The LTX 2.5 Smooth Sex workflow is our image-to-video ai sex video generator built for one thing: turning a single still image into a smooth, believable ~10-second sex clip with synced audio baked in. You upload one picture, add a short prompt, and the studio does the rest on our hardware. No GPU, no ComfyUI install, no driver hell. It runs in your browser.

This is the “smooth” specialist in our LTX 2.5 family. Where other workflows chase a money shot or a talking-head performance, Smooth Sex is tuned for one deceptively hard thing: continuous, non-jittery motion where two bodies actually move together and the sound follows the rhythm.

What it actually does

Feed it one image of a couple (or a solo scene) and a short description of the action. It generates a short video, typically around 10-15 seconds, in which the motion is fluid and the audio, moans, skin contact, breathing, tracks the on-screen thrusting rather than floating on top as a random loop. A verified output from this graph runs 1344×896 landscape, roughly 15 seconds, with an AAC audio track muxed directly into the file.

The payoff is that the clip doesn’t have the twitchy, strobing, “AI is guessing every frame” feel. Hips, hands, and hair move continuously. That smoothness is the entire point, and it is the hardest part to get right.

Why “smooth” motion is so hard, and how LTX 2.5 solves it

Most image-to-video models generate each frame semi-independently and then hope the frames agree with each other. They don’t, quite. You get micro-jitter, limbs that pop, and audio that has nothing to do with the picture. Sex footage is the worst-case stress test: repetitive, high-contact, rhythmic motion where the human eye instantly catches anything that stutters.

LTX 2.5 attacks this on three fronts. First, it generates picture and sound jointly in a shared latent space, so the audio is a native part of the same denoising process, not a bolted-on afterthought. Second, the graph runs a two-stage sample-then-refine pass plus optical-flow frame interpolation, so the final frame rate is filled in coherently rather than duplicated. Third, and this is the detail unique to Smooth Sex, it runs the full int8 safetensors build of the transformer instead of a compressed GGUF quant, which keeps motion consistency noticeably cleaner. More on that below.

How to make one in the studio

1. Open the AI Porn Studio and pick the LTX 2.5 Smooth Sex workflow.
2. Upload your starting image. This is the first frame the motion grows out of, so it matters a lot (tips below).
3. Write a short prompt describing the action and the sound you want (see prompt tips).
4. Hit generate. It queues on our GPUs and comes back as a finished MP4 with synced audio, ready to download or post to Shorts.

The number of tokens per clip depends on your tier, and it’s all hosted. You never touch a GPU, a model file, or a node graph. Browse the models and characters over on the AI Pornstars page if you want a starting subject.

Best starting images and prompt tips

Because this is image-to-video, your first frame sets the ceiling. A few things that reliably help:

• Clear, in-focus subjects. The model animates what it can see. Soft or noisy source images animate into mush.
• Bodies already in contact / mid-pose. A pose that already reads as “mid-action” gives the motion somewhere to go. A stiff, arms-at-sides pose gives it nothing to continue.
• Aspect ratio near 3:4 or landscape. The graph targets roughly a 1.2-megapixel canvas (dimensions rounded to multiples of 32); source images close to that shape need the least cropping.
• Prompt the rhythm and the sound. Describe the motion (“slow, deep thrusting”) and the audio you want (“soft moaning, skin contact”). Because audio is generated jointly, prompting it actually steers the sound.

Under the Hood: The Exact LTX 2.5 Stack

Here is exactly what runs when you hit generate. If you have used ComfyUI, this will read like home.

Base model and text

The transformer is ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors: the 22-billion-parameter distilled LTX 2.5 transformer in int8 (convrot), loaded through the standard UNETLoader node. This is the load-bearing distinction for Smooth Sex: it is not a GGUF quant. It’s the full int8 safetensors build. GGUF quants shave VRAM further but soften temporal consistency, which shows up immediately as jitter in repetitive motion. Running the full int8 build is a big part of why this workflow holds especially clean, consistent motion. Text conditioning loads via CLIPLoader, and guidance is handled by LTXVDualCFGGuider.

Native synced audio

The audio is not post-processed onto the video, it is generated in the same latent space. An empty audio latent (LTXVEmptyLatentAudio) is concatenated with the video latent via LTXVConcatAVLatent, and the two are denoised together so sound and picture share the same generative pass. Afterward they’re split apart with LTXVSeparateAVLatent and decoded by two separate VAEs:

• ltx-2.5-video-vae-conv-bf16.safetensors decodes the picture.
• ltx-2.5-audio-vae-bf16.safetensors decodes the sound (moans, skin contact, breathing) through LTXVAudioVAEDecode.

Because generation is joint, the audio tracks the thrusting rhythm instead of drifting. That’s the technical reason the moans line up with the motion.

Two-stage sampling and interpolation

Sampling runs in two stages. The base pass uses KSamplerSelect with euler_ancestral feeding a SamplerCustomAdvanced node driven by a ManualSigmas schedule. Then LTXVLatentUpsampler upscales the latent and a second refine pass adds detail. Finally, RIFE VFI (optical-flow frame interpolation) synthesizes the in-between frames, which is where the literal “smoothness” the workflow is named for comes from: real interpolated motion, not duplicated frames.

Prep, I2V, and output

The source image is prepared through LTXVPreprocess and ImageResizeKJ. Image-to-video is driven by LTXVImgToVideoInplace, and a ResolutionSelector targets a 3:4-ish frame at roughly 1.2MP (again, multiples of 32). The finished frames are assembled by CreateVideo and written out by SaveVideo. That is how a single still becomes a ~15s 1344×896 clip with an AAC track.

The exact LoRA stack, and why the strengths are low

Four LoRAs stack on the base model through LoraLoaderModelOnly nodes. Note how restrained the strengths are, that is deliberate. Pushing these harder produces exaggerated, jittery motion; keeping them low is a big part of why Smooth Sex stays smooth:

• LTX-2.3-22b-AV-LoRA-talking-head-v1.safetensors @ 0.88 — audio-video sync. The one intentionally strong LoRA, because tight A/V sync is the whole game.
• ltx2-3d-animations-12500-steps-k3nk.safetensors @ 0.5 — fluid, continuous 3D-style motion. This is the anti-jitter LoRA, held at a moderate strength so it smooths without stylizing.
• ltx23-ultimatedt-NSFW-sulphured_audio_v2_k3nk.safetensors @ 0.4 — explicit motion plus wet audio, kept low on purpose so the explicitness doesn’t overrun the motion.
• DR34ML4Y_LT3X_V3.safetensors @ 0.3 — NSFW realism, kept low so it adds grit without fighting the smoothing LoRAs.

The overall design philosophy is “one strong sync LoRA, everything else dialed back.” Combined with the full int8 transformer and RIFE interpolation, that restraint is what separates smooth from strobing.

Where it fits in the LTX 2.5 family

Smooth Sex is one of several LTX 2.5 image-to-video workflows in the studio, each tuned for a different job:

• Epic Cumshot — built around the finish rather than sustained motion.
• Sucking — oral-focused motion and pacing.
• Flirt & Dialogue — talking-head performance and lip-sync, leaning on the AV/talking-head side of the stack.
• First-to-Last Frame — you supply a start and an end image and it generates the transition between them.

If your goal is clean, continuous, rhythmic action from a single image, with sound that actually matches, Smooth Sex is the one to reach for. Pick a sharp first frame, prompt the rhythm and the sound, and let the hosted stack do the heavy lifting. Start in the AI Porn Studio.