How to create your own cartoon or film with AI

Generating a beautiful shot and making a film are different tasks

A hero enters a room. In the next shot, they have a different nose, the door has moved, and the cup has changed hands. Each clip may look expensive on its own. Together, the story falls apart.

That's why it's best to start with controlled shot sequencing: who acts, what they want, what changes, and how the viewer notices it. The generator helps obtain images and motion. Scripting, staging, shot selection, editing, and sound join them into a film.

And one more fundamental difference: video in a "3D cartoon" style is usually a finished moving image. It does not automatically provide a mesh, rig, skin, shaders, or animation keys for editing in 3D software. If a client needs a pose change in a specific frame or a different angle of the same scene, this must be built into the production method from the start.

Review method. Product references, primary company announcements, author credits, and festival websites were verified on October 24, 2026. Manufacturer advertising promises are not interpreted as independently measured quality. The educational fox story below is a fictional creation by the editorial team; we did not produce a finished film based on it.

What major studios are already doing — and how they create festival works

Lionsgate and Runway: a confirmed partnership, not "all of Hollywood at one button"

On June 11, 2026, the companies announced an expansion of their cooperation: Lionsgate acquired a stake in Runway, with plans for joint development of new projects, specifically in short episodic formats. The announcement mentioned previsualization, storyboarding, and working with final shots. This confirms a specific production direction but does not prove that any particular feature film by the studio was entirely generated. [1]

Netflix / Eyeline: generative AI within the VFX process

In October 2025, Netflix described the application of generative AI by the Eyeline team in a series The Eternaut. The publication concerns broader work with visual effects and production technologies; it does not reveal a universal public set of models. This does not imply that a subscription to any service will replicate a studio's internal process. [2]

Ukrainian example: Distance Between Two Points Of Me

The credits for the short film by Illia Dutsyk and HOLYWATER list Runway, Midjourney, Kling, Stable Diffusion, Sora, Flux, and Luma. Alongside them are the editor, art director, AI-generalist, composer, and voice actress. This is a historical list of the 2025 toolkit, without attributing today's versions to them. The official AIFF 2025 program confirms the Honoree distinction; duration — 2 minutes 27 seconds. A useful takeaway for your own film: different tasks may require different tools, while voice and music can remain authored. [3] [4]

Current festival snapshot: 2026

On the official Runway AIF 2026 Grand Prix website, noted for A Face Only A Mother Could Love Robert Gaudette, Gold — for THE WELL the duo Dorian & Daniel. BIFAN 2026 names the best AI film Cornered Odaira Falek and Davi Murri. Awards are confirmed by the festivals themselves. Tools from separate production descriptions are provided in the following table; for THE WELL, we do not attribute an unknown set here. [5] [6]

Specifically: which AIs were used to make the 2026 short films

These are declared toolsets for specific works from the Reply organizer's announcement and author portfolios. The list does not disclose every operation or model version and is not a quality ranking.

What to take for yourself: for images and scenes, authors use several environments, and for motion, they may combine different generators. Artistic animation can be built through approved frames and a shared style; a large-scale scene can be broken down into separate shots. Check the specific distribution of roles between services in the author's making-of, rather than inferring it solely from a list of logos.

Reply AI Film Festival is a separate competition in Venice. Its finale should not be called a selection for the main competition of the Venice Biennale Film Festival. This refers to the status from the finalist announcement on July 28, 2026, without attributing subsequent awards. [41]

There are also open process demonstrations from manufacturers: Google names Dave Clark, Henry Dobrez, and Juni Lau among the filmmakers who worked with Flow and other tools and techniques. This is a source for studying approaches, not a confirmation of festival wins for these specific works. [7]

Three AI films worth watching before your own

This is a selection based on confirmed festival recognition, not a ranking by view count. Each work has its own purpose: a personal story, a character drama, or a visual essay. Full videos can be opened here or on YouTube; the author may change availability or embedding permissions.

Tips under the video are editorial suggestions for learning. The mentioned tools are taken from credits and festival descriptions; they do not prove that every frame was created in the same way.

