Gnothi AI Video Generation

OCDevel

A good AI-generated shot still has to cut with the next one. Learn AI video generation as a production craft, from a first usable clip to coherent scenes and a repeatable process for finishing work. Begin with text-to-video and image-to-video prompts, reference frames, camera movement and targeted revisions. Then work on character and style consistency across shots, storyboards, dialogue and lip-sync, editing, color and sound. The course explains filmmaking terms as they become useful, so experience with a video editor is optional. Worked examples focus on decisions you can repeat and problems you can recognize, from identity drift to flicker, with attention to the cost of usable footage. Advanced topics include ComfyUI, APIs and automated production workflows, along with asset management, quality checks, licensing and client delivery. News and production shortcuts focus on changes that affect the shots you can make or the effort they take. For creators, marketers, indie filmmakers and small studios making ads, social video or narrative work. This show's audio is narrated by AI-generated synthetic voices.

Play

Open in

8 episodes
weekly
English

Episodes 8

Newest Oldest Longest

All8 Unplayed8 In progress0 Played0

October 25 · length unknown
8. Giving the Silent Scene Its Sound
A walkthrough of giving a silent, locked scene a believable soundtrack in Shotcut: spotting, sourcing and licensing effects, syncing them to frames, carrying ambience across cuts, and balancing levels to YouTube's loudness target. It closes with NC AI's new VARCO Sound and a one-shot test for its auto-synced effects. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Spotting the silent scene The courier scene's picture is locked, with nine shots and only an ambience bed under them. The soundtrack goes on in layers: ambience at the bottom, spot effects tied to single actions on top, then the mix and a final check. Muting the timeline exposes every silent action, from two knocks on the bakery door to the parcel landing on a wet step. Each one becomes a row on a spotting list. Offscreen sounds stay out unless they do a job. Finding sounds and checking licences Every file gets a licence check before it reaches the timeline. On Freesound, CC0, CC BY and CC BY-NC behave very differently on a monetized channel. CC BY files need full credit, and Freesound's terms govern the downloads. YouTube Studio's Audio Library sound effects are cleared for use inside YouTube videos with no credit, but they can't be repackaged. ElevenLabs' sound effects generator is free to test. The free plan is non-commercial, so keepers for a monetized release need a paid plan. Generation earns its place for the knock, whose rhythm has to match the courier's hand. Generated takes get listened to and rejected when they come back wrong, such as three knocks or a metallic ring. Syncing on the timeline Effects sit on their own audio tracks, with waveforms visible so each knock's transient can snap to its contact frame. When a sound can't land exactly on its frame, a hair late is safer than early. Ambience runs unbroken under the cuts, and short fades remove clicks. A crossfade into sheltered rain marks the courier's arrival at the lit door, starting slightly before the picture cut. Levels, loudness and listening Gain/Volume keyframes dip the rain at the reaction close-up. Library footsteps come down until the rain sits under them. The knocks peak a decibel or two below zero. YouTube turns uploads down toward −14 LUFS but doesn't raise quiet ones. Normalize: Two Pass applies one fixed gain change toward that target, and the loudness scope confirms the result. Headphones reveal detail and laptop speakers reveal balance. The exported file gets one last listen, and the source and licence records are kept for any later claim. VARCO Sound On October 20, 2026, NC AI released VARCO Sound, which generates synced multi-track effects from text, an image or video. It claims to suggest sounds and timing from the picture. It doesn't know the knock rhythm or the rain dip. The suggested test is to run the knock shot alone through it and compare its knocks against the contact frames. The licence terms on your tier still need checking.
October 18 · length unknown
7. Chaining Shots Into a Finished Scene
