AI short films: consistent characters and sets

Question: how close is AI to generating a multi-scene short film, taking as the benchmark Evillica's analog-horror YouTube Shorts (a recurring actress, a recurring apartment, 15 to 23 cuts in under 40 seconds)? And are there techniques by which characters and scenes can be generated deterministically and combined into a film, so that episode two has the same person in the same room as episode one?

Short answer: for a series, no, and it is a dead end for generation rather than a gap the next model release closes. A single convincing short is achievable today with a person curating several takes per shot. But the two things a series depends on, the same actor and the same location in every shot, are exactly what no 2026 video model can lock: reference images hold a face "recognisably the same" while costume, props and the room itself drift, and nothing hosted is repeatable. The only architecture that makes a multi-scene film both consistent and reproducible builds the character and the set as 3D assets, renders every shot's structure conventionally, and uses AI only as a finishing pass. At that point AI is not generating the film; a 3D production is, with a generative filter on top.

This is a snapshot of the landscape in September 2026. Figures come from vendor documentation, model cards, papers and practitioner write-ups accessed that month; section 10 lists what could not be verified against a primary source.

1. The three reference shorts are live-action, not AI: one actress, one real apartment, practical creatures, a post-production face warp, a VHS grade. The bar is therefore "fake a cheap camcorder short with the same actor and flat across episodes", which is harder for generative video than "make an AI-looking short".
2. Per-shot generation of a convincing single short in this look is possible today with a human curating 3 to 6 takes per shot, for roughly $40 to $200 hosted or $10 to $25 self-hosted. Keep rates are about 25 %; character drift shows after about ten shots; no hosted model repeats a clip.
3. Characters can be locked "recognisably the same" by reference images (every 2026 model has them, about 0.5 face similarity on the one published cross-vendor table) and locked exactly only by a trained LoRA or by owning the character as a 3D asset. A deliberately wrong face survives only the latter two: every zero-shot identity mechanism is trained to make faces look right.
4. Scenes cannot be locked by any video model: reference images condition on appearance, not geometry, and the room is re-invented once the camera turns past roughly a quarter turn. The door stays put only in a 3D set, a Gaussian splat, or a depth-locked control pass driven by one.
5. Determinism exists at exactly one place in the whole landscape: a 3D scene rendered with a fixed seed. Open-weight diffusion is bit-repeatable under a strict recipe on pinned hardware; every hosted API says "seed improves similarity" at best. Hosted models were also retired three times in 2026 (Sora, Gen-3, Gen-4 Aleph) with weeks of notice.
6. The technique that makes a multi-scene film both deterministic and combinable is to build the character and the set as 3D assets, render every shot's structure in a path tracer such as Blender's Cycles, let an open-weight model add photographic surface under depth and identity lock, add narration and a single VHS pass last, and treat the episode as a reproducible build with a lock file. Hosted models are inserts, not the spine.

Downloaded and stepped frame by frame (608×1080, 30 fps, AV1). What is on screen:

• One actress plays the daughter in all three: long dark hair, red and blue plaid flannel over a black tee, leopard-print phone case, identical across the three uploads.
• One real apartment recurs: galley kitchen with a white fridge, white six-panel doors, vertical blinds, grey plank floor, a round table with a cream cloth, a bedroom with grey bedding and a mirrored closet.
• Practical creatures with a post pass: "Mom" in a red bathrobe with a digitally warped face (stretched-open mouth, blank white eyes; a grinning variant through the window); a black morph-suit figure with white eyes and teeth under the bed and in the mirror; a rabbit-eared silhouette in a doorway. The audience calls the black figure "the black Dorito" across all three videos and compliments the "monster costumes".
• A VHS or camcorder grade over everything: chroma bleed, scanline shimmer, tracking noise.
• Burned-in dialogue captions ("Mom?", "Did you need something, honey?", "You okay?"); "Stay quiet" is narrated by a synthetic emergency-alert voice.
• Coverage: a static tripod for most shots, occasional handheld, shot/reverse-shot, an average shot of 1.5 to 2.5 s, a reveal beat about two-thirds of the way through.

The creator's side confirms it. Evillica is a videographer (182 K YouTube subscribers, 149 videos, 83 M views; 328 K on TikTok) whose two published interviews describe notes, then a storyboard, then filming, then Premiere, After Effects and Photoshop; they cite Henson and Krofft puppetry and treat AI as "a tool … it should never become the skeleton of the work". No synthetic-content label appears on any of the three videos; the TikTok "everything is fake" caption is a fiction disclaimer. The face warp on Mom is the only element that could plausibly be AI-assisted, and nothing public settles it.

So the question splits:

• Track A: can generation reproduce a live-action-looking single-actor, single-location camcorder short, with the same actor and flat recurring across episodes?
• Track B: what do the AI-native creators in the same niche do, and how good is it?

Track B is a known quantity: Angel Engine (58+ parts, generative AI admitted on its Steam page), content-farm clusters like Dead Signal VHS, viral one-offs like "That's not MOM.." (2.2 M views, 11 s). They change the conventions to dodge the hard parts: no recurring human actor, found-tape and emergency-broadcast framings that justify stills and text cards, narration over rotating subjects. YouTube's July 2025 and July 2026 "inauthentic content" rules were written against exactly that population. The rest of this document is about Track A, because that is where the engineering problems are.

The economic asymmetry underneath: AI cost is per shot and per reroll; live-action cost is per setup, and the recurring elements (actress, flat, costume) are free after episode one. A pipeline that makes the character and the room assets rather than prompts restores that asymmetry.