What to generate with: tools for different parts of the work

A service, a model, and a function are three different things. One service can have several models with different references, sound, pricing, and limitations. Compare the specific mode you need, not just the company name.

On a phone, the table can be scrolled sideways. The tool name leads to its breakdown below.

Google Flow: pay attention to the model inside

Flow should no longer be described merely as a wrapper for Veo. The August 27, 2026 update adds new capabilities Gemini Omni 1.1 Flash: start and end frames, quick drafts, and resolution enhancement. For Veo 3.1 Ingredients to Video Google describes working with character, object, and environment references separately. Not all combinations are available in every mode. [8] [9]

Our picks for testing: a scenario where it is important to quickly change the look or environment and compare options. During testing, Ukraine is on the list of supported countries. For individual access, the documentation specifies Google AI Plus, Pro, or Ultra and age 18+; this is not a promise of permanent free video generation. Separate access is described for relevant Workspace plans. [10] [11]

Runway: from approved look to acting

Gen-4 Image References helps transfer character characteristics into new images. The approved frame can then be used in video mode; the Gen-4.5 documentation describes text-to-video and image-to-video. This is convenient logic for an animation: first check the design statically, then spend credits on motion. [12] [13]

Act-Two solves a different task: it takes a performance video and an image or video of a character. This allows you to set the performance with your own recording rather than begging for a specific emotion with a long prompt. The documentation notes support for up to 30 seconds; this is a characteristic of this mode, not all of Runway. Still, check the accuracy of eyes, hands, and lips. [14]

Kling: references and multi-plane scenes

In the Kling 3.0 announcement from February 5, 2026, the manufacturer describes multi-plane generation, and for Video 3.0 Omni — reference video and individual plane settings. This is a candidate for testing when you need to combine a character, motion, and changes in shot scale. Multi-planarity does not guarantee successful editing: check every seam and the possibility of using fragments separately. It is impossible to determine the exact current price and free quota from this announcement — we do not replace them with old figures from reviews. [16]

Luma Ray 3.2: when the motion is already set

The official Ray 3.2 Modify Video documentation describes transforming an existing clip, separate control of structure and motion, and reference images at specific moments. For this V2V mode, up to 64 reference frames and a source of up to 20 seconds are specified. This is a different scenario than arbitrary text-to-video generation; do not apply these numbers to all Luma modes. [17]

Our picks for testing: to film simple acting motion or render a 3D draft, then check the stylization. If you need to approve a precise gesture, first establish it in the source and then compare whether the contact with the object has changed.

Adobe Firefly: useful control with important limitations

Adobe documents composition references, camera movement, and start/end images. At the same time, adding a first or last frame disables some other settings. Therefore, the list of features does not mean they can be applied simultaneously. For someone already editing in Adobe, this is a suitable candidate for inserts and controlled tests. Adobe models and third-party models within Firefly have different terms: the name of the shell does not make their licenses identical. [19] [20]

Midjourney: a strong candidate for looks, but check the version

For V8.1/V8.2, the documentation describes Edit Model with support for up to four references. For V7, there is Omni Reference. The old V7 parameter instructions should not be mechanically applied to V8.2. We consider Midjourney primarily for finding a look and preparing frames; a beautiful set of illustrations alone does not yet solve character performance. [21] [22]

ComfyUI, Wan, and LTX: more freedom, more technical work

ComfyUI allows you to save the generation workflow, and official examples explain how Wan2.2 works. In the Wan repository, there is a separate Wan-Animate for character animation using a video source. It should not be confused with the lighter Wan2.2 5B model: requirements and compatibility differ; the authors also warn against arbitrarily applying LoRAs trained on Wan2.2 to Wan-Animate. [25] [26]

LTX-2.5 — another open-weights option with documented multi-plane generation. Its community license has conditions, specifically an annual organizational revenue threshold of $10M; this is not equivalent to unrestricted use. For a local workflow, choose a specific checkpoint, its license, recommended VRAM, and a verified graph. Having a gaming RTX in a laptop does not, in itself, guarantee comfortable speeds. [27] [28]

What is better for a cartoon vs. realistic animation

There is no universal winner without identical testing. We recommend choosing based on the level of control required. Stylization can hide some texture flaws, but not an incorrect number of fingers, a lost silhouette, or an incomprehensible action.

