What is start-end frame generation?
Video generation guided by two keyframes, the first and the last image: how the interpolation works, which models support it, and what breaks when the frames differ too much.
Technical background for readers of tool reviews; legal references are general information, not legal advice.
Start-end frame generation is a video-AI mode in which the user supplies two images and the model produces the motion that connects them; Alibaba calls it first-last-frame-to-video (FLF2V). The first image fixes the opening frame, the second fixes the closing frame, and a text prompt describes the transition. It gives tighter control than single-image animation.
Keyframe-guided video modes at a glance
Checked against the linked sources on September 3, 2026; no editor scores here (how we rate).
| Term | What it does | Where used | Limits |
|---|---|---|---|
| Start and End Frames (Kling) | Two images become first and last frames of a transition clip | Kling image-to-video via “Add End Frame”; Kling 3.0 binds up to 3 elements | Images should share a theme, about 5 s; large differences trigger a shot switch |
| FLF2V-14B-720P (Wan2.1) | Open-weight first-last-frame model, April 17, 2025 | Local or cloud inference; Apache 2.0 | 720P only; 14B parameters need a large GPU |
| Start-end to video (Vidu) | API endpoint taking exactly two images | Vidu Q3 (1–16 s, 540p–1080p), Q2 (1–8 s), Q1, 2.0 | Aspect ratios within 0.8–1.25 of each other; 50 MB per image |
| First and Last Frame (Veo 3.1) | Transitions between two provided images | Veo 3.1; 8-second clips, 1080p or 4K, SynthID | Fixed length; hosted only |
| Image-to-video (I2V) | Animates a single first frame, no end constraint | Wan2.2 I2V-A14B, Veo, Vidu, Kling; undress apps | Ending unpredictable; subject drifts |
How does keyframe guidance work?
A video diffusion model denoises a block of frames together. In start-end mode the first and last positions are clamped to the encoded input images, and the model generates the intermediate frames so that motion, lighting and identity flow from one anchor to the other. The prompt describes what happens in between, but the anchors bound the result.
Both frames should therefore depict the same subject and setting. Kling’s guide warns that “significant differences may cause a lens switch,” meaning the model cuts to the second image as a new shot. Vidu enforces a related constraint at the API level: the aspect ratios of the two images must fall within 0.8 to 1.25 of each other.
Which models support start-end frames?
Kling exposes the feature as “Add End Frame” inside image-to-video and recommends transitions within 5 seconds. Alibaba’s Wan2.1 released FLF2V-14B-720P on April 17, 2025 as open weights under Apache 2.0, 720P only. Vidu’s start-end endpoint accepts two images across Q3 (1–16 seconds, 540p to 1080p), Q2 (1–8 seconds), Q1 and 2.0. Google’s Veo 3.1 offers First and Last Frame with 8-second output at 1080p or 4K, marked with SynthID. Entries: Kling, Wan, Vidu, Veo.
What are the limits?
Three failure modes dominate. First, the shot switch: when frames differ in pose, scale or background, the model cuts instead of interpolating. Second, morphing — hands, hair and fabric smear during fast transitions. Third, duration: open models produce about 5 seconds and hosted ones 8 to 16, so longer scenes are chained and small errors accumulate.
Where does it appear in undress apps?
No undress app in the current ranking exposes an end-frame control; the video modes reviewed in Ainudez, Deep Undress and Pornworks animate a single image, as described under image-to-video. The technique is still relevant: a before-and-after pair from an undress workflow is exactly the two-frame input FLF2V models animate, which is why reviews check whether a video feature accepts a second image. A clip made this way from a real person’s photo is a “digital forgery” under the US TAKE IT DOWN Act and synthetic video that must be marked under the EU AI Act; the consent rule of the Responsible AI policy applies to both frames. This page is general information, not legal advice.
Sources
All sources accessed September 3, 2026.
- Kling AI — Start and End Frames guide
- Wan2.1 — FLF2V-14B-720P, April 17, 2025 — GitHub
- Vidu API — start-end to video, 0.8–1.25 ratio
- Veo — First and Last Frame, SynthID — Google DeepMind
- Wan2.2 — I2V-A14B and TI2V-5B — GitHub
- TAKE IT DOWN Act, S.146 — “digital forgery” — Congress.gov
Corrections: editorial@clothoff.ai · Editorial policy
Frequently asked questions
What is FLF2V?
First-last-frame-to-video, the name Alibaba uses for its Wan2.1 model that interpolates between a start and an end image. FLF2V-14B-720P was released on April 17, 2025 as open weights under the Apache 2.0 license and currently outputs 720P only. The term is now used generically for start-end frame generation.
How similar must the two frames be?
Very. Kling recommends two images with the same theme and warns that large differences trigger a shot switch instead of a smooth transition. Vidu requires the aspect ratios of the two images to stay within 0.8 to 1.25 of each other. Matching subject, framing and lighting produces the most stable motion.
How long can a start-end frame clip be?
Kling suggests transitions within 5 seconds; Vidu Q3 allows 1 to 16 seconds and Q2 1 to 8 seconds; Veo 3.1 produces fixed 8-second clips; Wan2.1 FLF2V examples run about 5 seconds. Longer sequences are chained by reusing the end frame as the next start frame.
Is start-end frame generation available in undress apps?
Not as an explicit control in the tools tested so far. Ainudez, Deep Undress and Pornworks animate a single image. The editorial team checks at each review update whether a video feature accepts a second image, because a before-and-after pair is the natural input for this technique.
Can start-end frame output be detected?
Veo 3.1 output is marked with SynthID, and hosted platforms increasingly add C2PA metadata. Open Wan2.1 FLF2V output has no mandatory mark. From August 2, 2026 providers serving EU users must mark synthetic video in a machine-readable way under Article 50(2) of the EU AI Act.
Is it legal to make a two-frame transition of a real person?
Only with that person’s consent. An intimate transition of a real person made without consent is a “digital forgery” under the TAKE IT DOWN Act, with up to 2 years in prison for adult subjects and 3 for minors, and covered platforms must remove it within 48 hours. Both input frames fall under the same consent rule.