Video Generator: T2V, I2V & Audio Explained

Introduction

Most AI video generators present a single generation path. nocensor.ai's AI video generator surfaces four separate decisions — generation mode, audio, duration, and character consistency — and the wrong choice at any of those points produces a result that technically works but misses what the user was going for.

This guide covers what each option actually does under the hood and the specific scenarios where each one performs best.

Text-to-Video: Building a Scene From Pure Imagination

Text-to-video is nocensor.ai's fully open-ended generation path. A user types a prompt, and the Wan 2.2 model synthesizes motion from scratch — no reference image required. The model runs a dual-pass diffusion pipeline: a high-lighting model handles the first half of the denoising steps, and a low-lighting model refines the second half. This split-pass approach produces sharper edges and more consistent lighting than single-model video generation.

Text-to-video works best when the user has a specific visual concept but no existing image to start from. Cinematic establishing shots, fantasy settings, and explicit scenes where the exact composition matters are all strong fits. Because the model starts from pure noise, every element of the frame — environment, subject position, lighting direction — is shaped entirely by the prompt. That freedom is also the constraint: what the prompt doesn't specify, the model fills in on its own.

Prompt construction for text-to-video benefits from camera language alongside scene description. Phrases like "slow zoom," "tracking shot," "handheld," and "cinematic wide angle" give the model a motion vocabulary to work with. Describing the subject's action explicitly rather than just their appearance also produces more intentional motion: "a woman slowly tilting her head back" generates more coherent movement than "a woman with long dark hair" because the first phrase specifies what's happening rather than what's visible.

A negative prompt remains useful even though the Wan 2.2 model operates at a CFG value near 1 rather than the 5–10 range typical of standard diffusion models. Terms like "static," "blurry," "jitter," and "low quality" continue to steer the sampler away from degenerate outputs.

The main limitation of text-to-video is character consistency across multiple clips. Without a reference image anchoring the subject's appearance, slight prompt changes shift facial features and body proportions between generations. Users generating narrative sequences of the same character across multiple scenes should combine text-to-video with a LoRA character model — covered in the final section of this guide.

Image-to-Video: Animating a Still With Realistic Motion

Image-to-video takes an existing still as its starting point and generates plausible motion from it. The Wan 2.2 I2V model reads the pixel content of the source image and uses it as an anchor for both subject appearance and scene composition. The output begins from a state much closer to the uploaded image than any text prompt could specify.

This mode is the right choice when a user already has a satisfactory still — whether from nocensor.ai's image generator or an external source — and wants to bring it into motion without drifting from the established look. Because the model has concrete visual evidence of the subject's appearance, face and body consistency across the clip is significantly higher than in text-to-video.

Image-to-video also produces more naturalistic motion. Since the model extends a real-world pixel structure rather than constructing a synthetic one, subtle details like fabric movement, hair physics, and background parallax tend to track more coherently with scene logic. This makes it particularly effective for animating portraits and close-up compositions where surface detail matters.

The noise augmentation parameter, exposed as an advanced control, defaults to a value calibrated to preserve fidelity while generating visible motion. Increasing it gives the model more room to deviate from the source image — useful when the source image is relatively static and the goal is expressive movement rather than faithful animation. The default value works for most use cases without adjustment.

Image-to-video performs best with stills where the subject's body position suggests a natural next movement — reaching forward, turning, leaning back. Static, fully symmetrical poses sometimes produce subtle oscillation rather than directed motion because the model has fewer motion cues to extend.

For users who generate stills regularly on nocensor.ai, image-to-video is the fastest path to a finished video. Generate the still in the image workflow, confirm the composition and character look, then bring it into image-to-video. The visual quality of the source still sets the upper bound for the video output.

Audio Modes: How Sound Transforms the Final Video

nocensor.ai offers four audio modes — Ambient, Music, Voice, and Moaning — as optional add-ons to any generated video. The right choice depends less on personal preference and more on what the visual content of the clip actually calls for.

Character close-ups and dialogue-adjacent scenes get the most from Voice mode. A clip where a character appears to speak or respond gains a narrative dimension that static visuals alone don't provide. Voice is also the natural audio pair for users who have built companion characters with custom ElevenLabs voices — the audio mode extends that identity into the video output.

Music fits cinematic or wide environmental clips better than voice or ambient sound. These videos are typically longer than character close-ups and benefit from a composed score that gives the piece structural rhythm. The same landscape video with an ambient track versus a scored orchestral backing produces fundamentally different reactions in the viewer — one reads as atmospheric footage, the other as an edited short film. The music generated is specific to each output, so there are no copyright conflicts on social platforms.