Practical test: take three tasks — a close-up of an emotion, a hero turn, and interaction with an object. Compare two models using the same input materials. Don't count the best clip from social media, but the number of attempts required before your shot is editable.

How to maintain a consistent look for a single character

You need to maintain three types of consistency: identity — who it is; design — silhouette, clothing, materials; scene continuity — where they stand, where they look, and what they are holding. A face reference helps with the first, but doesn't cover everything else.

1. Create a character passport before video generation

Approve front, three-quarters, profile, back, full height, and several key expressions. Choose one main standard. Additional angles should explain it, not create new "almost similar" characters. If the generator makes a mistake on the angle sheet, fix it now.

Separately record unchanging features and allowed changes. For example: "paper red fox, one dark left ear, short turquoise scarf; no change in material or snout length; facial expressions and pose change." The more fine asymmetry and complex decorations, the more checks you will need.

2. Separate the hero reference from the style reference

A face photo answers the question "who," a stylistic shot answers "how it is depicted," and a composition sketch answers "where it stands." Mark these roles explicitly. In supported modes like Runway References, Midjourney Edit Model, Veo Ingredients, or Kling Omni, reuse approved sources. Do not provide a new random image of a similar person every time.

3. First, approve all starting frames

Collect them into a single contact sheet. Compare eye shape, nose line, height relative to doors, hair silhouette, fasteners, and hands. Only then bring them to life. This order allows you to catch discrepancies before expensive video attempts.

Do not build the entire film as a chain of "last frame of previous clip → next." This is convenient for continuing action, but an existing error can carry forward. Returning to the main reference and approving each shot separately reduces the accumulation of deviations.

4. Do not confuse seed, reference, and training

Seed — a randomness parameter in a specific process, not a personal ID. The same number does not transfer a character between different models and settings. LoRA can help reproduce an image in a compatible process, but does not guarantee correct hands, materials, and movement. You should train it on materials you have the right to use, testing it with new poses and lighting. This is already a technical route, not a mandatory first step for a beginner. [33] [34]

5. Separate appearance from performance

An approved image is responsible for appearance, while a video with acting is responsible for movement and emotion. In the appropriate Act-Two or video-to-video mode, you can provide both sources instead of describing everything in one text. Check the result against the reference: the transferred emotion should not change the face. Working with a constant 3D model and edited keys was covered separately for 3D animators.

Acceptance rule: if the character looks like different people on the contact sheet, do not try to hide it with music or quick cuts. Go back to the design or regenerate the specific frame.

Beginner's route: 30 seconds, one fox, six shots

Educational fictional story "The Last Tram." A fox finds a lost ticket, sees that a passenger is already departing, folds the ticket into a paper airplane, and returns it through an open window. At the end, the passenger waves, and the fox remains at the stop, satisfied. There is a goal, an obstacle, and a conclusion — without five locations and long dialogue.

1. Write the story in one paragraph. Decide what the viewer should understand without explanation.
2. Create a hero profile and six sketches. Edit them as static images with rough sound — that's an animatic.
3. Check the most difficult part first. Here, it's folding a paper plane. If it doesn't work, replace it with two parts and a glue joint instead of wasting the budget on the rest.
4. Generate short shots from starting images. Each should have one main action. Don't promise yourself that the first take will be the final one.
5. Assemble the edit, add sound, and review without pauses. If the clip is unclear, rewrite the scene, not just the prompt.

The first successful result is a finished little story that an outsider understands. In the next project, you can add dialogue or a second location.

Professional AI workflow: approval, takes, and refinement

For professional work through generators, reproducible solutions are needed: an approved hero, a shot list, quality control, and clear corrections. The main pipeline: references → key images → motion generation → selection → editing and compositing.

