Seedance 2.5 Face Restrictions Explained — and What Actually Works

TL;DR: Dreamina does not accept reference images or videos containing realistic human faces. This is a platform policy, not a quality problem with your photo — a face-detection pass runs before generation and rejects the upload outright. Better lighting, a bigger file or a different crop will not change the outcome. What works: stylised or illustrated faces, ByteDance's own avatar and asset-authorisation routes, or running Seedance 2.5 somewhere that doesn't apply that gate.

VidCella · Seedance 2.5

Reference images, without the face gate.

4–30s · 480p to 4K · Up to 30 reference images · Pay-as-you-go

If you have tried to give Seedance 2.5 a photo of a person on Dreamina, you have probably hit a wall that doesn't explain itself. The upload is refused, or the job dies immediately, and the message says nothing about faces. People reasonably conclude their image was too low-resolution, badly cropped, or the wrong format, and start re-exporting it.

That's the wrong diagnosis, and it costs a lot of wasted attempts. Here's what the rule actually is.

The rule, in ByteDance's own words

Dreamina's support account has stated the policy directly:

Due to platform compliance requirements, currently uploading creatives (both image and video creatives) containing realistic human faces is not supported on Dreamina. To safeguard user rights and content safety, the system will automatically block such creatives.

Three things in that sentence matter:

"Both image and video creatives." It isn't only reference photos. Video uploads carrying a recognisable face are covered by the same rule.

"Realistic human faces." The classifier is looking for photorealism, not for people. That's why the restriction feels inconsistent — see below.

"Automatically block." Nothing is queued for review and nothing is appealable in the moment. A detection is a refusal.

This tightened after the copyright dispute with major studios, during which ByteDance also pulled human reference input and Face-to-Voice from the product entirely.

What you'll actually see

The message you get depends on where the check fires, and none of them mention faces:

Message	When it fires
Not eligible	Moderation rejection — commonly the reference image, but the same wording covers prompt and output blocks
Visual restriction detected	Specifically a flag on the visual content of an uploaded image
The text you entered does not comply with the platform’s rules.	Prompt-side block, not the image
Not available in your region	Geographic restriction, unrelated to content

If you're seeing one of the first two after adding a photo, the face gate is the likely cause. If you're seeing the third with no image attached at all, that's a prompt problem and a different fix.

Why it catches things you didn't expect

The filter is a photorealism classifier, and that produces results that feel arbitrary until you see the logic:

Your own face is not an exception. There is no verification flow that says "this is me, I consent." The classifier can't tell whose face it is, so consent isn't a variable it can weigh.

Glasses, helmets and sunglasses don't defeat it. Partial occlusion generally doesn't drop a real photograph below the detection threshold.

Highly realistic AI-generated faces are frequently blocked too. This surprises people who assumed a synthetic person would be exempt. The classifier is judging whether the image looks like a real human face, not whether a real human was photographed — so a good synthetic portrait trips exactly the same wire.

Stylised and illustrated faces usually pass. Anime, painterly, obviously rendered or cartoon faces sit far enough from photoreal that the classifier lets them through. This is the single most reliable predictor of whether a given image will be accepted.

What actually works

1. Use a stylised face instead of a photograph. If your project can tolerate an illustrated or rendered character, this clears the restriction on Dreamina with no further effort. It's the path of least resistance when the aesthetic is flexible.

2. Use ByteDance's own sanctioned routes. They document a few: pre-set virtual avatars, AI-generated portraits intended for this purpose, and asset-id authorisation for licensed material. These exist precisely because the blanket rule would otherwise make legitimate commercial work impossible.

3. Describe the person instead of uploading them. Text-to-video never touches the image filter. You lose likeness control, which may or may not matter depending on the shot.

4. Run Seedance 2.5 somewhere the platform gate doesn't apply. The restriction belongs to Dreamina's product policy, not to the Seedance model. The same model served through a different provider doesn't inherit ByteDance's consumer-app face rule.

A note on what doesn't work, so you don't waste the attempts: re-exporting at higher resolution, cropping tighter, converting the file type, lowering the face's prominence in frame, or resubmitting the identical image hoping for a different result. The detection is on the content, and it is consistent.

Seedance 2.5 · Reference-driven video · Pay-as-you-go

Your reference photo, actually accepted

VidCella serves Seedance 2.5 through a different upstream to Dreamina, so the platform face gate that rejects your reference image isn’t part of the path. Up to 30 reference images, 4 to 30 seconds, 480p to 4K.

From 45 credits per second · No subscription

What that looks like on VidCella

We serve Seedance 2.5 through a different upstream to Dreamina, so the consumer-app face gate isn't in the path. Reference images work the way the model supports them — up to 30 images, 10 videos and 10 audio clips in one generation.

We run our own safeguards, as any serious platform does, and they're built for precision: aimed at genuinely prohibited material rather than sweeping whole categories of ordinary creative work. Uploading a reference photo of a person is not, by itself, what they're looking for.

What you get here:

4 to 30 seconds, 480p through 4K
Up to 30 reference images, 10 videos, 10 audio clips per generation
Text-to-video, image-to-video and video editing
From 45 credits per second, pay-as-you-go, no subscription

The responsibility that comes with that doesn't move. Real-person likeness carries obligations wherever you generate it — consent for identifiable people, and applicable law on publicity and defamation. A platform not blocking something is not the same as it being fine to publish.

FAQs

Why does Dreamina reject my reference photo? Because it contains a realistic human face. Dreamina does not support uploading image or video creatives containing realistic human faces, and the block is automatic.

Is this a problem with my image quality? No. Resolution, cropping and file format have no bearing on it. A face-detection pass runs on the content before generation and refuses the upload regardless of technical quality.

Can I upload my own face? No. There's no consent or verification path — the classifier detects a realistic face without regard to whose it is.

Why was my AI-generated portrait blocked? Because the classifier judges photorealism, not provenance. A synthetic face that looks convincingly real trips the same detection as a photograph.

What kinds of faces do get through? Stylised ones — illustrated, anime, painterly, clearly rendered. The further from photoreal, the more reliably they pass.

Does the restriction come from the Seedance model itself? No. It's Dreamina's platform policy, applied before the model runs. Seedance 2.5 served through another provider doesn't carry that consumer-app rule.