Works and making / AI video
Start with real finished works across comic drama, short drama, film, music video and commercial video, then follow production, distribution and business decisions.
Start with the work
Start with a serial drama, a film, a commissioned ad and a merchant's short. Each card identifies original imagery or an editorial illustration.
Watch original film ↗Commercial videoCase study
An unmanufactured plant-pot insert needed to become understandable. The maker delivered a story with a fictional apartment, character and see-through pot shot. The images explain a proposed design; they cannot prove how a physical product performs.
Watch season one on Hongguo ↗BitShovel illustration · not film footageAI comic dramaCase study
See why viewers return to the finished series, what its makers chose, what audiences questioned and how to run a small original test.
Watch original film ↗AI filmCase study
Some generated shots looked beautiful but made the character move unlike himself; the director discarded them.
Watch original film ↗Commercial videoCase study
An eyewear brand needed a younger audience and had four weeks to launch. An official case film, an unchosen concept, the production method and media results are public; a standalone complete ad and financial ledger have not been verified.
Watch original film ↗BitShovel illustration · not film footageAI comic dramaCase study
This AI comic drama has a traceable audience result. Its clearest production account is that the lead character's final look took repeated human revision.
Watch original film ↗AI short dramaCase study
An earlier series that did not take off, a trailer that attracted attention, and the difficulties of keeping a serial story on screen.
Search the platform for the film ↗AI comic dramaCase study
Its reported commercial result comes with a specific production trail: choose a story people still discuss, establish rights and characters, then organize ten people to deliver fifty episodes.
Open a related finished video on the brand's Douyin account ↗BitShovel illustration · not film footageCommercial videoCase study
The product became the story's reversal. Selection, rework and audience remixes shaped what the team could make next. The sales gains are maker claims without independent orders or a profit ledger.
Discover new work and continue earlier cases
Production evidence
These maker materials explain production. They do not automatically establish an audience, orders or income.
Watch the creator's 2:53 progress film on YouTube ↗AI filmProcess sample
Watch the girl, grandfather and redwood story, then inspect character inputs, a missed check and targeted rework. The portraits are references; the full soundtrack has not been reviewed here.
Watch original film ↗BitShovel illustration · not film footageAI filmProcess sample
Watch the film, then trace the shared character file, environments, storyboard, twelve-second four-shot prompt and edit decisions.
View original work ↗AI short dramaProcess sample
How a real sample was made from source text, storyboards and generated clips.
Watch original film ↗AI filmProcess sample
This team drew its characters by hand, used AI to animate the world around them, and joined the two shot by shot.
View original work ↗AI filmProcess sample
Open the finished film, compare three real storyboard directions, then trace the falling-dish shot from anchor frames and video prompt through the edit.
View original work ↗AI filmProcess sample
Open the complete demo file in the original repository, then compare the two rows of Shot 05 frames: how one revision balances the gate impact with a readable full-body hero.
Watch original film ↗BitShovel illustration · not film footageCommercial videoProcess sample
Start with the film, then trace the pour still, the maker's account of the failed motion, the ice-drop redesign and the edit seam. The fix was a new shot design, not just a revised prompt for the same pour.
Watch original film ↗BitShovel illustration · not film footageAI music videoProcess sample
Watch the finished video, then connect the maker's successful and failed clips to casting, lyric planning, framing, generation and editing.
Watch original film ↗AI filmProcess sample
A maker with media-production experience spent about 72 hours and $390 in tools on this film. Watch the result, then trace script, generation, voice and manual editing. Rework in continuity, pronunciation and on-screen text is documented in his account; failed clips are not public.
Watch official trailer ↗BitShovel illustration · not film footageAI filmProcess sample
Start with the trailer and festival listing, then follow the manager's account of visual anchors, shot review, rerouting failures and the limits of the release evidence.
Watch original film ↗BitShovel illustration · not film footageAI filmProcess sample
Start with the film and the director's published character and location material. See which shots demanded rework and which accidental angle survived. The reported inputs describe effort, not profit.
Watch original film ↗Commercial videoProcess sample
Watch the ferry cross through time, then return to the albums, family visit and creative choices: how a local business can turn its own memory into a distinctive film.