1. Fix the images and scenes. Prepare standards for the hero, location, and lighting. Check the starting frames side-by-side before generating video.
2. Choose a mode based on the action. For motion from a ready image — image-to-video in Runway/Kling/Flow; for acting from a video reference — Act-Two; for transforming filmed material — Ray 3.2. Inputs and capabilities of the modes differ.
3. Generate and approve individual shots. Save variations with a short reason for rejection: face, hands, contact, camera, or timing.
4. Fix the required area. Try video editing, a different starting frame, a shorter fragment, or an edit replacement. Regenerating the entire clip might change things you already liked.
5. Finish in the editor. Editing, masks, color, sound, and subtitles join shots into a cohesive work. Determine the final file format based on the publication platform.

For handovers between people, save the frame ID, model version, exact prompt, input files, parameters, seed if available, and the decision regarding the take. Example naming: tram_sh040_v007_review.mp4. Do not overwrite an approved file with the next attempt. In a cloud service, even the same request might not be reproduced bit-for-bit after an update.

What to check frame-by-frame: jitter, limb length changes, extra fingers, object grip, silhouette, gaze, and lighting direction. Then — at playback speed, to notice flickering and unnatural transitions.

Specifically for 3D animators: which AI tools to add to your workflow

This block is needed if you are already working with a model/rig in Maya, Blender, or another 3D program and want to obtain motion or a scene for editing. To generate cartoon video, mastering these programs is not mandatory. The capabilities listed below are based on manufacturer documentation; we have not conducted a comparative test on a single rig.

Google Flow and Autodesk Flow Studio are different products. We consider the first for video generation/editing, and the second here for AI mocap and CG scenes.

What to try first for your task

• Walk, run, guided path: Use MotionMaker as a source of base motion in compatible Maya. Then — retargeting, checking feet and character performance.
• Your own recorded action for your own character: DeepMotion or Flow Studio. For a test run, take a short clip with a clearly visible body and check that it transferred to the required format.
• Posing and motion refinement: Cascadeur. AutoPosing predicts poses based on the animator's actions; AutoPhysics provides physical assistance rather than being a full movie generator.
• Already set 3D animation that you want to stylize: Render a short source and check Ray 3.2 video-to-video. The result of this stylization is a video; keep the original scene with the rig separately.

Production example: Pigmalion Animation Studio

In an Autodesk case study from September 11, 2026, the studio describes a specific application of Flow Studio: phone video → obtaining performance → motion on an existing 3D character → returning to Maya for further lighting and rendering work. This is an example of AI assistance in character animation production; it does not mean the entire film was generated from text. We do not apply speed estimates provided by the manufacturer to other studios. [43]

Practical test for an animator: Take 5–10 seconds of motion with a turn and hand contact with an object. Check scale, frame rate, skeleton orientation, root motion, and foot sliding after import. Successfully opening an FBX does not necessarily mean correct motion on your rig.

For unchanging geometry, use the same 3D model and rig in all scenes. Generative refinement of its render may change the appearance again, so comparison with a reference is needed here as well.

Unreal Engine 5.8: what can already be generated and animated with AI

As of October 24, 2026: Epic released Unreal Engine 5.8 on June 23. In the September training selection, the company separately covered AnimGen, motion capture from video, and agent operation via MCP. For an animator, this is a path to a controlled 3D scene where the model, cameras, and animation data are preserved. [44] [45]

Video performance → motion of that same MetaHuman

MetaHuman Animator Markerless Motion Capture in UE 5.8 creates body or body and face animation from a single-camera video. The plugin works on Windows, processing is local; status — Experimental. This is performance capture, not inventing an action from text. [46]

Official workflow: import recording via Capture Manager in Live Link Hub → perform solve in MetaHuman Performance Editor → export animation sequence → add it to the character in Level Sequence. The hero's geometry remains your 3D model; check contacts, feet, and fine facial expressions on the resulting animation. [46]

