/director

AI Video Models (2026) — Selection Guide

A film director in 1925 chose a film stock, a lens set, and a lab. A director in 2026 chooses a model. The model is your stock, your lens, your camera operator, and — increasingly — your sound recordist all at once, and unlike celluloid it changes its capabilities every six to ten weeks. This chapter is the catalog and the decision logic: which generative video model to reach for, why, and how to wire them together into a controllable pipeline rather than a slot machine.

The single most important control surface in this entire chapter — the one Yuval should internalize before anything else — is the keyframe, specifically first-frame and last-frame conditioning. Text-to-video is a wish. Keyframed image-to-video is direction. Almost every craft principle in the sibling chapters (the precise camera move in 07-camera-angles-and-movement.md, the composition you locked in 06-shots-framing-composition.md, the look you graded in 08-lenses-lighting-color.md) only becomes executable when you can pin the start and end of a shot to images you control. So this chapter foregrounds, for every model, the answer to one question: can I give it a start frame, an end frame, or both?

Currency warning. This is the most perishable chapter in the bible. Every fact below was verified against live sources in June 2026 and is marked [verified Jun 2026] or, where I'm extrapolating from a slightly older source, [may have changed]. Model versions, prices, and clip limits move fast. Re-verify pricing before you quote a client a budget. Sources are listed at the end.

The mental model: three jobs, not one tool

Before the catalog, the framing that prevents 90% of wasted credits. A generative video model is really being asked to do up to three separable jobs, and no single model is best at all three:

1. Synthesis — invent pixels that look like a real (or stylized) world. This is the "photoreal vs. plastic" axis. Veo, Sora, Kling, Seedance lead here.
2. Motion / physics — make objects move with believable weight, momentum, and continuity across frames. This is where most models fail (the "morphing hands," the "spaghetti gait"). Kling and Hunyuan are physics-forward; many "pretty" models are motion-weak.
3. Control — obey your constraints: a start image, an end image, a camera path, a character's identity, a line of dialogue. This is the director's axis, and it's where Runway, Luma, Pika, and MiniMax differentiate.

A "hero shot" film is usually built by using a strong-synthesis model for the look, a strong-control model for the precise motion, and a specialist for lip-sync or upscaling. You are an assembler, not a button-presser.

The big comparison table

All figures [verified Jun 2026] unless flagged. "FF/LF" = first-frame / last-frame conditioning. Prices are rough, per-second or per-clip, and vary by tier and reseller; treat them as order-of-magnitude.

Model-by-model: the why behind the spec

Google Veo 3.1 — the audio-native realism flagship

Veo's headline differentiator is native, synchronized audio generated with the video — dialogue lip-synced to the character, sound effects matched to on-screen action, and ambient soundscape, all at 48 kHz [verified Jun 2026]. Most competitors bolt audio on or leave it to you; Veo treats sound as part of the diffusion target. It outputs 8-second clips at up to 4K (the 4K tier rolled out January 2026), supports portrait and landscape, image-to-video, up to three reference images, video extension, and — critically for us — explicit first-frame and last-frame conditioning [verified Jun 2026].

The 8-second ceiling is the catch. Veo is a shot machine, not a scene machine; you build longer pieces by extending or by cutting between generated shots in an editor (see 10-editing-theory.md). Pricing on the Gemini API runs roughly $0.40/s at 1080p, $0.60/s at 4K, with a Fast tier near $0.15/s and a cheaper Lite tier; consumer access is via Flow credits inside Google AI Pro ($19.99/mo) and Ultra ($249.99/mo) [verified Jun 2026].

→ AI APPLICATION. Veo is your default when a shot needs integrated diegetic sound or 4K finish. Generate a start frame and an end frame in an image model (Imagen/Nano Banana/Flux), feed both into Veo's first+last-frame mode, and write the camera move as plain cinematographer's language in the prompt ("slow 35mm dolly-in, shallow depth of field, the actress turns to camera on the last beat"). Because audio is native, put the line of dialogue and the SFX in the same prompt rather than scoring in post.

OpenAI Sora 2 — the long-coherence storyteller

Sora 2's edge is temporal coherence and world consistency over longer durations. Access is API-only as of this writing — the consumer Sora web app and all ChatGPT consumer access to Sora were discontinued on April 26, 2026; only the developer API remains until the September 24, 2026 sunset [verified Jun 2026]. Per OpenAI's API docs, both sora-2 and sora-2-pro generate 16- and 20-second clips, and via the extend mechanism (up to six 20 s continuations) reach 120 seconds while keeping a character and environment stable [verified Jun 2026]. Resellers report the selectable fixed durations as 4/8/12 s on standard Sora 2 and 10/15/25 s on Sora 2 Pro — the 25-second option is a Pro-tier figure, not a standard Sora 2 one. Audio is generated in sync (dialogue, SFX, music). Sora 2 Pro unlocks true 1080p (1920×1080 / 1080×1920); standard Sora 2 caps at 720p. You can supply an input reference image as the opening frame, and reuse uploaded character assets for consistency [verified Jun 2026].