Turning the last three stills of the courier scene into moving shots: chaining last frames across continuous cuts, using fresh reference frames or two-ended animation where the cut demands it, and catching drift and lighting faults in a full continuity pass. It closes with what Kling 4.0's longer passes and multiple keyframes might mean for the same hand-off. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Three cuts, three starting frames The courier scene's timeline holds nine shots, three of them still frames. Each still is turned into a moving shot, and the method depends on the cut. When action runs straight across a cut, the new shot starts from the last frame of the old one. This is frame chaining. When the framing changes, it gets a fresh first frame from Gen-4 References. When the shot must also end on a pose that matches an existing clip, it goes to Animate Frames. The arrival at the lit door is chained. The knock that opens the scene uses Animate Frames. The name plate insert gets a fresh References frame. All walkthrough shots, takes and credit figures are illustrative. Exporting and checking the last frame The final frame of the wide walk is found by stepping frame by frame in Shotcut and saved as a PNG with Export Frame, as described in the Shotcut user guide. Motion blur that looks natural in playback becomes a smear in a still. If the last frame is soft, step back a frame or two and trim the walk to match. The still should keep the project's landscape shape, since Runway's input rules cover accepted formats, and a mismatched image gets cropped. Prompting and inspecting the join Gen-4.5 Image to Video gets a motion-and-camera-only prompt. Describing the jacket or collar again invites the model to redraw them. Short takes drift less. Each take is judged across the cut, in this order: identity, collar, bag side, light, screen direction, eyeline, and any lurch at the start of the shot. One take is rejected when the bag switches shoulders. The cause traces to an extra action in the prompt. Drift, re-anchoring and the locked picture In long chains, small errors stack up from link to link. The fix is to re-anchor on a fresh References frame built from the reference sheet. A full continuity pass catches the name plate lit from the wrong side. The cause is missing lighting words in its first-frame prompt. It is repaired without touching any other shot. Logged cost per accepted shot comes in above the earlier estimate, driven by one rejected take and one repair. The picture is now locked, but the only sound is the ambience bed. Kling 4.0 Kling 4.0 adds three things: longer single passes, several keyframes along the timeline, and an Omni Reference system for multi-angle sheets. Its API and HDR output are still listed as coming soon. Free-tier output is watermarked. The suggested next step is a single test against the chained shots, to see whether the courier holds steady.
October 11 · length unknown
6. Keeping the Courier Consistent Across Shots
How to keep one character consistent across separately generated AI video shots, using a reference sheet, a one-page production bible and a targeted repair of a single failed shot. It closes with Vidu Q4's larger reference limit and how to test it on one shot. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Continuity written down once Seven new shots of the courier will each be generated separately, and each is a chance for her to come back different: a darker coat, the bag on the wrong shoulder, light from the wrong side. The answer is a habit. Write the continuity decisions down once and feed them into every shot. The walkthrough is illustrative, and its numbers are examples. Runway splits the work into two stages. Gen-4 References makes a still from up to three tagged reference images. That still then becomes the first frame of a Gen-4.5 Image to Video generation. The video model anchors on one frame, so identity is locked in the still and motion comes second. The chapter also compares the free, Standard and Pro tiers on Runway's plans against the earlier estimate for the seven shots. Reference sheet and one-page bible The reference sheet has front, three-quarter and profile views of the courier, made under flat light on grey so the model doesn't learn a light direction as part of the character. Each view is checked against frames exported from the accepted clips. One profile is rejected because the hair is tied up. Close crops of the collar, bag and parcel label sit alongside the views. The bible covers appearance, exact wardrobe words, props, fixed lighting terms, screen direction and one fixed description line. That line goes only into first-frame prompts. Video prompts describe motion and camera alone. Four shots, then a cut that failed The phone close-up, the sign insert, the reaction close-up and the wide walk each get five checks: face, collar, bag, light and eyeline. On the Shotcut timeline the reaction shot breaks continuity, because the collar lies flat after the phone shot showed it standing up. Tracing and fixing one error Working backward shows the first frame was already wrong, so rerunning the video stage would have wasted credits. The cause was a collar crop cut from a front view where the collar sat low. Swapping in a crop from the three-quarter