Watch the complete film on the maker's page ↗Commercial videoProcess sample
Watch the finished work, then follow how the maker fixed identities, revised a shot's landing and passed a frame from one sequence to the next.
Watch the official film and read its production account ↗Commercial videoProcess sample
Start with the visible result, then examine how shots, captions and sound can be changed separately.
Field research · past, present and what to test next
Models expand the shots people can make. A work still takes human choices about story, rights, production, revision and release. Market value adds another question: who pays, who receives the money and what the full costs are.
Clips, image and sound
Trace capability changes →Rights, script, shots and rework
Inspect a named project →Finished film, release and response
See response to a serial →Payment, commission or product sales; costs separate
Separate payer from recipient →These are public research and tool milestones from different teams. They do not establish which models any case on this site used.
Meta's Make-A-Video research demonstrated short clips generated from text. A clip is not yet a finished serial story.
Meta research announcement ↗OpenAI's Sora report showed videos up to one minute and noted limits in simulating physical interactions. This was a research report at the time.
OpenAI technical report ↗Google announced Veo 3 with native sound effects, ambience and dialogue. Editing and story choices remain production work.
Google I/O announcement ↗Wan2.2 released inference code and model weights, giving equipped developers an inspectable entry point. Open code does not establish its use in any particular film.
Wan2.2 original repository ↗Generation, delivery and business are different outcomes. Check each against a named work.
Tradecraft publishes a wrong character, revised description and four-shot opening prompt. Input and result can be compared; the full number of retries remains unknown.
Read the maker's prompt →Wanyao Tulu Zhuan has released seasons and visible audience feedback. Its makers discuss story choice, shots and early discarded material, but not shot-level inputs or costs.
See the work and response →The Tomorrow Is Monday team has discussed payment and cost recovery, without inspectable settlement records. Wanyao's platform heat cannot be converted into income either.
Check the payment claim →Over the next year, these are BitShovel's two questions to test against new evidence, not forecasts of income or release dates.
Serial rework may become cheaper. Compare usable-shot yield, total time per shot and character consistency across episodes under comparable scripts and quality thresholds, not just model demos.
Business paths become comparable only with contract or revenue-share terms, receipts, and production and promotion costs. Views, platform heat and reservations still measure attention.
Research assembled 26 September 2026. Model milestones link to original announcements; read case figures with each case's own checked date.
Serial works, character and shot consistency, audiences and sustained production.
Watch season one on Hongguo ↗BitShovel illustration · not film footageFeatured work
Wanyao Tulu Zhuan: choosing a story, handling discarded shots and building a seriesHongguo displays 49.59m heat and 81 episodesPlatform response · season one · Check result sourceFrom script and storyboard to finished video, release tests and the next episode.
Featured work
The Accident: how two students built a serial drama after an earlier attempt stalledFour episodes, roughly 30–40m views and over 2m likesReported audience response · interview date · Check result sourceLong-form storytelling, direction, compositing, sound and exhibition.
Featured work
A Face Only A Mother Could Love: keeping a character intact across an award-winning AI filmRunway AI Festival Film Grand Prix; $50,000 cash prize listedFestival result · 2026Check result source9 published cases
Process sample · Migration: bringing hand-drawn characters and AI backgrounds into one film →How a song becomes a watchable performance through characters, lip sync, timing and edit decisions.
Documented process
Welcome Home, Darling! — a finished AI anime music video and the takes that failedOne finished animated MV and three public process clipsDocumented outcome · personal work · Check result source1 published case
Client briefs, delivery, variants and payment in ads and branded films.
Featured work
Pit Viper × DEPT: from a brand problem to an AI ad campaign10.3 million reached and $2.94 CPM (DEPT-reported)Campaign performance · agency-reportedCheck result source7 published cases
Process sample · Born From Fire: when the pour failed, the maker redesigned the shot →Begin with one documented production choice
Watch the 61-second Tradecraft film, then inspect its character mistake and the maker’s original four-shot prompt for its twelve-second opening. Try a short segment with material you may use, recording your input, candidates and final choice. This is a BitShovel exercise, not the maker’s tutorial.
From attention to a payer
An audience is a reason to investigate. Economic value still depends on who pays, who receives the money and what a published figure actually measures.