A survey of about 25 models against their primary pages and the Artificial Analysis arena found the familiar picture of "Veo 3 vs Sora 2 vs Kling 2.x" obsolete. Six shifts:

1. Sora is dead. The app closed 2026-04-26; the API closes 2026-09-24. Its $1 B Disney deal never funded; the $30 M Critterz feature lost its renderer mid-production and missed Cannes.
2. Chinese models own the top of the arena. Text-to-video with audio, fetched 2026-09-16: Gemini Omni Flash 1233, Wan 3.0 1229, MiniMax H3 Max 1227, MiniMax H3 1220, Seedance 2.0 1210. Kling 3.0 is 12th (1095), Veo 3.1 14th (1088). Image-to-video: H3 Max, H3, Omni Flash, HiDream-O1-Video, Seedance 2.0.
3. The 8-second wall is gone at the top. Wan 3.0 and Seedance 2.5 do 30 s in one pass; Kling 3.0, Seedance 2.0, H3, HappyHorse and Grok do 15 s; Veo 3.1 stays at 8 s plus 7 s extensions (720p only, up to 20); Omni Flash does 10 s, extending to 40 s.
4. References and native multi-shot are universal. Veo 3.1 takes 3 images; Kling Omni 7 images or 4 plus a video, with voice binding; Seedance 2.0 takes 9 images, 3 videos and 3 audio clips, Seedance 2.5 takes 30, 10 and 10; Wan 3.0 takes 20 assets including documents; MiniMax H3 takes 9, 3 and 3. Kling (6 cuts per 15 s, per-shot framing), Seedance ("Shot 1:/Shot 2:" prose), H3, Wan 3.0, LTX-2.5 and Omni Flash all cut between shots inside one generation.
5. Open weights: two real 2026 options plus the Wan 2.2 workhorse. LTX-2.5 (22B, audio and video, native multi-shot, 2026-08-11, free under $10 M ARR, gated download) and MiniMax H3 (33B, audio, 9-image references, 2026-08-03; the model card routes US, EU, UK and Korean users through an application, and secondary sources describe a local-deployment exclusion, so read the licence). Wan 2.2 (Apache-2.0, no audio) plus Wan2.2-Animate-2 (2026-08-07) remain Alibaba's open ceiling; Wan 2.5, 2.6, 2.7 and 3.0 are API-only despite SEO claims. HappyHorse's "fully open" claim has produced no weights. MAGI-2 is Apache-2.0 but needs 8 Hopper GPUs.
6. Nobody hosted guarantees determinism (section 6.4).

3.1 Condensed table

Prices differ by host.

3.2 Failure modes that persist (vendor-admitted)

Consistency through edits and extensions ("maintaining complete consistency throughout edits … remains a challenge", Google's Omni model card); complex motion; garbled text (composite title cards and VHS timestamps in post, never generate them); ignored negative constraints (a hand appears when forbidden); hands, faces and physics, still scored dimensions on VBench-2.0 because they fail; morphing on held shots (practitioner consensus, not quantified); audio caveats (Kling: no native audio with video input); prose-timed cuts on Seedance; extension only at 720p on Veo; and non-determinism everywhere hosted.

4.1 The determinism ladder

4.2 Reference conditioning in video models

Every hosted model now has a reference slot: Runway one image; Hailuo one; Luma one plus 10 s of video; Veo three; Vidu seven; Kling O1 seven images, or multi-angle Elements of 2 to 4 images plus a 3 to 8 s video that also captures voice; Seedance and H3 mixed media. All vendors say "consistent"; none publishes a face-similarity number. The one cross-vendor table (HunyuanCustom, 2025) puts hosted models at 0.42 to 0.53 face similarity, i.e. the same person, far from copy-paste. Hailuo's launch note admits "environmental morphing"; OpenAI's cookbook warned "small changes in phrasing can alter identity". An 8-week, 300-clip Seedance 2.0 production test calls multi-shot character consistency "reliable, not 100 % of the time". The 2026 practitioner finding is sharper: faces mostly hold; costume, hairline, props, gait and palette do not ("the face passed every review; the costume failed eleven of twenty-four"), and drift becomes perceptible after about ten shots.

4.3 Training the character (LoRA)

Tools: ai-toolkit (MIT) and musubi-tuner (Apache) cover FLUX.2, Qwen-Image, Wan 2.2, LTX-2.x and MiniMax H3 on a 24 GB card; Lightricks ships ltx-trainer for LoRA and IC-LoRA. The recipe practitioners converge on for Wan 2.2: 5 to 100 captioned images or ~2 s clips, rank 32 and alpha 16, learning rate 2e-4, flow shift 2 to 3 for tight identity, latents cached first. Hosted: fal charges $4 per 1,000 steps for a Wan 2.2 14B LoRA and $2 per FLUX run with commercial rights. Roughly 1 to 3 h on one H100 (a recalled figure). Licence-safe bases: Wan 2.2, Qwen-Image-Edit-2511 and FLUX.2-klein-4B (Apache); FLUX.1 and FLUX.2 dev and klein-9B are non-commercial; HunyuanVideo excludes the EU, UK and Korea; Krea Realtime is CC BY-NC-SA. A LoRA also learns the dataset's lighting and lens unless the set is varied, which for a rendered character is a feature: render under a dozen HDRIs and focal lengths and train on that.

4.4 Zero-shot adapters and the "wrong face" problem

For stills: PuLID-FLUX, InfiniteYou (model CC BY-NC), DreamO, OmniGen2, and the older InstantID and PhotoMaker line. For video on Wan 2.1 and 2.2 (Apache): Phantom, SkyReels-A2, MAGREF (ICLR 2026), Stand-In (153 M parameters, composes with VACE and LoRAs, V2 announced 2026-08-10), VACE, Wan2.2-Animate and Animate-2, InfiniteTalk, MultiTalk. HunyuanCustom has the best published face similarity (0.627) under the territory-limited Tencent licence.