view fixes that one shot and leaves the others untouched. The bible gets a one-line note about the change. The log puts four accepted shots at roughly 63 credits each. It also shows that one rejected video take cost more than all the rejected stills combined. More references per generation ShengShu's Vidu Q4 preview accepts up to fifteen reference images and three reference audio files in a single setup. That would remove the three-image choice, though the sheet and bible would still be needed. The suggested next step is a single test: regenerate the failed reaction shot in Vidu Q4 and check its preview access and commercial licence terms. HiDream's vivago R1 and Higgsfield's reusable character assets also keep character consistency in one place.
October 4 · length unknown
5. Planning the Scene Before You Generate It
A courier's simple parcel drop becomes a short scene with a want, an obstacle and a change, then a nine-shot list tested as a Shotcut animatic and priced by counting every take. It ends with a zero-credit push-in on a still, keyframed in Shotcut. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website From an action to a scene The two clips on the Shotcut timeline, the courier crouching to set a parcel on the bakery step and rising to turn toward the street, show an action but not yet a scene. Generation is the expensive step and story is the cheap one, so the work comes first on paper. A scene needs a want, an obstacle and a changed situation, and each beat must leave the courier somewhere different. In the illustrative version, a parcel that needs a signature meets a bakery closed for the week, and a lit window next door gives her a new route. The existing crouch and rise clips stay, now with a reason behind them. A general chat assistant can draft the scene. Give it the fixed facts, then ask for less when the first draft gets too busy. Judging the draft is your job: with the sound off, can a viewer see what the courier wants, what blocks her and what changed? A seven-shot list with eyelines The scene becomes a numbered list. Each shot has frame size, one action, screen direction, eyeline, length, and either an existing clip or a first-guess generation method. The knock opens the scene in the same wide angle as the crouch, planned for Animate Frames so it ends on the crouch clip's first frame. The phone close-up, the sign, the rise, the wide walk toward the lit window and the arrival at its door follow. Only the crouch and rise already exist. Writing eyelines down catches mismatches before they are generated: down to the phone, up and left to the sign, right toward the street. Testing it as an animatic in Shotcut Rough stills held at their planned lengths show timing before any credits are spent. Set a default still duration in the Properties panel with Set Default, then adjust each clip by trimming its edges or by splitting with S and ripple deleting with X. Forum threads cover changing the default duration of images and why durations can't be typed once a clip is on the timeline. Add street ambience on an audio track, because silence makes held frames feel long. Watching the animatic exposes two gaps that the paper version hid. The courier never visibly reacts to the closed sign, and the lit window means nothing. The fixes are a reaction close-up and a name plate insert at the door, which brings the list to nine shots. Counting every take Pricing assumes two takes per new shot and three for the two face close-ups, with the sign and name plate planned as stills. At the course's recorded rates, the seven new shots come to roughly 460 credits, or about 66 per accepted shot. Budgeting only the takes you keep would cover less than half of that. A push-in that costs no credits A still can carry an insert if it moves. In Shotcut, the Size, Position & Rotate filter with a zoom keyframe at each end of the clip makes a slow push-in of about eight percent. Keep the move small so the enlarged still stays sharp, set both keyframes to Smooth, and judge it across the cuts on either side. If it still looks pasted in, budget a generated take.
September 27 · length unknown
4. Steering a Shot From Both Ends
## Planning the cut before you generate The accepted crouch clip is three seconds long. The courier crouches and sets the parcel down on the bakery step,… This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Planning the cut before you generate The accepted crouch clip is three seconds long. The courier crouches and sets the parcel down on the bakery step, and the light holds steady from start to end. Watched alone, it works. A film is never one clip watched alone, though. The next thing the viewer sees has to follow from it. She straightens up. She turns toward the street. She walks away. This chapter builds that next shot, and it builds it with one question in mind. What will the cut between these…
September 20 · length unknown
