the route and tier, not just the channel. The channel name alone tells you little.
App renders at 1K are common. They're untouched, but the app decides size and quality, so they aren't the same as API output.
A wiped credential isn't tampering. Untouched pixels and re-encoded files are different things.

Let Yir check for you

Spot checks find problems, but you can't run them on every call. Yir checks every delivery and attaches a grade to the result:

original: vendor API output
app: vendor app output, unchanged
equivalent: no valid signature, but evidence that nothing was upscaled or changed, such as a wiped credential with untouched pixels
altered: upscaled or changed
unverified: no evidence either way

Set a floor with routing.fidelity:

genuine (default): skips any route we've declared as upscaling or altering output, such as a route that delivers upscaled ChatGPT renders as 2K/4K.
original: vendor originals only, for showcases and comparisons where you present the image as the model's own output.
any: no filter. Upscaling channels are often cheaper, and that choice is yours.

{
  "model": "openai/gpt-image-2",
  "input": { "type": "text", "prompt": "..." },
  "parameters": { "resolution": "2K", "aspect_ratio": "16:9" },
  "routing": { "fidelity": "original" }
}

If no channel meets the floor, the request fails instead of quietly downgrading. When a channel's source changes, the checks on live deliveries surface it and the default genuine steers away from it; when a channel fails, Yir moves to the next available one. You are billed by the channel that actually ran the job.

fidelity lowers the odds of getting an upscaled image; it is not a guarantee. That's why each file's actual grade is in the result, so you can decide per image whether to use it.

FAQ

Can C2PA prove an image is AI-generated?
Only when the file carries a credential. It records what the vendor itself wrote about the file's origin. It doesn't judge the picture, and it doesn't read invisible watermarks like SynthID.

Is my file uploaded?
No. The C2PA Viewer parses files in your browser and sends them nowhere.

Why does my image show "no content credentials"?
Usually because it was re-saved, screenshotted, or passed through a chat app or social platform. Use the original file returned by the API.

How can I quickly tell whether a GPT Image 4K file was upscaled?
Check how it was generated and whether it changed since signing. Generated by ChatGPT plus a content mismatch almost always means an app render that was upscaled.

Is a ChatGPT app render a fake?
It isn't altered: the vendor signature is valid and the pixels are untouched. But the app decides size and quality, so it isn't the same product as an API render. On Yir it is graded app; genuine keeps it and original excludes it.

To check a file, open the C2PA Viewer. To compare channel prices, see the GPT Image 2 model page.

MiniMax H3 API Speed by Provider: 32 Seconds or 8 Minutes for the Same Clip

Reecho1977 — Fri, 30 Oct 2026 09:46:42 +0000

I work on Yir, an image & video API router; the timings come from jobs Yir sent to each provider.

If your MiniMax H3 jobs feel slow, the model is probably not the problem. The provider is. Over the last 30 days, the same 768p H3 video took a median of 32 seconds on fal and 8.1 minutes on KIE. That's a 15× gap for the same model, and the slowest provider isn't even the cheapest one.

These numbers come from jobs Yir actually sent to each provider, not from a benchmark we ran once. "Time" here means the median from the moment a job is handed to the provider until the provider says it's done, which includes any queue on the provider's side. Our own queueing and the file download aren't counted.

Render time and price at 768p

For a 10-second 768p clip, as of October 30, 2026:

Provider	Typical time	Price per second	10-second clip
fal	32 sec	$0.06	$0.60
Metaso	45 sec	$0.0155	$0.16
APIMart	3.1 min	$0.0571	$0.57
KIE	8.1 min	$0.04	$0.40
MiniMax official	not measured	$0.08	$0.80

The official MiniMax API is listed as a price reference only; Yir doesn't route H3 jobs to it, so we have no timing for it.

Times are pooled across aspect ratios and durations at the same resolution. A 15-second clip will usually take a little longer than a 5-second one on the same provider, but the order between providers holds.

At 2K the gap is smaller but still there

Provider	Typical time	Price per second	10-second clip
Metaso	1.6 min	$0.0259	$0.26
KIE	5.9 min	$0.065	$0.65
APIMart	not enough jobs yet	$0.0914	$0.91
Higgsfield	not enough jobs yet	$0.13	$1.30
MiniMax official	not measured	$0.13	$1.30

fal doesn't offer H3 at 2K. If you need 2K and don't want to wait minutes per clip, Metaso is the only provider in this list with a measured time under two minutes.

