Image-to-video is the adult creator's workflow, and the reference frame decides more of the result than the prompt does. This guide covers how upload classifiers score your image, why they reject clothed pictures, and the preparation steps that make Wan 2.7 and Seedance 2.0 animate a frame cleanly.
Short answer: in image-to-video, the reference frame does most of the work. On filtered services an upload classifier scores the frame before the model sees it, and it scores pose, skin area and context, not just nudity, which is why fully clothed images get rejected. On a host with adult content enabled that classifier is gone, and the quality of the clip comes down to five properties of the frame: sharpness, lighting, framing, pose simplicity and the absence of text. Get those right and both Wan 2.7 and Seedance 2.0 animate reliably.
Why the frame matters more than the prompt
Text-to-video asks the model to invent a subject, a setting and a motion at once. Image-to-video removes two of those: the subject and the setting are given, and the model only has to invent motion. For adult work that is decisive, because identity, anatomy and lighting are exactly the things that go wrong when the model invents them. In our runs, the same prompt produced usable adult clips roughly twice as often in image-to-video as in text-to-video, on both Wan and Seedance.
The flip side: a weak frame produces a weak clip no matter how good the prompt is. Every artefact in the frame is animated.
How upload classifiers score an image
Filtered services (Kling, hosted Seedance, Veo, most API resellers) run an image classifier on the upload. From testing across several of them, the score is driven by more than nudity:
Signal | Effect on the score | Practical consequence |
|---|---|---|
Visible skin area | Strong | Swimwear and lingerie often rejected even with no nudity |
Pose | Strong | Lying down, arched back, spread limbs score high regardless of clothing |
Context objects | Moderate | Beds, showers and some props raise the score |
Face expression | Weak to moderate | Open-mouth or “suggestive” expressions add to the score |
Actual nudity | Decisive | Rejected outright |
This is why creators report that a clothed image was refused: the classifier is estimating intent, and it errs on the side of refusal. There is no reliable way to engineer around it; a frame that scores low enough to pass is usually one that cannot produce the clip you wanted.
On a platform with adult content enabled there is no upload classifier, and the rest of this guide assumes that setting. MyBabes runs Wan 2.7 and Seedance 2.0 that way, and all the test clips here were generated there.
The five properties of a frame that animates well
1. Sharpness
Soft input produces soft output plus motion smear. Enhance the frame before uploading; an uncensored enhancer at low creativity (see our Krea analysis) is the fastest way. Check hands, eyes and hairline at 100% zoom.
2. Lighting
Even, directional light with visible shadow shape gives the model a lighting model to carry through the clip. Flat, shadowless frames produce clips where the subject appears to float. Hard rim light and heavy backlight both cause flicker.
3. Framing
Leave room for the motion you are about to request. If the prompt says “she leans forward”, a frame cropped at the forehead will crop her out. Medium shots with head room and a little space on the motion side animate best; extreme close-ups animate worst on both models.
4. Pose simplicity
The model has to infer the 3D body from a 2D frame. Crossed limbs, foreshortened arms and heavy occlusion are where it guesses wrong, and a wrong guess becomes a limb that grows or vanishes mid-clip. Start from a clear, readable pose and let the prompt add the complexity.
5. No text, logos or borders
Anything with lettering gets animated into garbage, and borders are read as part of the scene. Crop them out. This is the single most common avoidable failure we see.
Preparing a frame: the checklist
Generate or select the frame at the highest quality available.
Crop to the target aspect ratio of the video (usually 16:9 or 9:16). Do not let the video model crop for you.
Enhance at low creativity; check anatomy after.
Remove text, watermarks, borders.
Confirm the pose reads clearly at thumbnail size. If you cannot tell what the body is doing at 200 px, neither can the model.
Save at the model’s native input resolution or higher.
Matching prompt to frame
The prompt must agree with the frame. Contradictions are resolved randomly and usually badly.
Describe the subject as they appear, not as you wish they appeared. Changing hair colour or clothing via prompt in image-to-video produces flicker.
Describe motion that is physically reachable from the pose. A seated subject can lean, turn and reach; asking them to walk produces a morph.