Published case · Tomorrow Is Monday: taking a rights-cleared story to a paid 50-episode comic drama
Mingri Zhouyi turned a licensed story into a 50-episode platform series. The maker's interview connects story choice and team production to a reported paid-viewing result.
Published case · Pit Viper × DEPT: from a brand problem to an AI ad campaign
Pit Viper commissioned DEPT for a brand campaign. The team published the brief, creative choices and campaign results; a larger media budget is not an agency revenue figure.
Published case · Snow Mountain Fox: how a duck became a story and a store visit
Chuanwa Sauce Duck put its own product into the Snow Mountain Fox story twist and published a film on its brand account. The maker reports store-sales changes, but the product, store and comparison periods differ or remain unspecified; the film's contribution cannot be isolated.
How a work reaches its audience
Use this map to locate questions. The cases supply the actual decisions and results.
Who is it for, and why did the client or creator choose this idea?
Can the source be used, and what becomes a filmable scene?
What did models and tools enable, and what did the makers decide?
How do characters and scenes remain recognizable across shots?
Which errors occurred, and how were they found and corrected?
How are shots selected, trimmed, voiced and delivered?
Where is it seen, and what does the response change?
Who pays, and are costs and receipts documented?
Shared in public
Creator samples, open projects and process notes are grouped by use. Check each item’s terms before reusing it.
Who is it for, and why did the client or creator choose this idea?
Public source and usage notes · Planning to final video
The project connects planning, writing, storyboards and output as another workflow to inspect; no case here is claimed to use it.
Start here: Trace how one script becomes shots and tasks, then read the repository's commercial terms; do not attribute this workflow to a BitShovel case maker.
Can the source be used, and what becomes a filmable scene?
Open pipeline and maker's sample film · Script, staged assets, storyboard, voice and render
The maker provides a playable AI Hot News sample and describes chapter adaptation, stage-aware character and scene assets, storyboards, voice and LTX/ComfyUI rendering. Its first/last-frame vision check may accept a flawed clip once the retry limit is reached. The repository does not publish per-shot failed clips, actual production bills or revenue.
Start here: Watch the repository's Demo 1 film, then read its five-stage process and retry limit. Treat it as an output and method sample, not evidence of per-shot success rates.
The pinned October 22 README separates reference-image review and a short warm-up from reviewing generated footage. Its migration record lists 61 completed clips and three queued while remote H3 and whole-film assembly remain unverified. Existing files and local mocks do not establish a complete film from the new backend. We have not run the project; the original Demo 1 retains its original-version context.
What did models and tools enable, and what did the makers decide?
Official code, weights and examples · Image and text to video
The official repository lists models, examples and runtime requirements; compare version and compute needs before trying shots.
Start here: Check the chosen weights, memory needs and example inputs, then try one short shot of your own; the official demo does not establish which model an existing case used.
How do characters and scenes remain recognizable across shots?
New format source and production conditions; not our film recreation · CAD input, coverage, narration, cuts and design checks
The October 22 promo format prioritizes Blender renders of the actual CAD, then word-timed cuts, design checks and mixing. An October 23 commit continues fixing per-take design checks and other issues; a command does not establish automatic acceptance. The maker links a SITKA promo. We inspected the fixed project, format and changes, without fully auditioning that film or closing The Oldest Tree’s audio follow-up.
Start here: Start with the fixed format contract, then compare sitka_demo’s promo.json, coverage.json and design.json. Separate inputs, plans, checks and human approvals. CAD GLBs and environment assets are not committed with the template; supply assets you may use. The example is not a complete one-click reproduction bundle.
Open workflow and maker's failure log · Prompt, storyboard frame, failure diagnosis and revision
For City Dating Show EP01, the maker logs failed and passing prompt revisions for grids 14, 25, 27 and 33 and distinguishes InvalidNode, moderation failure and a no-URL timeout. This is a written account without every original frame or a playable final film. Its character threshold comes from a small local test, not a universal platform rule, and no audience or revenue evidence is supplied.
Start here: Read grid 14's three prompt versions and reported results, then the three error signals at the top. For your own failed frame, keep the input, error signal and output before deciding what to revise.
