Turn a sentence — or a photo — into a video clip
A fast-moving split between proprietary frontier models you prompt and pay per second (Sora, Veo, Runway, Kling) and open-source models you self-host on a rented GPU for free (Wan, HunyuanVideo, LTX-Video). Clip length, audio, and resolution all vary sharply — no vendor has solved long-form video yet, so the goal is to understand the trade-offs rather than memorize a winner.
Text-to-video and image-to-video have split into two camps: proprietary frontier models (Sora, Veo, Runway, Kling) that you pay for by the second or by subscription and just prompt, and open-source models (Wan, HunyuanVideo, LTX-Video) that are free to use but need you to rent or own a serious GPU. Clip length, audio support, and cost per second all vary sharply between them — no proprietary model reliably produces long clips yet, and audio is inconsistent across the board. This space changes monthly, so the goal is to learn the category and the trade-offs, not to memorize one vendor as "the" answer. → Quick pick: cinematic quality → Sora; fast with native audio → Veo; pro VFX control → Runway; cheap with good physics → Kling; free on your own GPU → Wan, HunyuanVideo or LTX-Video.
Sora is the proprietary option that consistently delivers cinematic‑grade visuals, though you pay per second for its output.
Open‑source options like Wan, HunyuanVideo, and LTX‑Video run on your own rented GPU at no per‑second fee, but require you to manage the hardware yourself.
Veo offers native audio alongside video generation, while other services may provide limited or inconsistent sound support.
Sora 2 is an AI model that turns text or images into high‑resolution videos with native audio.
Create and download a 6‑second 1080p video from a simple prompt using the Sora 2 web interface.
A concise cinematic prompt that splits into subject, camera movement, lighting and mood layers for Sora 2.
Create a sub‑50‑word prompt that clearly defines those four layers
Sora 2 Storyboard is a feature that lets you define several timed scenes in one prompt.
Produce a 25‑second video composed of three distinct scenes using the storyboard timing option.
A single prompt that tells Sora 2 where to place “cut to” commands for automatic scene transitions.
Generate a clip containing at least two automatic cuts between shots
A workflow that uses Sora 2’s draft editor to tweak one element—lighting, camera angle or pacing—without restarting the whole generation.
Iteratively improve a clip until its mood and framing match your vision
Sora 2 can generate synchronized dialogue and sound effects when the prompt describes audio cues.
Enhance an existing short video with realistic background sounds by specifying them in the prompt.
Sora 2 offers HD/4K export options that produce watermark‑free, commercially usable files.
Generate a 10‑second 4K video and verify it is ready for commercial distribution.
23 outcomes in all — one per recipe below.
Tap the plus icon, choose Add Yours, and upload a clear photo of yourself. Set who can see the cameo (just you, friends, or everyone) and then reference your name in the text prompt to place the avatar in any scene.
You need a ChatGPT Plus or Pro subscription and an invite code, which can be found on the official OpenAI Discord, X, or shared by the community. Install the iOS app or visit sora.com, link your subscription, and enter the code.
Open the draft in the drafts folder, tap Edit to reopen the prompt, and change one element such as lighting, camera angle, or soundtrack. Regenerate the clip; this iterative tweak saves generation credits and aligns the output with your vision.
Start with the video's purpose (e.g., "TV ad for a jewelry brand"), then add explicit camera directions (like "cinematic close-up"), lighting details (such as "soft natural daylight"), and any material or texture cues. Keeping the prompt under 50 words helps Sora focus on the key visual cues.
Only have a short idea and images
Uploading OpenAI’s official Sora 2 documentation into ChatGPT trains the LLM on the exact terminology, formatting, and structural expectations of the video model. ChatGPT then acts as a prompt architect, converting brief concepts and reference images into highly detailed, timecoded scripts that specify camera movement, lighting, pacing, and mood. This bridges the gap between vague text inputs and the model’s actual rendering logic.