Keep it to two or three beats and one camera instruction. Short prompts win on both models.
Model-specific notes
Wan 2.7
Tolerates busier frames and contact poses better than Seedance. Prefers a static or slow-push camera. Strong at keeping the reference identity through the full clip. Best for explicit beats.
Seedance 2.0
Wants a cleaner, more cinematic frame: good light, clear composition. Rewards camera instructions and produces the more polished result on medium shots. Slightly more likely to lose a limb in close contact. Best for build-up and atmosphere.
The two are complementary; many creators generate the same frame on both and cut between them. Details in the 2026 model ranking.
Common failures and what they mean
Failure | Usual cause | Fix |
|---|---|---|
Face changes mid-clip | Soft or small face in the frame | Enhance; use a closer medium shot |
Extra or vanishing limb | Occluded or foreshortened pose | Simplify the pose in the frame |
Flicker in hair or clothing | Prompt contradicts the frame | Describe what is actually there |
Smeared motion | Prompt requests fast motion | Slow the beats; slow camera |
Subject drifts out of frame | No room on the motion side | Re-crop with head room and lead room |
Garbled patches | Text, logo or border in the frame | Crop it out |
Related in the Lab
Key takeaways
In image-to-video the frame decides more than the prompt; a clean frame roughly doubles usable output.
Upload classifiers score pose, skin and context, not only nudity, which is why clothed frames get refused.
Sharpness, directional light, room for motion, a readable pose and no text are the five properties that matter.
Prompt what is in the frame and motion that is reachable from it.
Wan 2.7 for contact and identity, Seedance 2.0 for cinematic motion; both need the same frame preparation.
FAQ
Why was my clothed image rejected by an AI video tool?
Upload classifiers estimate intent from pose, skin area and context, not just nudity. Lingerie, lying poses and bedroom settings all raise the score.
What resolution should the reference image be?
At or above the model’s native input resolution, sharp at 100% zoom. Enhance rather than upscale if the source is soft.
Can I change the outfit or hair with the prompt in image-to-video?
Not reliably. Contradicting the frame causes flicker. Change the frame instead, then animate.
Which model is better for NSFW image-to-video?
Wan 2.7 for close contact and identity consistency, Seedance 2.0 for camera movement and lighting. Both are available with adult content enabled on hosts that support it.
Does the reference image need to be explicit for an explicit clip?
No. A clean, non-explicit frame with a readable pose plus an explicit prompt usually produces a better clip than an explicit but cluttered frame.
Last tested: August 2026.
Models covered
Skip the filter
Run these models with adult content enabled
MyBabes runs Wan 2.7, Seedance 2.0 and Krea inside its own generation stack, so the prompts refused on official apps generate as written. No GPU, no setup.
Open MyBabes
Builds MyBabes, runs the Lab
Jiri builds MyBabes and runs the Lab’s model testing: the same prompt set on every image, video and chat model, on the vendor’s official surface and on MyBabes, re-run after every release. He writes up what the models actually do, not what the launch posts say.
Follow on X
More from the Lab
Latest articles
All articles
Analysis
AI Porn and the Law in 2026: What's Actually Illegal, What's Regulated, and Where the Lines Sit
2025 and 2026 turned AI adult content from a legal grey zone into a regulated space with real enforcement: a US federal crime for non-consensual imagery, deepfake laws in 48 states, UK fines for missing age checks, EU labeling duties. Here is the map — and the one principle that organises all of it.
·
5 min read
How-To
Why AI Porn Looks Off: Every Artifact Explained, What Causes It, and How to Fix Each One
Everyone can tell when an AI image is off, few can say why. Each artifact has a specific technical cause and a specific fix — hands, limbs, faces, skin, teeth, text, seams, smear, flicker. This is the field guide, for images and video.
·
5 min read
Analysis
ChatGPT and NSFW in 2026: The Adult Mode That Never Shipped, What's Actually Allowed, and Where People Went Instead
"Treating adults like adults" was the promise in October 2025. Ten months later the adult mode is paused with no timeline, ChatGPT still refuses explicit content, and the search interest hasn't gone anywhere. Here is the full story, the current policy line in practice, and what the alternatives actually are.
·
4 min read