All of these are supervised by an ArcFace-style embedding and increasingly by human-preference rewards (Identity-GRPO, Avatar V). ArcFace encodes who and discards what is off about the face; the diffusion prior then regresses toward an average, symmetric, attractive face. The WithAnyone paper names the complementary failure: adapters "copy-paste" the reference pixels and fail under new pose, expression and lighting. Large yaw in close-up remains the documented weak spot (FaithfulFaces, MoFE). The consequence for horror: eye spacing normalises, the asymmetric smile symmetrises, the dead eyes get a catch-light. The only generative object that can learn a defect as identity is a LoRA trained on images that contain it; the only thing that reproduces it exactly is a mesh.

4.5 Character sheets and keyframes

The workflow every vendor recommends, including Google's own Veo 3.1 guide: a canonical sheet, then a per-shot keyframe made by editing, then image-to-video from the keyframe. Editors: Nano Banana Pro (14 references, 5 people, SynthID on every output), FLUX.2 (10 references; dev non-commercial, klein-4B Apache), Qwen-Image-Edit-2511 (Apache, seedable), GPT-image-1.5 (input_fidelity: high; the cookbook says to repeat "do not change her face" in every call because drift occurs without repeated constraints), Seedream 4.5, Midjourney --oref ("does not promise exact copying"). All of them beautify; the defect must be spelled out every time.

4.6 3D-first: build once, render exactly, then use the render

Build the character as a rigged, photoreal 3D asset (body, blendshape face, strand hair, skeleton) and render it in Cycles. The render can then feed a generative model in three ways.

• (a) The render as the image-to-video first frame. Works with every model in section 3. Frame 0 is pixel-exact; identity decays at the model's native rate, slowly for a locked-off close-up, fast for a head turn (the model invents the unseen side from its prior). Give both ends where the model allows (Veo first+last, Luma keyframes, LTX multi-anchor), rendered from the same head.
• (b) The render as a control signal. Wan 2.1 and 2.2 VACE take a reference image plus a depth, pose or edge control video plus masks (Apache; 1.3B at 480p, 14B at 720p); LTX-2.3 and 2.5 IC-LoRA Union does depth, canny and pose in one adapter. Stand-In composes with VACE, so a frontal reference and a pose track ride together. Blender bridges exist: Pallaidium (GPL-3; a depth pass output added 2026-07-10, FLUX.2 klein multi-reference, LTX-2.3 IC-LoRA, ComfyUI and fal remote backends), ComfyUI-BlenderAI-node, the older controlnet-render addon. Geometry, camera and timing are then exact; the face is whatever the reference plus LoRA make it.
• (c) Video-to-video restyle of the render. Runway Aleph 2 ($0.28/s; its own editing guide now steers away from new angles); Luma Ray3 Modify with an "Adhere" strength that retextures and relights while keeping edges (the closest hosted thing to a deterministic render pass); Krea Realtime 14B (non-commercial); Wan2.2-Animate-2 replacement mode (put the character into a driving video and match its lighting); LTX IC-LoRA. RealMaster (arXiv 2603.23462, March
  2026. is the proof of concept: GTA V engine renders lifted to photoreal video "preserving the geometry, dynamics and identity specified by the original 3D control", distilled into an IC-LoRA that needs no anchors at inference. "Goodbye Drift" (arXiv 2605.20476) shows control-conditioned video-to-video holding for 40+ minutes when generated sparse-to-dense from anchors.

What survives: geometry, camera and timing survive (b) and (c)-Adhere almost completely; lighting survives Adhere; identity survives only as well as the identity conditioning you add; the wrongness survives only if it is in the LoRA or the restyle strength is low enough that the render dominates. A licence note: MetaHuman content "cannot be used to train or enhance AI models", which forbids training a LoRA on renders of a MetaHuman head; MakeHuman/MPFB and ICT FaceKit carry no such clause (ICT's exact text unverified). TRELLIS.2 is MIT but has no human-specific path.

4.7 Dialogue on a held close-up, ranked by identity safety

1. Animate the 3D head from audio with NVIDIA Audio2Face-3D (open-sourced September 2025; SDK MIT, models under the NVIDIA Open Model Licence; ARKit-style blendshape output; Maya and Unreal plugins but no Blender plugin, so a small importer onto the head's shape keys is needed) or with iPhone ARKit capture, then render. Zero drift by construction.
2. Relip the render: LatentSync 1.6 (Apache, a 512² mouth region, keeps everything outside the mouth) or InfiniteTalk (Apache, Wan 2.1, unlimited length; the FusionX speed LoRA reduces identity). LongCat-Video-Avatar-1.5 (MIT, 2026-05-21) is the newer open option.
3. Hosted single-image talkers for medium shots: Hedra Character-3 (6.25 ¢/s at 1080p, "character angle libraries" recommended), OmniHuman-1.5 (16 ¢/s).
4. Native dialogue in the video model (Veo, Kling, Seedance): the least controllable identity; dialogue realism degrades after about 8 s per clip and profile shots are problematic.

The genre's convention is captions, not audible lines, which removes most of this.