3. Starting From a Still
Why text-to-video takes of the same courier shot could not hold her face, and how supplying a still as the first frame fixes face, wardrobe, composition and light. Covers Gen-4.5 image-to-video setup and costs, two typical failures with the five available repairs, why regenerating proves nothing about cause, and a cheaper way to conform frame rates. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website Why the courier kept changing Text alone could not hold the courier's face. In the first take, a medium shot in a grey rain jacket, her face drifted as she straightened, her hand passed into the parcel, the puddle light jumped and the bakery window grew not-quite-letters; the two later takes fixed the contact by framing close on her hands, and lost who she was. The cause is not a badly written prompt. A sentence names a category, not a person — lengthening it narrows the cloud without ever reaching one specific face. What the image owns and what the prompt owns Supplying a still as the first frame fixes what a prompt can only describe. The image anchors facial structure, hair, wardrobe cut and fabric, composition at the start of the clip, and the direction of key, fill and rim light. The prompt owns action, contact, timing, camera path and duration — none of which exist in a still. Once the picture is fixed, stop describing it: Runway's image-to-video prompting guide advises complementing the still rather than duplicating it, using plain nouns, and phrasing everything affirmatively because negative instructions work poorly. Setting up the shot Gen-4.5 in Image to Video mode requires a paid plan, takes JPEG or PNG stills, outputs 720-line silent video, runs two to ten seconds, and bills twelve credits a second. Bring the still in at 16:9 or the interface will demand a crop — or widen it first with Expand Image, remembering that the new edges are invented. Motion Brush belongs to the deprecated Gen-2; camera moves are named in words. First-and-last-frame work lives in a separate keyframes app. Two failures and five repairs An illustrative take holds the face and wardrobe but pushes a hand through the parcel for four frames, and drifts the shadow direction once the dolly reveals unrendered doorway. The repairs, ranked: change the reference, subtract from the prompt, shorten the action, trim in the edit, or regenerate and hope. Regenerating proves nothing about cause — one clip against one clip is a different seed. Fix the seed and change one variable, or run five to ten under identical settings and count how often the failure appears. Failed generations are not refunded unless the system itself broke. Where the labour moves Text-to-video still wins when you do not yet know what the thing looks like. The image-led path costs you the still: ten to thirty candidates, retouching, sometimes a small reference sheet of angles. Conforming frame rates upstream A frame-rate enhancement endpoint added in mid-September conforms video up to 300 seconds to targets from 24 up to 120, including the fractional broadcast variants, at one credit per two seconds. The next step needs no code: note your clip's actual rate and your destination's.
September 19 · length unknown
2. Cutting Your First Clip Into a Sequence
A hands-on chapter on turning a single generated clip into a short watchable sequence in Shotcut: project setup and Video Mode, trimming two shots into one readable action, adding a title and ambience, licensing the assets you bring in, exporting and checking the result. It closes with Runway's new Premiere Pro and After Effects plugins, which generate straight onto the timeline. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website What an editor is for A single generated clip isn't a film. A cut is a relationship — one image giving way to another — and the meaning lives in the pairing. The work here is a short exported file of two trimmed shots, the courier clip and a still of the step, that someone else can watch and understand. Why Shotcut, and why not Resolve yet Shotcut is free, GPL version three, with no watermark and no export resolution cap; editing at 1920x1080 wants around four cores and eight gigabytes of memory. DaVinci Resolve is the more powerful program and will come up later, but it expects sixteen gigabytes (thirty-two for compositing), a separate graphics card, and splits work across six named workspaces. Builds for Windows, macOS and Linux come from the Shotcut download page. Getting Video Mode right before anything else Name a project folder and keep the media in it — the project file is only a list of instructions pointing at where the files live. Leave Video Mode on Automatic, which reads