The control gap: Sora has no native last-frame target the way Veo/Kling/Luma do. Its control philosophy is storyboard + extend + cameo/character upload rather than first+last interpolation. API pricing: ~$0.10/s standard (720p), $0.30–0.50/s Pro depending on resolution, with a 50% batch discount [verified Jun 2026]. Important: the Sora 2 / 2 Pro API is scheduled to sunset on September 24, 2026 [verified Jun 2026] — assume a Sora 3 transition is coming; [may have changed] by the time you read this.

→ AI APPLICATION. Reach for Sora when you need a continuous performance — a character walking and talking through a space for 15–20 s without a cut, where Veo's 8-second wall would force an edit. Lock the opening frame with input_reference, upload a character asset for identity, then use the storyboard panel to steer beats. Because there's no last-frame pin, design the shot so the ending pose is implied by the action, not by a hard target.

Kuaishou Kling 3.0 — the motion and lip-sync king

If a shot lives or dies on motion realism — a fight, a dance, a sprint, cloth and hair physics, a face that actually emotes — Kling is the strongest dedicated answer in 2026. Kling 3.0 (launched Feb 4, 2026; Motion Control launch March 2026) generates 3–15 second clips at up to native 4K, 60 fps (up from Kling 2.6's 48 fps), with native audio (dialogue, SFX, music) and physics simulation [verified Jun 2026]. The widely-used Kling 2.6 (10 s, 1080p, 48 fps, native lip-sync) is still common because resellers lag [verified Jun 2026].

Two director-grade features: start-frame and end-frame control that chains into longer continuous scenes, and Kling 3.0 Motion Control, which extracts a motion sequence from a reference video (3–30 s) and transfers it onto a static image or new clip — full-body mocap, hand gestures, facial expressions, choreography [verified Jun 2026]. Note that the Motion Control mocap-transfer path outputs at up to 720p, distinct from the 4K base text-to-video/image-to-video pipeline. That is the closest thing to performance capture in a consumer text-to-video tool.

→ AI APPLICATION. For an action or dance hero shot: shoot (or generate) a reference motion clip, generate your stylized character as a still, then use Kling Motion Control to drive the still with the reference motion. For dialogue, Kling's native lip-sync means you can drive mouth shapes from a script line directly. Use start+end frames to chain two Kling clips into one continuous camera move across a cut point.

Runway Gen-4 / Gen-4 Turbo — the control-and-edit workbench

Runway's bet is post-generation control, not just generation. Gen-4 and the cheaper Gen-4 Turbo support image-to-video with References (feed character/style/location images for consistency), camera control, and character performance [verified Jun 2026]. The standout is Aleph — an in-video editing model: change weather, relight, remove or insert objects, restyle, all on already-generated footage. Pricing: Gen-4 Turbo = 5 credits/s, Gen-4 = 12 credits/s; plans run Standard ($12/mo, 625 cr) → Max ($76/mo, 9,500 cr) [verified Jun 2026]. Note: Gen-4 Aleph (v1) is deprecated, sunsetting July 30, 2026; migrate to Aleph 2.0 [verified Jun 2026] — [may have changed].

Runway's first-frame support is solid; its "last-frame" story is handled through References and Aleph editing rather than a pure interpolation slider — slightly less literal than Luma/Pika but more editable after the fact.

→ AI APPLICATION. Use Runway when the workflow is iterative and fixable: generate a rough shot, then Aleph-edit it toward the storyboard (relight to match 08-lenses-lighting-color.md, swap a prop, change the time of day) instead of re-rolling and praying. References is your character-consistency lever across a multi-shot sequence.

Luma Ray3 / Ray3.14 — the keyframe choreographer

Luma is the keyframe specialist and the one most aligned with Yuval's first/last-frame priority. Ray3 accepts start and end keyframes and generates the transition between them; the Multi-Keyframe feature — which landed with Ray3.2 — lets you set up to 16 keyframes inside a single clip, directing what changes, what holds, and how the shot resolves [verified Jun 2026]. Ray3 also delivers native 16-bit HDR (exportable as 16-bit EXR for pro grading pipelines) and Ray3 Modify brings start/end-frame control into a video-to-video workflow for performance editing. Ray3.14 added native 1080p, 4x speed, better adherence, and ~3x lower cost — but traded away character reference (and native audio) for that speed/cost win, so reach back to an earlier Ray3 sub-version when you need identity locking [verified Jun 2026]. Audio is not the model's focus — treat Luma as a picture tool and score separately.