Why the same model is fast on one provider and slow on another

Every provider runs, or resells, the same MiniMax model, so the rendering work itself is similar. What differs is everything around it: how many jobs are ahead of yours, how much GPU capacity the provider has bought, and whether it forwards your job to someone else.

We can't see inside a provider's queue, so we won't guess at the exact cause for each one. What we can see is the pattern. KIE and APIMart are consistently minutes, not seconds. fal and Metaso are consistently under a minute at 768p. That pattern has held across the whole 30-day window, so it isn't one bad afternoon.

For a background batch that runs overnight, eight minutes per clip doesn't matter much. For a product where a user clicks "generate" and stares at a spinner, it is the difference between a feature people use and one they abandon.

Metaso: the fast one that's also the cheapest

Metaso is the odd result in the table: second fastest at 768p, fastest at 2K, and about 80% below the official price. The obvious question is whether it's really H3.

Metaso runs MiniMax H3 on its own servers. We ran the same prompts on Metaso and on the official MiniMax API at 768p and 2K. Resolution, 24 fps, frame count, duration and the stereo audio track all matched, and Metaso encoded at a higher bitrate.

The one real difference: Metaso's files carry Metaso's own AI-content label instead of MiniMax's signed content credentials. If you need files that MiniMax itself signed, for example to show them as the model's untouched output, Metaso isn't for you. On Yir, a request with routing.fidelity set to original skips it automatically.

H3 Max: fal's faster variant

fal also sells MiniMax H3 Max, a version of H3 that fal post-trained for speed. It's a separate model with its own page, not a setting on H3.

H3 Max provider	768p time	Price per second
fal	17 sec	$0.052 promo, $0.08 list
APIMart	22 sec	$0.0571

fal's promo price ends October 31, 2026; after that it lists $0.08 per second, the same as official H3. H3 Max tops out at 768p, so it doesn't help if you need 2K.

Which provider to pick

A user is waiting on screen: fal H3 Max or fal H3 at 768p, or Metaso if you don't need MiniMax's own signature.
Lowest cost per clip: Metaso, at both 768p and 2K.
You need 2K: Metaso if speed matters, KIE if you can wait and want a second option.
Batch jobs where nobody waits: any of them; price matters more than time.

Remember these are medians. Even on a fast provider some jobs take much longer: across all H3 jobs on Yir in the same 30 days, the end-to-end median was 1 minute 17 seconds and the 95th percentile was 4 minutes 31 seconds. Build your UI for the slow tail, not the median.

Letting Yir pick by speed

You don't have to hard-code one provider. Yir routes each job across the providers that serve the model and fails over if one breaks. By default it picks the cheapest available provider. Ask for speed instead:

{
  "model": "minimax/minimax-h3",
  "input": { "type": "text", "prompt": "..." },
  "parameters": { "resolution": "768P", "duration": 10 },
  "routing": { "preference": "speed" }
}

With preference: "speed", Yir orders providers by their recent median time for the same spec, the same measurement as the tables above, and only ranks a provider by time once it has enough measured jobs. You're billed at the price of the provider that actually ran the job, so the fast route may cost more than the cheap one.

Live prices and times for every provider are on the MiniMax H3 model page and the H3 Max page. For a breakdown of what clips cost with reference images, see MiniMax H3 API pricing.

FAQ

How long does MiniMax H3 take to generate a video?
It depends on the provider far more than on the clip. A 768p clip took a median of 32 seconds on fal, 45 seconds on Metaso, 3.1 minutes on APIMart and 8.1 minutes on KIE over the last 30 days.

What is the fastest MiniMax H3 API?
At 768p, fal's H3 Max at 17 seconds, then fal's H3 at 32 seconds. At 2K, Metaso at 1.6 minutes.

Why is my MiniMax H3 job stuck for several minutes?
Most likely it's waiting in the provider's queue. If you're on a provider whose typical time is several minutes, that's normal for it. Switching providers changes it more than changing your prompt or settings.

Is the faster provider more expensive?
Not necessarily. Metaso is both one of the fastest and the cheapest. KIE is the slowest but costs less than fal and APIMart.

Are these numbers guaranteed?
No. They're 30-day medians of real jobs and they move as providers add capacity or get busier. The model page shows the current numbers.