4.8 Measuring identity instead of eyeballing it

• ArcFace or AdaFace cosine similarity (insightface buffalo_l) between each generated frame and a canonical render set. Published operating points run from 0.33 to 0.82 depending on the false-accept rate; a recent ICLR paper treats < 0.3 as a different person and ≥ 0.5 as the same person with variation; thresholds are dataset-dependent and must be calibrated.
• VBench "subject consistency": DINO feature similarity first-frame-to-frame and frame-to-frame. DINO sees within-class change; CLIP does not.
• A proposed gate: about 50 Cycles renders of the character (yaw ±60°, pitch ±20°, 6 HDRIs, 5 expressions); set the shot gate at the 5th percentile of the within-set ArcFace distribution (expect about 0.55 to 0.65 frontal); require DINO frame-to-frame similarity at least equal to the raw playblast's. ArcFace is blind to an authored defect by design, so pair it with a landmark-ratio check (eye spacing, mouth width, asymmetry, from MediaPipe or the mesh topology) that measures the specific wrongness.

4.9 Ranking for "the same wrong face on demand"

1. Author the wrongness in the 3D head and render in Cycles. Deterministic at any angle, light and expression; dialogue via Audio2Face-3D or ARKit. The raw render is already uncanny in the way analog horror wants (perfect skin, too-still eyes), so the AI pass may be optional.
2. Train a character LoRA on those renders (Wan 2.2 14B via musubi-tuner or fal; Qwen-Image-Edit or FLUX.2-klein-4B for stills; all Apache) with a caption token naming the defect. Use it for photoreal keyframes at low denoise, inside VACE and Stand-In for restyled shots, and inside LatentSync or InfiniteTalk for dialogue.
3. Image-to-video from a rendered first (and last) frame with minimal camera motion; gate with ArcFace-to-render plus the landmark check; record every seed and weight hash.
4. Hosted multi-reference (Kling Elements from four rendered angles plus a voice element; Veo ingredients from Nano Banana keyframes) for establishing and medium shots where the face is small and the wrongness is carried by posture and framing.

Routes 1 to 3 give a repeatable wrong face today. Route 4 gives a repeatable normal face and will silently fix the wrongness in close-up. Nothing on the market lets you upload a defect and have it respected without training.

5.1 Nothing in the video-model camp keeps the door where it was

Veo Ingredients, Runway References, Kling Elements (scenes are a first-class element type) and the rest condition on an image of the room, not on its geometry. They keep the look, palette and rough furniture set; they do not carry a floor plan. The cleanest negative evidence is from real-estate staging (Inman, April 2026): tools that treat each photo in isolation "place different furniture in different locations across the photos … no memory of layout". VideoGPA (ICML 2026) states the mechanism: denoising objectives "lack explicit incentives for geometric coherence". No vendor publishes a rotation budget; the practitioner rule is that the visible half of a room holds for pans of about 30° to 45° from one reference, and beyond that each generation invents the other half differently.

The strongest video-model-only primitives:

• First + last frame between two known views (Wan 2.1 FLF2V-14B, Apache; Veo 3.1; Kling start/end; Wan 2.7). Both endpoints are fixed, so the door is where you put it at both ends; hold focal length, angle, distance and key-light direction constant, or the model "commits to an interpolation path early" and artefacts mid-clip. If the endpoints are renders of one 3D set, the middle is usually plausible because it is a real camera move.
• Kling 3.0 Omni multi-shot: up to six cuts inside one 15 s generation share one latent, which is where the location is most likely to stay coherent; no control over where the cuts land.
• Extend modes condition on the tail of the previous clip and hold the room only while the view overlaps; not a way to get a reverse angle.
• Reference propagation: the last clean frame of clip N as the reference for clip N+1.

5.2 Generated 3D worlds are now real assets

Generation is not repeatable; the exported file is. Generate once, commit the file with a hash, and never regenerate it: the same discipline as a package lock file.

5.3 3D-first plus AI

Build the set (Infinigen, hand-dressed, Marble, or a capture), place the camera and character, render beauty, Z, normal and segmentation passes in Cycles, then:

• Image-to-video from the rendered first frame: the layout is exact at frame 0 and decays with camera motion; enough for the tripod shots that are most of this genre; first+last frame from two renders is the upgrade.
• Depth-locked control (Wan 2.2 VACE; LTX-2.3 and 2.5 IC-LoRA Union): the silhouette and depth ordering of every wall, door and prop is pinned per frame, so the door cannot move because its depth edge is in the control video. The RunComfy "Blender → ComfyUI AI Renderer 2.0" workflow is exactly this; its rules: identical aspect ratio and fps between Blender and the graph, 4n+1 frame counts, recheck the first and last 10 frames.
• Restyle (Luma "adhere", Aleph 2, Krea Realtime, Decart Lucy): structure approximately kept; the live restylers drift over long runs.
• Per-frame enhancers (Magnific, Krea) flicker on sequences; use a temporal upscaler (Topaz, SeedVR2) for video and per-frame enhancers for keyframes only.

Camera-control re-shoot models (GEN3C with an explicit 3D cache, NVIDIA OML; Uni3C, Apache; TrajectoryCrafter; ReCamMaster, MIT but documented to break across occlusions; Stable Virtual Camera, non-commercial; CameraAnything, July 2026) re-render a video from a new camera and hallucinate what the source never saw. They are redundant when you own the set: Blender is the camera-control model. They earn their place for re-shooting a hosted clip you like.

5.4 Lighting continuity

Trivial in 3D: one light rig, with the lamp keyframed per time of night. Splat sets have baked illumination. AI relighters: IC-Light V2 weights were never released and are non-commercial; Light-A-Video (Apache, Wan 2.1 backbone) and NVIDIA DiffusionRenderer (G-buffers from a splat or real video, then relit under a new HDR map; ~2.4 s clips) are the usable video relighters; the practical route without a set is to relight the keyframe and generate video from it. LTX-2.5 has a Relight IC-LoRA in beta.

