The LTX 2.5 Sucking workflow turns a single still image into a short, realistic ai blowjob video — complete with wet, synced audio that matches the motion on screen. No GPU, no ComfyUI install, no juggling model files. You upload a picture, tweak a prompt, and the whole graph runs hosted on VirtuaVixen’s AI Studio. In about ten seconds you get a clip where the sound of the blowjob actually lines up with the head-bob, because both the picture and the audio are generated together in the same pass.

This article walks through what the workflow does, exactly how the synced audio and motion are produced, how to run it yourself, which starting images and prompts give the best results, and how to chain it with the other LTX 2.5 workflows on the studio. There is also a full “under the hood” breakdown for anyone who wants to know the precise model, LoRAs, and ComfyUI nodes doing the work.

What the LTX 2.5 Sucking Workflow Actually Does

Image-to-video (I2V) is the core trick: you give it one frame, and the model imagines the seconds that follow. What sets the Sucking workflow apart from generic I2V is that it was tuned specifically for oral motion — the head movement, the throat depth, the rhythm — and it generates the matching wet audio in the same generation, not as a separate voiceover bolted on afterward.

That joint generation is the whole point. Most “AI video with sound” pipelines make the video first, then try to slap audio on top, and the timing never quite lands. Here the audio and the picture are denoised as one connected latent, so the slurp and suck cues are locked to the actual motion frame-for-frame. The result feels real because the sound was born from the same math as the movement.

How the Synced Audio and Motion Work

Under LTX 2.5, audio isn’t an afterthought — it lives inside the same latent space as the video. When you hit generate, an empty audio latent is created and concatenated directly onto the video latent, so the two travel through the sampler as a single joined tensor. The denoiser shapes picture and sound at once, which is why the wet sounds sync to the head-bob motion instead of drifting.

After sampling, that joined latent is split back into its two halves. The picture half is decoded by the LTX 2.5 video VAE; the audio half is decoded by a dedicated audio VAE that turns the audio latent into the actual wet, breathy sounds you hear. Two decoders, one shared generation. That shared origin is the entire reason the timing feels believable rather than dubbed.

On top of the base motion model, a stack of specialized LoRAs steers the behavior toward deepthroat depth, realistic skin and body motion, and — critically — the correct audio cues. One of those LoRAs was literally trained on “sulphured” audio so the throat sounds land at the right moment. The exact stack and strengths are listed in the “Under the Hood” section below.

How to Make One on the Studio

The whole thing runs in your browser. Here’s the flow:

1. Open the AI Studio. Head to the AI Porn Studio and pick the LTX 2.5 Sucking workflow from the list.
2. Upload a starting image. Drop in a still that already shows the two subjects positioned for the act — the closer the pose is to the moment before contact, the better the model extrapolates.
3. Write a short prompt. Describe the action and the mood in plain language. You don’t need to describe the audio; the workflow handles that automatically.
4. Generate. Roughly ten seconds later you get a short clip with synced wet audio, ready to download or post to Shorts.

There’s no GPU to rent and nothing to install. The generation cost per clip depends on your membership tier, so heavier users on higher tiers get more runs. Because the model is hosted, the same graph — the exact GGUF quant, VAEs, and LoRAs — runs identically for everyone; you never have to source model files or match versions.

Best Starting Images and Prompt Tips

The starting frame does most of the heavy lifting. A few things that consistently help:

• Clear framing. Both faces and the point of contact should be visible and unobstructed. Cluttered or ambiguous frames make the model guess, and guesses look off.
• A pose near the action. Because I2V predicts what happens next, a frame that’s already close to the act gives the motion model an easy on-ramp. Start mid-scene, not five seconds early.
• Good lighting and resolution. The workflow renders around a 3:4 frame near 1.2 megapixels, so a sharp, well-lit source upscales and animates far more cleanly than a soft or dark one.
• Keep prompts about action and mood. Say what’s happening and how it feels — pace, intensity, expression. Skip audio descriptions; the audio LoRAs handle the wet sync on their own.