AnimGen: training a controller on your own animations

AnimGen — an experimental ML plugin for UE 5.8. It allows you to create and train character animation controllers in the editor and compute them during project runtime. Epic's tutorials cover custom training data, the Game Animation Sample, and integration with markerless capture. This is a technical tool for animators and developers who need data, settings, and result verification. [45] [47]

Editor Agent: MCP for working with the project

Built-in Experimental MCP plugin enables LLM systems to interact with Blueprints, levels, materials, meshes, and other editor systems. It helps automate work with assets and the project. The chosen model and connected client are separate components; the MCP itself is not a video generator. [44]

Useful for film and animation, but not AI in itself

In UE 5.8 Movie Render Graph received Production-Ready status, while Toon Shader for stylized painted animation — Experimental. Sequencer is responsible for staging and the timeline, PCG for procedural environment assembly. Rule-based generation and scene rendering are not the same as generative AI. [44]

Cinematic quality: not a list of words like "8K, epic, cinematic"

A cinematic look begins with answering: why the camera is placed here and what the viewer should feel? Motion for the sake of motion quickly turns a scene into a generator demo.

• One action and one emphasis. "The fox hesitates before jumping" is more useful for staging than ten effects in the same frame.
• Shot size must serve a purpose. Wide shows space, medium shows action, close-up shows an important object, extreme close-up shows a reaction.
• Light must have a source. A warm lamp in the foreground and a cold city background explain the color. After a cut, their direction must remain clear.
• The camera supports the scene. A static camera heightens anticipation; a slow zoom-in can bring a reaction closer. A continuous orbit is not needed for every dialogue.
• Write observable behavior. For example: "the gaze stops on the ticket, the paw freezes, then picks it up."
• Edit for causality. Gaze → object → decision. Changing scale without new information only wastes time.

Including a lens name in a prompt can suggest the desired image character, but it does not guarantee physically accurate optics or alignment with the 3D scene's camera. Similarly, upscaling to 4K does not restore a lost character identity or correct anatomy.

Ukrainian voiceover, lips, music, and silence

For your first work, you can record the voice yourself or with an actor: this makes it easier to direct pauses, emphasis, and specific emotions. If you need synthesis, ElevenLabs documents Ukrainian support. However, the presence of a language does not guarantee the correctness of every accent — check names, numbers, apostrophes, colloquial lines, and language switching. [29]

Keep dialogue, ambience, footsteps/objects, and music on separate tracks. It should be clear from the start who is speaking. During lip-syncing, check consonants, pauses, and whether the character starts "talking" before the phrase begins. Using someone else's voice requires proper consent; a recording of an acquaintance should not be automatically assumed to be permission for cloning.

The free ElevenLabs version is intended for non-commercial use with attribution; commercial rights are provided in paid plans according to the service terms. Voice rights do not replace rights to music or third-party video footage. [30]

DaVinci Resolve is particularly suitable for filmmaking: editing, color, compositing, and sound are all integrated into one environment. A free version exists, though some features require Studio. Other simple and professional options are covered in our separate comparison of video editors for PC and mobile. [31]

Free and paid tools: where the real difference lies

"Can try for free," "can produce regularly for free," and "can be used commercially" — three separate checks. All amounts below are estimates from official pages on 24.10.2026, not guaranteed prices for Ukrainian carts; taxes, currency, and billing terms may vary.

For your first project, it is more practical to take one video generator, one way to prepare images, and one editor. There is no need to buy five subscriptions just to test a complex shot. If you are working on a phone, browser-based generation reduces some local GPU requirements, but selecting duplicates and editing a longer story is often more convenient on a PC.

How much will a finished minute cost?

The monthly subscription price does not reflect this. Calculate generated seconds, rejected duplicates, images, sound, upscale, and post-processing. A hypothetical example, not a service tariff: for 30 seconds of final editing, four full versions of each shot, and a rate of 6 credits/s, it comes to 30 × 4 × 6 = 720 credits just for video. Generations longer than needed will increase costs even for images and sound.