5.5 The VHS layer goes last, once, over the whole cut

ntsc-rs (open source; an OpenFX, After Effects, Premiere and Resolve plugin plus a standalone app) simulates the actual NTSC and VHS signal path. Apply it after generation and any upscale, as one parametric operator over the finished cut, because it (a) collapses the different "film stocks" that different generators, seeds and even clips of one generator produce below the tape's noise floor; (b) low-passes exactly the high-frequency texture shimmer where AI video flickers most; (c) turns clip joins and restyle-strength changes into tape dropouts and tracking errors, the genre's own vocabulary; and (d) is deterministic, so a regenerated upstream shot never needs the grade re-approved. Prompting "VHS look" into each generation gives each clip its own invented VHS, another axis of inconsistency. Put on-screen timecode before the tape emulation so it degrades with the picture.

5.6 Where the ceiling sits

6.1 Three architectures

A. Per-shot hosted generation with a reference kit (what the AI-native niche does). Script beats, then a Nano Banana character sheet and location keyframes, then image-to-video per shot on Kling, Veo or Seedance with the sheet as an ingredient, motion-only prompts, the camera locked off, two variations per shot, 10 to 15 frames trimmed from each end; then ElevenLabs or captions; then CapCut with a VHS overlay stack. Consistency is probabilistic and non-reproducible; drift shows after about ten shots; the practitioners' fix is a frozen reference kit ("the moment you 'improve' the hero shot in week two, every earlier shot becomes inconsistent") plus a per-shot checklist and last-frame chaining.

B. Native multi-shot in one call. Kling Omni (six structured cuts), Seedance 2.5 (30 s, 50 references, scene-level prose), Wan 3.0 (30 s), LTX-2.5 (native multi-shot, open), Omni Flash (conversational edits). For a 30 to 60 s Short this is 2 to 4 calls instead of 8 to 12, and continuity only has to survive across those seams. Timing control is prose on Seedance and structured on Kling; no public benchmark isolates cross-shot identity.

C. The 3D-first film build. Character and set are assets; every shot's blocking, camera, lens, lighting and hold time are keyframes; Cycles renders beauty and passes with a fixed seed; an open-weight pass adds photographic surface under depth and identity lock; the edit, captions and VHS grade are deterministic post. Only the surface is stochastic, and with pinned open weights and a seed even that repeats. This is also what Autodesk Flow Studio's 3D Editor and Canvas (August 2026) productise: 3D for "performance, staging, composition and camera movement", generative models for "lighting integration, atmosphere … and cinematic finishing".

6.2 Where each architecture breaks

6.3 Audio, captions, assembly

Narration is safer as a separate voice track than as generated dialogue; the genre's captions remove lip-sync entirely. Native audio in the video models (Veo, Kling, Seedance, H3, LTX-2.5, Vidu) is a scratch track: keep it for Foley timing, and replace narration and music with locked assets, because it is tuned for dialogue, not for a knock three frames before a cut. Text on screen must be composited, never generated.

No commercial product accepts a JSON shot list end to end. Descript's API takes only a natural-language prompt and publishes to a web link; Eddie for Agents offers CLI, SDK and MCP operations over real footage; Runway's agent runs Workflows from chat over MCP; CapCut is driveable only through an unofficial draft-file MCP; LTX Studio, Flow, Higgsfield, Hailuo Agent and Vidu Agent are UI-first. Google Flow deletes generated videos from the server after two days; Sora had an export deadline. The assembly layer must be self-built, and hosted output archived on receipt. The VHS pass goes last (section 5.5).

6.4 Determinism end to end

Bit-repeatability exists only self-hosted: torch.Generator(device="cpu") even on GPU, CUBLAS_WORKSPACE_CONFIG=":16:8", cudnn.benchmark=False, torch.use_deterministic_algorithms(True), a pinned GPU SKU, driver, CUDA, PyTorch and attention kernel (an H100 run will not bit-match a B200), and a fixed scheduler and step count; MAGI-2 ships a --deterministic flag. Diffusers' own guide: results are not guaranteed "across PyTorch releases, individual commits, or different platforms". Hosted: Veo's seed "doesn't guarantee determinism, but slightly improves it"; Runway's gives "similar results"; Seedance's "minor variation may still occur"; Omni Flash and Sora expose no seed; fxguide (August 2026) documents cloud models changing output under identical input and seed without changelogs.

A reproducible episode borrows the shape of a package manager: a manifest, a lock file and a build script.

film/<slug>/
  film.json        # source of truth: cast (character id → asset hashes), locations (set hashes),
                   # shots[] {id, duration, framing, camera, light state, action, dialogue, sfx[],
                   #          surface model, params, seed}, audio {narration, music, sfx}, edit,
                   #          degrade preset, loudness target
  film.lock.json   # per stage: inputs-hash → {provider, model version, request, seed, output sha256,
                   #          cost, time, accepted_by: auto|human, qc scores of accepted and rejected takes}
  build/           # content-addressed artefacts, git-ignored or LFS

Rules that make it reproducible in practice, each grounded in a published half-example:

