Most AI video tools give you one image and ask the model to guess where it goes. VirtuaVixen’s LTX 2.5 First-to-Last Frame workflow flips that: you hand it a start image and an end image, and the model is forced to travel from one to the other. This is genuine keyframe control — the single most requested feature for people who want a specific reveal, not a random hallucination. It’s the difference between hoping and directing.

What “First-to-Last Frame” Actually Means

A first to last frame ai video is built from two pinned endpoints. Frame one is exactly where your clip begins. The final frame is exactly where it ends. Everything in between — roughly 10 seconds of motion — is generated to smoothly connect those two states. You are not describing a transformation in words and praying; you are showing the model the “before” and the “after” and letting it fill the gap.

That control unlocks the reveals people actually want:

• Clothed → naked. Start dressed, end nude. The strip happens on-screen.
• Soft → hardcore. Begin with a tease, land on penetration.
• Position → position. Cowgirl to doggy, standing to on-the-bed — a real transition instead of a hard cut.
• Calm → orgasm. Composed face at the start, mid-climax expression at the end.
• Before → after cumshot. The classic money-shot arc, with a defined beginning and payoff.

How the Keyframe Guides Work

Here’s the part that separates this from ordinary image-to-video. In a normal I2V workflow, one image seeds the first frame and the model drifts wherever it likes. In the First-to-Last Frame graph, two LoadImage nodes feed two LTXVAddGuide nodes — one guide pinned to the opening frame, one pinned to the closing frame. Those guides constrain the diffusion at both ends of the timeline, so the model literally cannot ignore your endpoints. It has to start at your first image and arrive at your last.

That dual-guide pinning is the defining technical difference. A single-image workflow only knows where to begin. This one knows where to begin and where it must finish, which is why the motion in the middle reads as an intentional transition rather than a slow wander. After sampling, an LTXVCropGuides node strips the guide latents back out so they don’t leave artifacts in the decoded result.

Native Synced Audio, Not a Bolted-On Track

The moans and wet sounds aren’t dubbed on afterward. This graph generates audio in the same pass as the picture. An LTXVEmptyLatentAudio node creates an audio latent that gets stitched to the video latent by LTXVConcatAVLatent, the two are denoised together as one joint latent, then split apart by LTXVSeparateAVLatent. Because sound and image are sampled jointly, lip movement, thrust rhythm, and audio land in sync instead of sliding out of time.

Under the Hood: The Exact LTX 2.5 Stack

For the technically curious, here is the precise stack that powers this workflow. Note that this graph differs from VirtuaVixen’s other LTX 2.5 workflows — it is the only one built around dual keyframe guides, and it deliberately skips frame interpolation.

• Base model: LTX-2.5-Distilled-Q8_0.gguf — a Q8_0 GGUF quantization of the 22B distilled LTX 2.5 transformer, loaded through UnetLoaderGGUF. The distilled variant is what makes ~10-second clips viable on hosted hardware without a quality collapse.
• Text encoding: handled by a CLIPLoader node feeding the prompt into the graph.
• Guidance: LTXVDualCFGGuider, which steers the joint audio-video denoise.
• Keyframe guides: two LTXVAddGuide nodes (first frame + last frame), cleaned afterward by LTXVCropGuides.
• Audio path: LTXVEmptyLatentAudio → LTXVConcatAVLatent → joint denoise → LTXVSeparateAVLatent.
• Sampling: SamplerCustomAdvanced driving a SamplerEulerAncestral with an explicit ManualSigmas schedule — tight, deterministic control over the denoise curve.
• Decoding: two separate VAEs. Video is decoded by ltx-2.5-video-vae-conv-bf16.safetensors; audio is decoded by ltx-2.5-audio-vae-bf16.safetensors through an LTXVAudioVAEDecode node.
• Output: a CreateVideo node muxes the decoded frames and audio, then SaveVideo writes the file.
• Resolution: a 9:16 portrait ResolutionSelector at roughly 1.0 megapixel, snapped to a multiple of 32 — built for vertical shorts and phone screens.
• No RIFE: unlike the sibling LTX 2.5 graphs, this one does not run RIFE frame interpolation. The keyframe guides already anchor the temporal path, so motion is generated natively end-to-end.

The LoRA Stack (First-to-Last Frame)

Three LoRAs are stacked on top of the base transformer via LoraLoaderModelOnly, tuned specifically for this keyframe pipeline:

• LTX-2.3-22b-AV-LoRA-talking-head-v1.safetensors @ 0.88 — the audio-video sync LoRA that keeps mouths, moans, and motion locked together.
• DR34ML4Y_LT3X_V3.safetensors @ 0.8 — NSFW realism and motion, the LoRA that keeps skin, bodies, and movement believable through the transition.
• ltx23-ultimatedt-NSFW-sulphured_audio_v2_k3nk.safetensors @ 1.0 — deepthroat detail plus wet audio, run at full strength here so the sound is as explicit as the picture.

How to Make One in the Studio

You don’t touch a single one of those nodes. The whole graph runs hosted on VirtuaVixen’s GPUs — no local install, no VRAM ceiling, no downloads.

1. Open the AI Porn Studio and select the LTX 2.5 First-to-Last Frame workflow.
2. Pick your first frame. Each of the two frame slots has its own Library picker, so you can pull the opening image straight from your own generation history — a T2I still, a previous video frame, anything you’ve made.
3. Pick your last frame the same way, from the second slot’s Library picker. This is your destination.
4. Add a short prompt describing the action between the two frames, then generate. About 10 seconds of synced, audio-included video comes back, ready to save or post to Shorts.

Tips for Choosing Good Start and End Frames

• Keep the same character. The closer your two frames match in identity, lighting, and camera angle, the smoother the interpolation. Two frames of the same AI pornstar in the same scene work far better than two unrelated images.
• Make the change readable. The most satisfying results come from a clear, single transformation — clothed to naked, soft to hard — not five things changing at once.
• Match the framing. If your start frame is a medium shot and your end frame is an extreme close-up, the model has to travel a long way; keep them in the same rough composition for the cleanest motion.
• Use the portrait aspect. This workflow is tuned for 9:16, so choose or crop frames that suit a vertical shot.

Who It’s For

First-to-Last Frame is for anyone who wants to direct a reveal instead of gambling on it. If you have a killer start image and know exactly where you want the scene to end, this is the workflow that connects them. Pair it with the sibling LTX 2.5 tools — Epic Cumshot, Smooth Sex, Sucking, and Flirt & Dialogue — to cover every beat of a scene, from the first tease to the final frame.

Two images in, one directed 10-second clip out. No GPU, no setup, and generation is included with your tier. Head to the AI Porn Studio and pin your first and last frames.