One bad shot sometimes consumes more attempts than the rest of the video. Therefore, set a limit for duplicates in advance: once reached, change the setup, source, or method. You can track all expenses in the table below without uploading your materials to our site.

Four templates for your own film

These are starting tasks, not "magic prompts." They only work together with the appropriate model inputs. The English version for motion is provided as a convenient template; we have not verified language superiority across all generators. Replace the example names with your own files.

Help plan an original 30-second short animation. One main character, one location, no dialogue. Character: [my description]. He wants [goal], but [obstacle]. Ending: [what changes]. Propose 6 shots with a total duration of exactly 30 seconds. For each: ID, shot size, one main action, camera, sound, start and end state. Mark complex hand/paw/object contacts. If an action is hard to show, suggest a simplification through editing. Do not add new characters unless necessary. At the end, explain whether the story is understandable without text titles.

Input A — the main approved design of my original character. Input B — a reference for the desired style; it does not define a new face. Invariant features: [head shape, silhouette, proportions, material, hair/ears, clothing, colors, asymmetry]. Allowed changes: [pose, expression, lighting]. Prepare separate images: front, three-quarters, profile, back, full body, and two expressions. Neutral lighting, simple pose, clean background. Maintain the design from A. Do not invent new accessories. I will manually reject angles that differ in face or proportions.

Use the approved input frame as the visual starting point. One continuous medium shot of the same original paper fox at the tram stop. The fox notices the ticket on the ground, pauses, then slowly turns its head toward the departing tram. Keep the paper material, head proportions and teal scarf consistent with the input image. The tram moves from left to right. Locked camera. Warm light from the shelter, cool blue city light in the background. One restrained action, no cut. End with the fox looking toward the tram. Intended usable duration: 5 seconds; select an available duration in the tool and trim in editing if needed.

I will add a contact sheet of the approved character and my own shots/clips. Help find deviations: face, silhouette, clothing, materials, number of limbs, gaze, contact with objects, direction of movement, lighting, and editing causality. For each note, specify the shot ID and the visible feature; if you only have access to still frames, do not claim to have checked motion, lips, or sound. Distinguish an explicit error from a questionable area that needs human review. Suggest the minimal fix instead of full regeneration. Leave the final decision to accept with me.

How to prepare work you can be proud to submit to a festival

First, choose a festival and read its specific rules: definition of AI work, duration, participating countries and age limits, premiere status, subtitles, rights, and technical format. For example, AIF 2026 it required 3–15 minutes, the use of generative video, and a completed linear story; the form asked for a brief description of AI techniques. Its submission deadline was April 27, 2026 — this is an example of rules from a past cycle, not a currently open invitation. Our 30-second training video would not have fit such a duration. [5] [32]

• Final film: verified export, without accidental black frames, cut-off titles, or audio errors.
• Material package: synopsis, authors, poster, stills, subtitles, and a description of the process according to the specific event requirements.
• Rights confirmation: image sources, music licenses, voice and human likeness permissions, model and plan terms.
• Honest credits: what humans did, what was generated, which models and techniques were used. Do not call selection or submission a victory.

Tell the story to a person who doesn't know how you created it. If, after watching, they discuss the character's actions rather than just the generator's artifacts, you already have a useful direction for the next iteration.

Short answers before starting

Is it possible to make the entire cartoon for free?

Small tasks can be attempted with limited free generations, your own voice, and free editing. Regular volume, privacy, commercial rights, or required models may require payment. Local weights do not eliminate GPU costs.

Which neural network preserves a character best?

There is no single proven winner in this review. Compare reference modes on your own hero. For accurate geometry and posing, a stable model/rig is a more reliable foundation; the generative result following it must also be checked.

Do you need a powerful graphics card?

For cloud generation, the service handles the main computations. For local Wan/LTX, requirements are determined by the specific model and configuration; do not apply the requirements of a lightweight checkpoint to all others.

Can you get editable 3D animation from a finished clip?

Not automatically from the fact of video generation. Motion recovery or 3D scene reconstruction is a different task with separate tools and errors. If you need keys and a rig, include them in the production source.