1. Every stage is keyed by a hash of its inputs (prompt, params, seed, upstream artefact hashes, model ID and version string). A rebuild reuses the cached artefact when the key matches; this is the only way a hosted, non-deterministic stage becomes "reproducible". ftl-studio (MIT; Veo 3.1, Gemini and Claude) frames its plan.json as "a reproducible recipe" while admitting the APIs expose no seed.
2. Human acceptance is recorded in the lock, not in chat: which candidate was chosen, and the QC scores of the rejected siblings. ftl-studio's qc.json scores identity, wardrobe, set match and manifest compliance from 0 to 100 with a vision model before any video spend; ViStoryBench and MSVBench publish open scorers for character consistency, style, prompt alignment and copy-paste artefacts.
3. Generate the establishing shot first and reference it everywhere: GroundShot (2026) finds "the visual quality of this initial appearance sets the consistency ceiling for all that follows".
4. Prefer local, seeded models for anything iterated on (sheets, plates, degrade, music) and pin hosted hero shots by hash. The published Krea-2 multi-shot node uses seed + N - 1 per shot, the cleanest example of per-shot seed discipline.
5. Emit the edit as .otio plus a Remotion or ffmpeg render script, both derived from film.json; render under bitexact with a pinned ffmpeg.

No public project implements the lock-file half; the "video as code" repositories are declarative timelines without generation provenance.

6.5 Assembly frameworks and agentic pipelines

Native multi-shot in one call (the strongest inter-shot consistency available without 3D):

A 45 s Short is 3 to 4 multi-shot calls or 6 to 12 single-shot calls. Multi-shot buys consistency inside each call and pushes the problem to the seams, where a locked sheet, a locked plate and the establishing-shot-first rule matter. It also fights a 3D lock, because it decides framing for you; use it only where the shot list says "montage".

Agentic and open frameworks, by how usable they are today:

• Runnable and relevant: Wan2.2-Animate-2 (Apache, driving video plus reference character), the bridge from a Blender playblast to a photoreal shot without regenerating identity; the RunComfy Qwen-Image-Edit + Wan 2.2 "cinematic coherence" graph; ComfyUI-Wan-VACE-Prep for transitions and extensions; MoneyPrinterTurbo (124 k stars, MIT: LLM script, TTS, clips, burned subtitles, ffmpeg, 9:16, API, CLI and agent modes) as the Shorts half minus the locks; Vanta (MIT: a JSON timeline rendered by Remotion, WhisperX captions, Wan 2.2 and LTX, ACE-Step, LatentSync, licence-audited); OpenCut (MIT, a TypeScript timeline to MP4); Nomi (AGPL, 514 stars: storyboard, references, generation, then an "editable first cut on a real timeline"; 23 MCP tools so a coding agent can drive it; local ComfyUI as one of about 12 providers; no seeds, no audio stages, no documented project format).
• Closest to the 3D-first build: FilmAgent (a Unity sandbox with 15 locations, 272 camera shots and 21 Mixamo actions; LLM director, screenwriter, actor and cinematographer roles; a deterministic 3D render) and Cutscene Agent (arXiv 2604.25318, MCP-driven engine agents producing editable engine-native cinematic assets). These are the "shot list, then 3D scene, then render" pattern in a game engine.
• Research, not yet runnable: STAGE (a structural storyboard of start and end frame pairs per shot plus a multi-shot memory pack), GroundShot (training-free, model-agnostic entity memory; its method is usable by hand), ViMax, Captain Cinema (keyframe planning plus interleaved synthesis, 1,000 s films, no code), and MSVBench's finding that current systems are "visual interpolators rather than true world models".
• Historical: MovieAgent (a stale SVD stack), StoryAgent, Anim-Director, VideoGen-of-Thought, DreamFactory and Mora, superseded by native multi-shot models.

Reference pipelines published in 2026 with code: ftl-studio (one sentence to a directed multi-shot film on Veo 3.1; canon stills approved once, every frame composited from them and chained to the previous frame; vision-judge QC before video spend; an ffmpeg finish with a unified grade, grain, crossfades and loudness; no narration; about $6 per 64 s pass on Veo Fast); Krea-2 multi-shot stills with an identity LoRA into LTX Director (seeded stills, hosted video); MoneyPrinterTurbo plus Vanta (fully local possible, the highest determinism, the lowest craft). No verifiable write-up of an AI analog-horror pipeline specifically could be found.

7.1 What has been made

Festival tier: the Runway AIFF 2026 Grand Prix winner "A Face Only A Mother Could Love" is 7:50; winners run 3.5 to 11.6 min; the festival requires generative video but no minimum AI percentage; process is rarely disclosed. Studio tier: Netflix's one generative AI shot in El Eternauta; Amazon's House of David with 72 of 850 VFX shots AI in season 1 and 350 to 400 in season 2, plates and inserts, never a recurring AI character. The only place AI narrative is mass-produced is China: 470 AI dramas per day, RMB 500 to 1,000 per finished minute (about $30), 10-person teams doing 30 episodes in 20 days, a 0.117 % breakout rate, the lowest willingness to pay of any AI format, and the top revenue still with human-acted titles. Creator tier: production logs from a vendor's own agent show a 3-minute episode at 164 generations, 41 kept (25 %), 17 of the 41 final shots stitched from two or more takes, 2 people, 2 days, about $950; a 90 s horror short at about 400 generations, 2 days, $870.

7.2 Residual gaps

A keep rate of about 25 % (3 generations per usable shot, some 8+); identity drift perceptible after about ten shots, mostly in costume, hairline, props, gait and palette rather than the face; dialogue realism that degrades after about 8 s per clip, and poor "quiet, layered" emotion; extension chains that accumulate colour and identity walk; references that bias but do not constrain; blocking and camera direction that obey prompts loosely (the reason 3D-first hybrids appeared); and nothing repeatable. In one line: faces are mostly solved, everything around the face is not, and nothing hosted is repeatable.