If your source image is thin, generate one first. The T2I workflows on the studio (and the AI Pornstars gallery) are a fast way to produce a clean, on-model still to feed into the Sucking workflow.

Pairing With Other LTX 2.5 Workflows

The Sucking workflow is one member of a family of LTX 2.5 clips that all share the same audio-synced engine, which makes them easy to chain into a longer scene. The most natural pairing is Epic Cumshot: run the Sucking workflow, grab its final frame, and feed that frame into Epic Cumshot as its starting image. Because both are first-to-last-frame I2V graphs, the handoff is seamless — the last frame of one becomes the first frame of the next, and the motion continues without a visible cut.

The other siblings expand the range:

• Smooth Sex — for continuous, rhythmic intercourse motion.
• Flirt & Dialogue — leans on the talking-head side of the model for lip-synced dialogue.
• First-to-Last Frame — the general-purpose bridge for stitching any two frames into a continuous clip, ideal for building multi-shot scenes.

Chained together, these turn a single starting image into a full sequence — foreplay, act, and finish — all with synced audio the whole way through.

Under the Hood: The Exact LTX 2.5 Stack

For the technically curious, here is precisely what runs when you press generate. The base model is LTX-2.5-Distilled-Q8_0.gguf, a Q8_0 GGUF quantization of the 22-billion-parameter distilled LTX 2.5 transformer, loaded through the UnetLoaderGGUF node. Text conditioning comes in via CLIPLoader, and guidance is handled by LTXVDualCFGGuider, which is what makes the dual-conditioning (prompt plus image) behave.

The synced-audio magic is a chain of dedicated nodes. An empty audio latent is created by LTXVEmptyLatentAudio, then joined to the video latent with LTXVConcatAVLatent so both denoise together. After sampling they’re split with LTXVSeparateAVLatent. The picture half is decoded by the video VAE ltx-2.5-video-vae-conv-bf16.safetensors, while the audio half is decoded through LTXVAudioVAEDecode using ltx-2.5-audio-vae-bf16.safetensors — the decoder that produces the actual wet sounds. Because both halves come out of one joint generation, the audio syncs to the head-bob motion by construction.

Sampling runs in two stages. The base pass uses KSamplerSelect with euler_ancestral, driven through SamplerCustomAdvanced with a ManualSigmas schedule. A second refine pass runs through LTXVLatentUpsampler to sharpen detail, and RIFE VFI interpolates frames for smoother playback fps. Image prep is handled by LTXVPreprocess and ImageResizeKJ; the I2V injection itself happens in LTXVImgToVideoInplace. Output resolution is set by a ResolutionSelector at 3:4 near 1.2 megapixels (dimensions kept as multiples of 32), and the finished clip is assembled with CreateVideo and written out by SaveVideo.

The behavior — deepthroat depth, realistic motion, and correct audio cues — is steered by a stack of LoRAs, each applied through LoraLoaderModelOnly at a specific strength:

• LTX-2.3-22b-AV-LoRA-talking-head-v1.safetensors @ 0.88 — the audio-video sync backbone; keeps mouth and throat motion locked to the generated sound.
• ltx23-ultimatedt-NSFW-sulphured_audio_v2_k3nk.safetensors @ 0.6 — the “ultimate deepthroat” LoRA, trained with sulphured audio. It drives both the deepthroat/throat motion and the wet audio cues, and it’s the single most important piece for making this specific act read correctly.
• DR34ML4Y_LT3X_V3.safetensors @ 0.4 — general NSFW realism and motion, keeping skin, bodies, and movement believable.
• LTX2_3_NSFW_furry_concat_v2.safetensors @ 0.3 — extends coverage and coherence so the scene holds together across the full clip.

That combination — a Q8_0-quantized 22B distilled transformer, joint audio-video latents through paired VAEs, a two-stage sampler with upsampling and frame interpolation, and a four-LoRA stack weighted for deepthroat motion and wet audio — is what produces a believable, synced ai blowjob video from a single image in about ten seconds. And because it’s all hosted, you get that exact stack in your browser without ever touching a model file.