Awkward dismounts or static cameras in AI video
AI video models rarely produce perfect results on the first generation due to complex physics and motion synthesis. By treating ChatGPT as a cinematographer that expands your directorial notes, you can isolate specific flaws like awkward dismounts or static cameras and request precise shot-level fixes. Regenerating after each refinement cycle or simply re-running the prompt often yields the final polished clip.
Text prompts keep changing characters and settings
Text-only prompts often cause AI video models to hallucinate inconsistent characters, props, or environments. Uploading reference images alongside your prompt gives the model concrete visual anchors, allowing it to extract accurate details for lighting, texture, and composition. This dramatically reduces random variations and keeps the generated video aligned with your original vision.
Not sure what style my video should have
Telling Sora the intended use (e.g., TV ad, YouTube explainer) gives the model a clear creative boundary, focusing its attention on relevant visual and tonal cues. This prevents vague outputs and aligns the generation with professional standards. The model uses the stated purpose to infer pacing, polish level, and compositional priorities.
Unsure what view the AI will use
Sora relies heavily on explicit camera direction to determine composition. Describing angles (close-up, wide shot, tracking shot) directly dictates the viewer's visual experience and emotional engagement, replacing guesswork. The model maps these directional cues to specific framing and depth-of-field behaviors.
Lighting feels flat in AI video prompts
Lighting is the primary driver of perceived realism and cinematic quality in AI video. Explicitly stating light sources, times of day, or quality (soft, harsh, directional) gives Sora a concrete reference for shading, reflections, and atmosphere. This reduces the model's reliance on default, often flat, lighting setups.
5‑second clip limit stops long scenes
Sora's 5-second limit and predictive nature struggle with complex, multi-step actions in a single prompt. The Storyboard feature chains sequential prompts, maintaining character and environment consistency while maximizing action within each clip. This turns isolated generations into a continuous narrative flow.
Need a custom video transition
The Blend feature merges two separate videos using adjustable timing curves. The Transition mode acts like a manual keyframe system, letting users control exactly when and how one clip morphs into another. This bypasses rigid preset transitions and creates unique, user-directed morphs.
Need a looping background video
Sora's Loop feature extends a clip by blending its start and end frames. It works best on repetitive, non-directional motion because forward-moving objects create jarring discontinuities when merged. This makes it ideal for generating ambient backgrounds or motion graphics.
Videos look flat and unrealistic
Sora struggles with complex physics but excels at rendering surface properties. Explicitly naming materials grounds the scene and tricks the model into generating convincing light interactions and weight. This anchors the video in physical reality, making abstract or simple motions feel tangible.
My AI videos turn out random and low quality
Sora 2 responds best to structured, cinematic prompts rather than vague descriptions. Using templates like scene plus camera movement plus style gives the model clear directional cues for lighting, framing, and pacing. This structure mimics film directing, resulting in predictable, high-quality outputs.
AI video keeps showing generic versions of my product
AI video models often hallucinate generic versions of specific objects or places. Uploading a reference photo before generation anchors the model to the exact visual details of your subject. This dramatically improves accuracy for client work, product demos, and branded content.
Need a single video with built‑in scene changes
Instead of generating separate clips and editing them together, you can instruct the AI to produce a sequence with internal cuts. Using the phrase cut to between scene descriptions tells the model to transition between shots while maintaining continuity. This saves time and creates dynamic storytelling clips.
When AI video glitches on splashing water or flowing fabric
AI video models still struggle with complex physics like splashing liquids, flowing fabric, or gravity-defying actions. Prompting for simple, static, or gently moving subjects prevents weird artifacts and unnatural movement. This strategy ensures reliable, watchable outputs without wasting generations.
Video is close but the mood’s off
First-generation AI videos are often 80 percent correct but miss the exact mood or framing. Instead of rewriting the entire prompt, tweak one variable at a time (lighting, camera angle, pacing) and regenerate. This iterative approach builds on successful outputs and rapidly converges on the desired result.
Can't access the limited rollout app
Sora 2 is currently in a phased rollout that requires a ChatGPT Plus or Pro subscription plus an invite code. Users can obtain codes by monitoring the official OpenAI Discord, searching X, or checking community shares as OpenAI gradually expands access. This gatekeeping manages server load while gathering early creator feedback.