7.3 Cost for a 45-second, 8 to 18 shot vertical short

An independent estimate for an Evillica-shaped short (18 shots, 4 rerolls each, about 400 generated seconds) agrees: about $37 to $46 on Kling 3.0 at 1080p, $48 on Veo 3.1 Fast, $20 on Veo Lite, $160 on Veo Standard, and half a day to a day of solo work.

7.4 Trend line and forecast

In twelve months: native clips went from 8 s to 15 s to 30 s, with a 3-minute beta; native audio everywhere; references from 3 to 50; multi-shot storyboards; conversational editing; prices down 3 to 5× per second at equal or better quality (while Sora ran at about $1 M a day against $2.1 M of lifetime revenue); MiniMax H3 the only open model in the arena's top 15. No Veo 4, Sora 3, Gen-5 or Seedance 3 has been announced. The frontier moved from "better clips" to directability, which is the right direction for a series.

A sober forecast:

• A human-curated 45 s analog-horror Short with a consistent character and room: now (an afternoon, $40 to $200, keeping 1 take in 4). Analog horror is the most forgiving genre for these defects.
• Agent-curated end to end, publishable most of the time: 6 to 18 months; the remaining work is the judgment (which take, where to cut, when drift is a feature) that a 25 % keep rate says models cannot yet self-assess.
• A human-curated 3 to 10 minute dialogue short: now, at 1 to 3 people for 1 to 3 weeks, with heavy stitching.
• An automated 10-minute dialogue short with subtle performance and 100-shot continuity: 2 to 4 years. Of its two blockers, only reproducibility has a known engineering path (open weights, a LoRA and 3D structure); performance does not.

The question was whether AI can generate a film like these. For a one-off, the answer in section 7.4 is yes, with a person curating. For a series, which is what makes a channel like Evillica's work, it is no, and the reason is structural rather than a matter of model quality:

• The recurring actress and the recurring apartment are the two things that make live action cheap after episode one, and they are the two things no generative video model can lock. References improve "recognisably the same"; they do not produce "the same".
• Every hosted route is non-repeatable and exposed to model retirement, so a show built on one cannot reshoot a scene next month, let alone next year.
• Everything that does lock identity and layout moves the work out of generation: a trained LoRA needs a consistent source to train on, and a depth-locked pass needs a 3D set to render from. Follow that to the end and the film is a conventional 3D production with a generative surface pass, which is architecture C. It can be a good way to make a film, but it is not AI generating one, and its up-front cost is building the character and the set.

8.1 What a 3D-first episode build would contain

For anyone who does take architecture C, these are the stages it needs.

The raw Cycles render with the VHS pass and no surface stage is a legitimate product on its own: the genre's aesthetic absorbs a clean CG interior and a too-still face, and it is bit-exact end to end. The surface stage is an upgrade whose value should be judged against that baseline, not assumed.

8.2 Licences

Keep to Apache, MIT and BSD wherever a model touches shipped pixels: the Wan 2.2 family (VACE, Stand-In, Animate-2, InfiniteTalk), LatentSync, Qwen-Image-Edit-2511 or FLUX.2-klein-4B for stills, Infinigen (BSD-3), ntsc-rs, and Audio2Face-3D (SDK MIT; models under the NVIDIA Open Model Licence, which needs a read). LTX-2.5 is usable under $10 M ARR and must not present machine output as human-made. Avoid without reading the licence: HunyuanVideo, HunyuanCustom and HY-World (EU, UK and Korea exclusion), MiniMax H3 (territory application), FLUX dev and klein-9B (non-commercial), Krea Realtime, Stable Virtual Camera and InfiniteYou's model (non-commercial), and MetaHuman (no AI training on its output). Marble needs the Max plan for commercial rights.

8.3 Experiments that would settle the open questions

None of these were run. They are the cheapest tests of the claims above that rest on vendor statements or practitioner consensus.

1. Determinism: run LTX-2.5 distilled and Wan 2.2 VACE twice on one H100 under the recipe in section 6.4 and check bit-equality; then across an H100 and a B200 (expect a mismatch, and decide which SKU is canonical). Measure seconds per 5 s and 10 s 720p clip.
2. Identity under restyle: render 50 canonical frames of a character (a yaw, pitch, HDRI and expression sweep), calibrate the ArcFace gate on them, then run one held 10 s close-up through VACE at three control strengths and through Luma "adhere"; plot face similarity and the landmark-ratio defect against strength. This answers whether a deliberate wrongness survives.
3. Layout under restyle: one Infinigen room, a 90° pan rendered in Cycles, the same pan through depth-locked VACE and through image-to-video from the first frame; measure depth-edge IoU against the Z pass per frame.
4. The drift benchmark nobody publishes: a 10-shot, same-character, same-kitchen sequence on Kling Omni, Seedance 2.5 and LTX-2.5 from the same keyframes, scored with the section 4.8 gates. This decides whether hosted inserts are worth having at all.
5. Licences: read the MiniMax H3 community licence and the NVIDIA Open Model Licence in full.

A rough cost: single-digit dollars of rented GPU time per experiment, plus a Kling or Seedance credit pack (about $40) for the fourth.

US Copyright Office, Part 2 (2025-01-29): prompts alone are not authorship; a human's creative selection, arrangement and modification of AI output is registrable with the AI portions disclaimed, so the edit of a short is protectable and the raw clips are not. YouTube: since 2025-07-15, "inauthentic content" (templated, mass-produced AI) is demonetised, while a series with recurring characters and distinct storylines is explicitly allowed; the July 2026 clarification adds an "unsatisfying or off-putting … designed to shock" bucket that an AI horror channel must avoid; realistic synthetic scenes need the disclosure label (which does not affect monetisation); C2PA metadata auto-labels; and from 2027-02-01 the Shorts Creator Pool requires 10 M Shorts views per trailing 90 days. Practical-plus-post work like Evillica's needs no label. Veo and Omni carry SynthID. Seedance carries live US legal exposure (studio complaints, a Senate letter).