→ AI APPLICATION. Luma is the precision tool for exactly the move you storyboarded. Generate 3–16 keyframe stills (opening composition, mid-beat, ending pose), drop them on Luma's timeline in order, and the model interpolates a single continuous shot that hits every mark. This is the cleanest way to get "a precise camera move that ends on a specific pose" — see the decision tree below.

MiniMax Hailuo 02 / 2.3 — the cost-efficient identity-keeper

Hailuo's reputation is best-in-class physics-per-dollar and a strong subject-reference mode: feed a reference face/character image and it maintains that identity across the clip [verified Jun 2026]. It outputs native 1080p (Pro; standard offers 768p/512p), 6–10 second clips, with image-to-video and first-frame support; newer 2.3 adds audio [verified Jun 2026]. The value proposition is volume: it's one of the cheapest credible models, which makes it ideal for exploration and consistent-character batch work.

→ AI APPLICATION. When you need the same character across twenty shots on a budget, lock a reference image and run Hailuo for identity continuity, then promote your hero shots to Veo/Kling for the finish. Use it as the cheap "draft pass" of a two-tier pipeline.

Pika 2.2 / 2.5 — the effects and transition playground

Pika is the creative-effects model. Pikaframes is its keyframe system: define a start and end image and it interpolates between them, chaining up to 5 keyframes for up to ~25 seconds [verified Jun 2026]. Its signature toys — Pikaffects (melt, explode, inflate, "cake-ify"), Pikadditions (insert an AI element into real footage), Pikaswaps, and Pikaformances (audio-driven performance) — make it the go-to for stylized transitions and viral effect shots rather than photoreal drama. Output is 1080p, up to 10 s per base gen [verified Jun 2026].

→ AI APPLICATION. Use Pikaframes for transformation shots — a product morphing, a face aging, an impossible match-cut between two stills you generated. Define both ends precisely and let Pika handle the impossible middle. For straight realism, prefer Veo/Kling.

ByteDance Seedance 2.0 / 1.0 Pro — native multi-shot in one prompt

Seedance's unique trick is native multi-shot generation: a single prompt yields a clip with 2–3 shot changes (long → medium → close-up) and natural transitions [verified Jun 2026]. 1.0 Pro does 1080p, 10 s; 2.0 extends to 15 s with multi-shot storytelling and audio. Combined with ByteDance's scaling infra (the merged Pixeldance + Seaweed efforts), it's a cost-leader for micro-scenes.

→ AI APPLICATION. When a beat needs internal coverage (establishing → reaction → detail) but you don't want to direct three separate generations, prompt Seedance for the whole mini-scene. Use it to rough out coverage, then re-shoot the keepers in a higher-control model.

Open source — Wan 2.2, LTX-Video 13B, HunyuanVideo 1.5

For local, private, unlimited-iteration work on Yuval's own GPUs (relevant given the local-AI-stack memory): Wan 2.2 is the first open-source MoE video model with sub-models for text-to-video, image-to-video, character animation, and audio — best open photoreal humans [verified Jun 2026]. LTX-Video 13B is the speed/accessibility champion and the only one that fits comfortably on a 16 GB card — ideal for volume iteration [verified Jun 2026]. HunyuanVideo 1.5 has the best open physics (water, smoke, cloth) but is VRAM-hungry: 60 GB+ for full precision, ~24 GB quantized [verified Jun 2026]. All three support image-to-video and, via ControlNet-style conditioning, first-frame (and varying degrees of last-frame/keyframe) control. NVIDIA is the path of least pain; ROCm support was still unstable as of April 2026 [may have changed].

→ AI APPLICATION. On the 4090 in the local stack, LTX-Video is your fast draft loop and Wan 2.2 your photoreal-human pass; reserve cloud Veo/Kling for finals. Open models also give you programmatic keyframe control (scripted batches, frame conditioning) that hosted UIs don't expose — the right tool when you're building a skill, not clicking a website.

Specialists — lip-sync and upscaling

Two jobs the generalists do poorly. Lip-sync / talking avatars: HeyGen (Avatar IV) and Hedra turn a photo + audio into a synced talking head; independent tests rate Hedra's lip-sync ~9/10, though Hedra video export peaks at 720p (4K via its "Super Genius" upscaler) [verified Jun 2026]. Upscaling / de-artifacting: Topaz — and HeyGen's upscaler powered by Topaz Starlight Precise 2.5 — is purpose-tuned for generative footage, reconstructing texture and fixing the softness/aliasing that naive upscalers amplify [verified Jun 2026].