Ambient is the practical choice for looping clips and footage where the sound needs to repeat seamlessly. Environmental audio — rain, fire, wind, ocean waves — has no audible beginning or end, which means it loops without the audible cuts that music produces. A two-second character animation played on repeat with ambient sound is barely noticeable; the same clip with a music track announces its loop every cycle.

Moaning is explicit adult audio designed for explicit visual content. Unlike the other modes, it is not an aesthetic choice — it is a content-specific pairing. If the video is sexually explicit, this mode matches the audio to the intent of the video. nocensor.ai is among the few AI video platforms to include this as a native generation option rather than requiring external post-production.

All four modes are optional. Generating video without audio completes faster and costs fewer credits. For iterative workflows — generate, review, refine, generate again — silent output on the early passes keeps the feedback loop fast, and audio can be added to the confirmed final output either during generation or retroactively via the gallery.

Video Duration and Resolution: Matching Settings to Your Goal

The practical framing for duration is not "how long do I want the video" but "how complete does the motion need to be." A two-second clip captures a single gesture; a five-second clip follows an action through its full arc. Users running iterative tests benefit from short duration — the clip either demonstrates that the motion concept works or it doesn't, and spending credits on a five-second clip that misses the mark is waste. Once a prompt has proven itself at short duration, a longer generation captures the full movement.

At nocensor.ai's frame rate of 16 fps, 33 frames run approximately two seconds; 49 frames run around three seconds; 81 frames run just over five seconds. Choosing longer duration doesn't just extend the clip — it changes what the model is being asked to do. The model must maintain temporal consistency across more frames, which increases the probability of subjects or backgrounds drifting between exposures. At 81 frames, minor inconsistencies that would be invisible at 33 frames become readable as motion artifacts. Users who notice that their longer clips have more visual instability than short clips are not encountering a prompt problem — it's a property of the model's temporal consistency capability, and the solution is either a shorter clip or a simpler scene setup.

Resolution should follow a similar tiered approach. Standard resolution is faster and cheaper; HD captures the maximum detail the model can produce. Running first iterations at standard resolution and switching to HD for the confirmed final generation avoids burning credits on exploratory runs at a quality level that doesn't change whether the motion concept is working.

For clips at the maximum 81-frame duration, HD resolution adds meaningful generation time. Users who need both maximum duration and maximum resolution should expect the generation to take proportionally longer than shorter or standard-resolution jobs.

Using LoRA Characters in Video: Keeping Faces Consistent Across Scenes

LoRA models — small fine-tuned neural network weights that specialize the base model toward a specific character or visual style — apply to video generation the same way they apply to image generation. A user working with a custom character LoRA can inject it into the video pipeline to preserve facial and body consistency across clips.

The video pipeline uses LoRA weights that are format-compatible with the Wan 2.2 architecture. When a LoRA is active, the denoising process draws character-specific guidance from the LoRA's learned representations in addition to the text prompt. A character trained on a consistent set of reference images will appear with their established features — face shape, hair color, body type — across different settings, poses, and lighting conditions.

In text-to-video, LoRA injection is the most impactful tool for character consistency, because without it the model has no visual anchor for the subject's appearance. The LoRA acts as a visual specification that sits alongside the text prompt, reducing frame-to-frame variation that would otherwise be driven entirely by the denoising noise seed. Users generating multi-clip sequences of the same character across different scenarios will find LoRA-assisted text-to-video produces significantly more recognizable outputs than unassisted text-to-video.

In image-to-video, LoRA injection and source image anchoring stack. The source image sets the initial pixel state; the LoRA reinforces the character's identity throughout the motion. This combination — a strong source still plus an active character LoRA — is the highest-fidelity path for character-consistent AI video on the platform.

LoRA-assisted generation requires the standard Wan 2.2 model. The fast video model does not support LoRA injection, and selecting it automatically clears any active LoRA from the generation queue. Users who need character consistency should use the standard model.

Conclusion

The text-to-video versus image-to-video decision is the most consequential choice in nocensor.ai's video workflow. Text-to-video gives maximum creative control at the cost of greater prompt responsibility; image-to-video anchors character fidelity at the cost of full compositional freedom. Audio, duration, resolution, and LoRA injection layer on top of that choice to shape the final output's feel, quality, and runtime.

For most users starting out, the recommended path is to generate a strong still in the image workflow first, then bring it into image-to-video at short duration and no audio. Once the motion reads correctly, extending duration and adding audio captures the final version at full quality. Users already working with custom LoRA characters will find that pairing them with image-to-video produces the highest-fidelity character video output the platform currently offers.

All modes described in this guide are available at nocensor.ai's video generator — no content restrictions, no approval queue between the prompt and the output.