What to do if hands and objects are constantly breaking?

Simplify the staging, split the action with editing, or create a precise motion source. For important contact, use traditional animation or compositing instead of endless full regenerations.

Does a festival tool guarantee festival quality?

No. The set of services in the credits shows the process of specific authors. Your result is determined by story, staging, material selection, sound, and polish.

Primary sources and fact-checking

Verified 24.10.2026. These are official references, company announcements, festival programs, and the film's author page. Manufacturer statements about 'best' quality are not used as an independent rating. Rates and availability change; Ukrainian prices are not derived from US ones without checking the cart. No affiliate links.

1. Runway and Lionsgate: partnership expansion, 11.06.2026
2. Netflix: Eyeline, VFX, and generative AI applications, 15.10.2025
3. HOLYWATER: authors and digital tools Distance Between Two Points Of Me
4. Runway: official AIFF 2025 winners
5. Runway: AIF 2026 winners and criteria
6. BIFAN: official 2026 festival results
7. Google: collaboration with filmmakers and launch of Flow, 20.05.2025
8. Google: new Flow tools with Omni 1.1 Flash, 27.08.2026
9. Google: Veo 3.1 Ingredients to Video, 13.01.2026
10. Google Flow: requirements, subscriptions, and models
11. Google Flow: supported countries, including Ukraine
12. Runway: Gen-4 Image References
13. Runway: Gen-4.5, text and image as input
14. Runway: performance transfer via Act-Two
15. Runway: rates and one-time free credits
16. Kuaishou: Kling 3.0, Omni, references, and storyboarding, 05.02.2026
17. Luma: Ray 3.2 Video to Video, structure, motion, characters, and limitations
18. Luma: plans, commercial use, and watermarks
19. Adobe: Firefly generation, composition, motion, and mutually exclusive parameters
20. Adobe Firefly: free and current paid plans
21. Midjourney: Edit Model for V8.1/V8.2 and up to four references
22. Midjourney: Omni Reference for V7
23. Midjourney: pricing and privacy
24. Midjourney: free trial limitations
25. Wan2.2: official repository, models, and Wan-Animate
26. ComfyUI: official Wan2.2 examples
27. LTX-2.5: open weights and multi-plane generation
28. LTX: license terms and organizational revenue threshold
29. ElevenLabs: supported speech synthesis languages
30. ElevenLabs: commercial rights for paid plans and free limitations
31. Blackmagic Design: DaVinci Resolve, editing, color, and sound
32. AIF 2026: official rules, deadlines, and description of AI techniques used
33. Hugging Face Diffusers: seed and reproducibility limits
34. Hugging Face Diffusers: LoRA and compatible training
35. Autodesk: MotionMaker, driven motion, keys, and Maya 2026.1
36. DeepMotion Animate3D: export formats, Freemium, and commercial terms
37. Cascadeur: which functions use machine learning
38. Cascadeur: free export limitations and plans
39. Autodesk Flow Studio: AI mocap and transfer to 3D editors
40. Autodesk Flow Studio: free tier, credits, and export
41. Reply AI Film Festival 2026: finalists and tools used, 28.07.2026
42. Odair Falek: author's description of Cornered and tools used, 2026
43. Autodesk: how Pigmalion Animation Studio uses Flow Studio, 11.09.2026
44. Epic Games — Unreal Engine 5.8 is now available, 23.06.2026
45. Epic Games — September learning: AnimGen, markerless capture, MCP, 30.09.2026
46. Epic Games — Animation from Mono Video Capture: UE 5.8, Windows, Experimental
47. Epic Games — AnimGen, official tutorial series
48. HOLYWATER / Illya Dutsyk — Distance Between Two Points Of Me, YouTube
49. Robert Gaudette — A Face Only A Mother Could Love, YouTube
50. Jacob Adler — Total Pixel Space, YouTube
51. Epic Games — MetaHuman markerless motion capture demo, YouTube

Editorial principles and correction notice.