The ones that matter: the "slow push-in breathing" failure mode is practitioner consensus, not a measurement; no public benchmark isolates cross-shot identity, and VBench-2.0's 2026 tables would not render; the MiniMax H3 geographic clause comes from secondary sources; LTX-2.5's seconds per clip on one H100 and its run-to-run bit-equality are vendor claims until experiment 1 runs; Reddit and Instagram were unreachable, so community reroll counts come from vendor and course pages; the ICT FaceKit licence text was not re-read; and whether the Mom face warp in the reference shorts is AI-assisted is unknown.

The primary pages relied on most, by topic. Every claim above was checked against these or marked as recalled or unverified.

• Reference shorts and creator: youtube.com/@Evillica; the horrortoculture.com interview (2025-12-28); the houseofscream.com BUFF piece (2026-04-13); the three watch pages' metadata and comment threads.
• Models: ai.google.dev Gemini API docs for Veo, Omni and pricing; the deepmind.google Omni model card; kling.ai Omni and Element Library guides; Runway's API docs (api.md, pricing.md, changelog); seed.bytedance.com Seedance 2.5; huggingface.co MiniMaxAI/MiniMax-H3; huggingface.co Lightricks/LTX-2.5 and github.com Lightricks/LTX-2; github.com Wan-Video/Wan2.2 and Wan-Animate-2; huggingface.co Wan-AI; huggingface.co sand-ai/MAGI-2-preview; the artificialanalysis.ai video leaderboard; the-decoder.com on the Sora shutdown; the Hugging Face diffusers reproducibility guide.
• Characters: github.com WeChatCV/Stand-In, Phantom-video/Phantom, MAGREF-Video/MAGREF, Tencent-Hunyuan/HunyuanCustom (face-similarity table and LICENSE), ostris/ai-toolkit, kohya-ss/musubi-tuner, bytedance/LatentSync, MeiGen-AI/InfiniteTalk, NVIDIA/Audio2Face-3D; arXiv 2510.14975 (WithAnyone), 2510.14256 (Identity-GRPO), 2605.04702 (FaithfulFaces), 2603.23462 (RealMaster), 2605.20476 (Goodbye Drift); Google's Veo 3.1 prompting guide; OpenAI's GPT-image-1.5 cookbook; cgchannel on the MetaHuman licence.
• Scenes: the worldlabs.ai Marble blog and mesh-export docs; github.com Tencent-Hunyuan/HY-World-2.0 and its LICENSE; github.com princeton-vl/infinigen; github.com Kiri-Innovation/3dgs-render-blender-addon; the Wan2GP VACE doc; docs.ltx.io IC-LoRA adapters; the runcomfy.com Blender to VACE workflow; the flick.art Blender AI filmmaking guide; docs.lumalabs.ai Modify Video; github.com nv-tlabs/GEN3C, alibaba-damo-academy/Uni3C, KlingAIResearch/ReCamMaster; ntsc.rs; inman.com on multi-angle staging (2026-04-27); arXiv 2601.23286 (VideoGPA); the PyTorch randomness note.
• Assembly: the fal.ai Kling v3 pro API schema (multi_prompt, no seed); the Emily2040/seedance-2.0 api-status notes; huggingface.co MiniMaxAI/MiniMax-H3 and Comfy-Org/MiniMax-H3; docs.comfy.org LTX-2.5; github.com yihao-meng/HoloCine, SandAI-org/MAGI-1, showlab/MovieAgent; arXiv 2512.12372 (STAGE), 2606.20799 (GroundShot), 2602.23969 (MSVBench), 2604.25318 (Cutscene Agent); huggingface.co papers 2501.12909 (FilmAgent); thecinema.ai (Captain Cinema); github.com uby174/ftl-studio, CodingWithShahzaib/ComfyUI-Krea-MultiShot-Stills, harry0703/MoneyPrinterTurbo, itsjwill/vanta, floomhq/opencut, aqm857886159/Nomi, idreesaziz/swimlane; the elevenlabs.io TTS API reference (seed) and Music docs; the OpenAI TTS guide; ai.google.dev speech and music generation; github.com resemble-ai/chatterbox, ace-step/ACE-Step, Tencent-Hunyuan/HunyuanVideo-Foley, hkchengrex/MMAudio, m-bain/whisperX, AcademySoftwareFoundation/OpenTimelineIO; the ffmpeg formats and codecs docs (bitexact); the remotion.dev flickering and openai-whisper docs; docs.descriptapi.com; heyeddie.ai/devs; github.com atx-guy/capcut-mcp-server; the blog.google Flow tips (2-day retention).
• Readiness and policy: cartoonbrew.com and thenextweb.com on Critterz; techcrunch.com on the Sora shutdown; hellochinatech.com and techtimes.com on China's AI dramas; invideo.io generation-count FAQs; screenweaver.ai on character drift (2026-08-24); fxguide.com on on-premises open weights (2026-08-28) and Autodesk Flow Studio (2026-08-24); hackernoon.com on retired models (2026-09-10); aif.runwayml.com; copyright.gov/ai; YouTube Help answers 1311392 and 14328491; tubefilter.com and techcrunch.com on the July 2026 YouTube clarification; runpod.io/pricing; fal.ai model pages; eesel.ai on Kling pricing; elevenlabs.io/pricing.