→ AI APPLICATION. Pipeline order matters: generate the picture (Veo/Kling), do any restyle/edit (Runway Aleph), then upscale last (Topaz/Starlight) so you're not amplifying artifacts you'll later edit. For a presenter saying exact lines, drive lip-sync in Hedra/HeyGen rather than fighting a text-to-video model's mouth.

The decision tree

Read top-down; stop at the first match.

1. Need a presenter speaking exact scripted lines (talking head)? → Hedra (best sync) or HeyGen (avatar ecosystem). Upscale with Topaz after.
2. Need integrated dialogue + SFX + a 4K hero shot? → Veo 3.1 (first+last frame, native 48 kHz audio, 4K).
3. Need a precise camera move that ends on a specific pose? → Luma Ray3 multi-keyframe (up to 16), or Veo 3.1 first+last, or Pika Pikaframes for stylized morphs. Pick Luma for realism, Pika for effects.
4. Need heavy motion / action / dance / mocap transfer? → Kling 3.0 (Kling 3.0 Motion Control, best physics, native lip-sync).
5. Need a single continuous 15–20 s performance, no cut? → Sora 2 Pro (extend to 120 s, world coherence).
6. Need internal coverage (multi-shot mini-scene) from one prompt? → Seedance 2.0.
7. Need to fix a shot after generating (relight, swap prop, restyle)? → Runway Gen-4 + Aleph 2.0.
8. Need the same character across many cheap shots? → MiniMax Hailuo subject-reference.
9. Need local / private / unlimited iteration? → LTX-Video (16 GB, fast) → Wan 2.2 (photoreal humans) → HunyuanVideo (physics, big VRAM).

→ AI APPLICATION (chapter-level): the keyframe-bridge pipeline

The recommended generation strategy — and the one this bible's skill should default to — is a two-stage, keyframe-bridge pipeline that separates composition from motion:

Stage 1 — Author keyframes in an image model. Use a high-control still-image model (Nano Banana 2 / Gemini image, Flux, Imagen, Midjourney) to generate your first frame (the opening composition, lit and graded to your 06/08 specs) and your last frame (the ending pose/framing). For complex moves, generate intermediate keyframes too. You now control composition, character identity, palette, and pose deterministically — none of it left to the video model's imagination.

Stage 2 — Interpolate with a video model that accepts those keyframes. Hand the frames to a model whose control surface matches the shot:

Stage 3 — Finish. Optional Aleph edits for continuity, lip-sync specialist for any to-camera dialogue, then Topaz/Starlight upscale last.

The reason this works is the same reason storyboards work in traditional film: the expensive, uncontrollable step (motion synthesis) is bracketed by two cheap, fully-controllable steps (the still frames). You are not asking the model to invent your shot; you are asking it to connect two images you already approved. That is the difference between directing and gambling — and as of mid-2026, Veo 3.1, Luma Ray3, Pika, Kling 3.0, Seedance 2.0, and the open Wan/LTX models all support it natively. Sora is the notable exception (no literal last-frame), which is precisely why you reach for it only when you want continuity over a long take rather than a pinned ending.

Sources

• Veo 3.1 — Google DeepMind
• Generate videos with Veo 3.1 — Gemini API docs
• Gemini API pricing
• Veo 3.1 Pricing Guide 2026 — AI Free API
• Sora 2 — OpenAI
• Video generation with Sora — OpenAI API docs
• Sora 2 Model — OpenAI platform
• Sora 2 API Pricing & Sunset Guide — CostGoat
• Sora 2 release notes — OpenAI Help
• Kling 3.0 Motion Control Release — KlingAI
• Kling 2.6 — Media.io
• Kling video models essentials — Scenario
• Runway API pricing & costs
• Runway pricing
• Gen-4 Aleph pricing & specs — CloudPrice
• Ray3 — Luma
• Ray3 Modify (start/end frame) — Luma
• Luma Ray3 HDR — CineD
• MiniMax Hailuo 02 — MiniMax
• Hailuo 02 image-to-video — fal.ai
• Pika v2.2 Pikaframes — fal.ai
• Pika 2.2 release — AIbase
• Seedance 1.0 — ByteDance Seed
• Seedance 2.0 — Artlist
• Open-source AI video comparison 2026 — AI Magicx
• Local AI video: Wan 2.2, LTX, Hunyuan — Local AI Master
• HeyGen vs Hedra lip-sync 2026 — lipsync.com
• HeyGen AI Upscaler (Topaz Starlight)