resolution, frame rate, aspect ratio and colour space off the first item imported. So the generated video goes in first: with an image or music first, Shotcut falls back to 1920x1080 at twenty-five fps, and a mismatch means duplicated or dropped frames and a small stutter in the exact moment you care about. Different aspect ratios get black bars and a notice in Properties. Generated files often carry variable frame rate; accept Shotcut's offer to convert to edit-friendly, which writes a constant-rate copy and leaves your original alone. Two clips, two trims A still of the bakery step, same sixteen-to-nine shape, becomes the companion shot. In and out points by dragging the trim bracket, or by splitting with S and ripple-deleting with X. Cut about a third of a second after the parcel settles and the hand starts to withdraw; run the still about a second and a half. Titles, sound and permission Gen-4.5 clips carry no sound, so the sound bed is yours to add; the Properties panel lists the streams in any file and is the check for clips from other services. A CC0 street ambience from Freesound needs no credit and allows commercial use, and quiet street tone across the cut makes it feel intentional. Titles go on with the Text: Simple filter, or on a second video track. Synchronisation is a separate right: commercial music needs clearance for both composition and master, and the Creative Commons letters matter — NC, ND, SA, and attribution done properly. Keep a record of title, source, rights holder, licence version, proof of purchase, platform clearance code and attribution text. Export, then watch it The YouTube preset gives MP4, progressive scan, libx264, quality-based variable bitrate, stereo AAC at forty-eight thousand. A green tick means the render didn't crash, nothing more: watch the file in a player with the sound up, read its codec information, and reopen it in Shotcut to confirm the chain held. Generation inside the timeline On the eighth of September twenty twenty-six, Runway released plugins for Premiere Pro and After Effects on macOS and Windows, querying Gen-4.5, Kling 3.0 Pro, Veo 3.1 and Aleph 2.0 and inserting results at the playhead. Video-to-video restyling pulls anchor frames from the sequence and returns clips cut to the original's exact length, so the cuts either side don't move. It needs a Premiere subscription and a paid Runway workspace, and nothing here depends on it.
September 10 · length unknown
1. Your First Usable Shot
An illustrative three-take exercise uses Runway Gen-4.5 text-to-video to make and critique a five-second courier shot at sixty credits a take. It teaches prompt structure, the model's cost and silent output, deliberate visual checks and keeping useful production notes; no listener completion is assumed. This episode is narrated by AI-generated synthetic voices. Episode page & show notes Visit website A first shot and a repeatable review The chapter demonstrates a five-second widescreen text-to-video shot in Runway Gen-4.5: an illustrative bike courier in a grey rain jacket leaves a parcel on a bakery step. Gen-4.5 requires a Standard plan or higher; a free account does not unlock this exercise. It bills twelve credits a second, so each five-second take costs sixty credits and the three takes cost 180. The output is a silent 1280x720 MP4, so sound is a separate step in the editor. Runway was chosen over Veo 3.1 or Kling, which generate sound with the picture, because the same web app also holds the image-to-video, first-and-last-frame and reference-image tools that later shots need. In Tool Mode, choose Video and Gen-4.5, leave the image input empty, enter the prompt, set five seconds and check the credit charge before generating. The prompt specifies subject, framing, one action, setting, light and camera movement. Three illustrative takes expose hand contact, identity drift, lettering and lighting problems. Review each take deliberately, change the framing while keeping other instructions as similar as possible, and record why each clip was kept or rejected. Stochastic variation means one pair of clips cannot prove that a prompt change alone caused an improvement. Keep the clips, prompts and settings rather than promise identical regeneration. The text-to-video prompting guide supports describing the desired visible action and camera behavior. The next production questions The same Gen-4.5 model can also take an uploaded image as its first frame, fixing the look a prompt can only describe. The companion news covers MiniMax H3 and Kling VIDEO 3.0 Turbo, which generate sound in the same pass as the picture; when comparing such workflows, check the exact model and its audio controls. This chapter teaches choosing and reviewing one usable shot, not a complete film-production pipeline.

Showing 1–8 of 8 episodes