Want your own face and voice in AI videos
Sora 2 lets you train a temporary digital avatar of yourself using a few seconds of video and voice. You grant permission, set visibility (private, friends, or public), and then reference your name in prompts to insert yourself into any scene. The model maps your facial features and voice onto the generated character, enabling personalized video creation.
Need a video that follows my exact scene, actions and music
Sora 2 generates video and synchronized sound from a single text prompt. By explicitly describing the scene, action, camera movement, and background music in one prompt, the model aligns the visuals and audio track. This unified prompting approach reduces the need for separate editing tools and leverages the model’s physics and audio synthesis capabilities.
Want broadcast‑quality video and a bit more length
Pro subscribers unlock a higher-tier model that prioritizes visual quality and longer clip duration over speed. The interface exposes advanced settings for resolution (high or standard) and duration (up to 15 seconds), giving creators more control over the final output. This tier is designed for users who need broadcast-ready quality or longer narrative clips.
Need to adjust a generated video's music, angle or style
After generation, videos land in the drafts folder where you can reopen the prompt box to make targeted edits. Instead of regenerating from scratch, you can tweak specific elements like the soundtrack, camera angle, or visual style by modifying the prompt. This iterative workflow saves time and helps fine-tune the output to match your vision.
Want your own face in AI‑generated videos
Sora 2 lets you upload a photo to create a digital cameo that the model can animate in new scenes. The app processes the image into a recognizable avatar, which you can then insert directly into text prompts to control who appears in the generated video.
Instead of starting from scratch, Sora 2's main feed includes a Pick a Mood filter that lets you type specific vibes or genres to surface existing AI videos. Swiping left on any video reveals alternative prompts or variations used to recreate it, providing instant inspiration for your own generations.
My AI video isn’t perfect
Initial generations often need tweaking. Sora 2 saves outputs to a drafts folder where you can reopen the prompt, modify details, and regenerate without losing the original concept. This iterative loop improves alignment with your vision and saves generation credits.
The same set on /recipes, filtered by tool and role.
Shows beginners how to use Sora by selecting appropriate video styles, managing the 5‑second clip limit for smooth multi‑step flows, and adding realistic lighting with shadows and highlights.
Shows beginners how to use Sora 2 to generate exact product videos, create multi‑shot clips with automatic cuts, and fine‑tune the mood of AI‑generated footage
Shows how to create synced AI video clips, add personalized digital cameos, and produce higher‑detail videos up to 15 seconds
Shows how to create and edit AI‑generated videos with Sora 2, add a personal cameo avatar, and generate video ideas matching a desired mood
Upload a clear, well‑lit reference photo of the exact product or place and attach it to the reference image field before generating. The model uses that image as an anchor, reducing generic hallucinations and matching the visual details you need.
Use multi‑cut prompting by describing each shot and inserting “cut to” between them (e.g., "person walks forward, cut to close‑up of hand, cut to wide room view"). The model then creates internal transitions, giving you a single clip with built‑in scene changes.
Complex physics like liquid splashing or heavy wind often cause artifacts. Follow the physics‑safe strategy: replace those actions with simple, static or gently moving subjects (e.g., a bottle on a table instead of pouring wine) and use gentle camera moves to keep the output clean.
Apply the remix‑and‑iterate workflow: identify one element that’s off (like lighting or camera angle), change only that part of the prompt, and regenerate. Repeat tweaking single variables until the video matches your desired mood and framing.
Yes, Sora 2 lets you create a cameo avatar by uploading a photo (or short video) of yourself, setting visibility permissions, and then referencing your name in the text prompt. The model animates that digital version within any scene you describe.
Upload a clear, well‑lit reference photo of the exact product and attach it to the reference image field before you generate. The model uses that image as an anchor, so the resulting clip matches the real product’s details instead of hallucinating a generic version.
Use a cinematic formula that lists the scene, camera movement, and style in under 50 words. For example, start with the camera direction (like “wide tracking shot”), then describe the subject’s action, followed by the visual mood or lighting.
Follow the physics‑safe strategy: choose simple, static or gently moving subjects instead of complex actions like pouring liquids or heavy wind. Specify gentle motions such as a slow rotation or soft drape to keep the output clean.
Yes, use the draft editor: open the generated clip in your drafts folder, tap Edit, modify only the part of the prompt that describes lighting or camera perspective, and regenerate. The rest of the video stays the same.
Use multi‑cut prompting by writing a single prompt that includes “cut to” between scene descriptions. List each shot’s subject and action in order, and Sora 2 will generate internal transitions for you.
Upload a clear, well‑lit reference photo and attach it in the reference image field before you generate. The model uses that image as an anchor, so the resulting clip matches the exact visual details of your subject.
Use structured cinematic formulas such as “scene + camera movement + style” and keep the prompt under 50 words. Fill in specific details like location, camera action, and mood; this gives the model clear directional cues for lighting, framing, and pacing.
Follow the physics‑safe strategy: replace complex actions (like liquid splashing or heavy wind) with simple, static or gently moving subjects. Specify gentle camera moves such as slow rotation to keep motion natural and prevent unnatural artifacts.
Yes, use multi‑cut prompting by writing “subject + action + cut to + next subject + action…”. The phrase “cut to” tells the model to insert internal transitions, producing a single clip with built-in scene changes.
Apply the remix and iterate workflow: identify one element that’s off (e.g., lighting or camera angle), change only that part of the prompt, and regenerate. Repeat tweaking single variables until the clip matches your desired mood and framing.
Upload a personal photo in Sora 2’s cameo feature; the app creates a digital avatar from it. Then reference your name or the cameo in the text prompt, set visibility permissions, and generate the video with you appearing in the scene.
Use a structured, cinematic formula that includes the scene, camera movement, and style, keeping it under 50 words. This gives clear direction for lighting, framing, and pacing, which helps the model produce high‑quality clips.
Upload a clear, well‑lit reference photo in the reference image field before generating. The model anchors its visuals to that image, reducing hallucinations and keeping the product appearance consistent throughout the clip.
Yes, use multi‑cut prompting by writing “cut to” between scene descriptions. This tells Sora 2 to insert internal transitions, producing a single video with built‑in cuts and continuous storytelling.
Complex physics such as liquid splashes often cause artifacts. Prompt for simple, static or gently moving subjects—like a bottle on a table instead of pouring wine—to keep the motion natural and artifact‑free.
+ 14 more in the library.
ChatGPT
CameoPick a Mooddraft editorStoryboardBlend toolLoop toolMulti‑Cut PromptingPhysics‑Safe PromptingCinematic Prompt FormulasReference Image Integrationreference image fielddrafts folderPick a Moodcameo permissionsinvite codeChatGPT PlusMulti-Cut PromptingPhysics‑Safe PromptingBlend featureLoop toolreference image fielddrafts folderPick a Mood filterCameo permissionsinvite codeChatGPT Plus or Pro subscriptionBlend featureLoop buttonStoryboard buttoncut to phrasephysics‑safe prompting strategySora 2 appreference image fieldmulti-cut promptingphysics-safe promptingcameo avatarPick a Mood feeddraft editorinvite codeChatGPT PlusPro subscriptionBlend featureLoop toolStoryboard buttonreference image fieldcut todraft editorinvite codeChatGPT Pluscameo permissionsPick a Mood filterStoryboard featureBlend featureLoop toolhigh resolution settingphysics‑safe promptingCinematic Prompt FormulasReference Image IntegrationMulti-Cut PromptingPhysics‑Safe Prompting StrategyRemix and Iterate WorkflowPick a Mood feeddraft editorinvite codeCameo avatarBlend toolLoop toolStoryboard featurereference image fielddrafts folderPick a Mood filterCameoinvite codeBlend featureLoop toolStoryboard buttoncut toPhysics-Safe Prompting StrategyAsk, share, or report — over on the Heidelberg AI community forum.