Maker-shared film, breakdown and production files · Story, characters and sets, prompt, edit
The maker shares his 61-second Tradecraft film, a breakdown and six planning files. A celebrity head-shape analogy produced an overly muscular character, so he moved to direct facial description; the Seedance file gives the actual four-shot prompt for the opening 12 seconds. He reports 155 saved outputs, including 23 videos, and under $50 of Weavy credits. That excludes other subscriptions and labor, and no income is documented.
Start here: Watch the film and the wrongly generated character in the breakdown. Then open the character prompts and the four-shot Seedance opening file; compare descriptions, reference-image roles and shot limits. The 155 figure counts saved outputs, not all attempts or a success rate.
Open production, whole-film rework and handoff method · Shot approval, whole-film review and recorded decisions
The maker shares the 2:53, 49-shot The Oldest Tree progress film, character inputs and repair scripts. The published case uses three portraits to explain adult redesign, the refs/identity gap, and three shots restarting from stills with two at animation. The v3 is explicitly linked to its YouTube ID; project reports retain failures and unchecked items, and full audio remains unreviewed.
Start here: Read the case's inputs and repair choices, then check the README's Mac, Python, ffmpeg and model-stack requirements. The starter .sf route differs from the full judge/director pipeline; --dry parses only and produces no film. BitShovel has not run this generation workflow.
Maker's prompts and rework review · Prompt, review and pickup shot
The maker shares the original prompt for Rainy Night's second segment and a nine-second pickup prompt, recording an unexpectedly opened umbrella, soft ending focus and wardrobe drift. The review distinguishes reshooting, grading, trimming and accepting flaws. Playable failed and revised clips are absent, so this is not a controlled visual A/B test.
Start here: Compare the original and nine-second pickup prompts, then read the decision for each flaw. For your own shot, record the input, failure, acceptable limit and retry cost.
Maker's public video loops and prompt library · Shot design, prompt and generated result
The maker groups 25 camera moves, including pans, zooms, tracking and orbits. Each loop sits above the camera prompt used for that Seedance 2.0 result. It shows how distinct shot intentions are written, not a controlled one-word A/B test on the same input; per-shot failure rates are absent.
Start here: Choose a static, dolly or orbit goal, watch its loop and read the prompt. Try one move with a start image you may use, then log the actual result, reruns and cost.
Earlier work, spatial references and camera prompt · Spatial and character references, camera choreography and generation
Published August 2 and updated August 3, the maker's account gives aerial views, the driver's seat, cabin props and camera path distinct references and timed instructions. He reports 22 references and a 30-second plan; the linked YouTube upload, NY Road Rage, played continuously muted to its roughly 28-second end in this review. Image approval and prompt revisions precede the reported single video generation; the final generation and its earlier preparation and rework have separate scopes.
Start here: Watch the path from aerial view to rear, front and cabin, then compare the driver's-seat reference with landmarks such as headlights and window in the prompt. For references you may use, specify who sits where and where the camera passes. The maker plans with Claude Code, generates with Seedance 2.5 in Dreamina and upscales with Topaz inside Higgsfield; these are this project's conditions, not settings that guarantee identity or camera accuracy.
Maker's film, rework account and credit figures · Product reference, video generation, failed-shot cutoff and edit
The maker shares an uncommissioned Corona spec ad, boards and keeper/failed prompts. He reports just over 900 Higgsfield credits in all; six failed attempts at a twirl-to-overhead transition consumed 270 before he stopped. The credit figures are his account of billing history, with no displayed bill; they do not establish project profit or client income.
Start here: Watch the film and read the failed-transition account, then inspect the two prompts marked FAILED. Set a retry limit for your own experiment and log stills, video, upscale and discarded runs separately.
Maker's documented process · Storyboard to generation and edit
Three storyboard directions, shot prompts, six clips and a final film; one shot has traceable anchor frames and output, with documented stitch fixes.
Start here: Compare the three storyboard directions, then trace Shot 04 from anchor frames and prompt to output and the documented stitch fix.
Open workflow with real output · Script to edit
Public script, continuity plan, storyboards, prompts, checks, edit sheet and a 24-second sample let you trace one shot.
Start here: Pick one shot in the public EP001 example and follow it through script, visual plan, storyboard and video prompt; check identity and duration before generation.
Open production tool · Characters to voice
The public project covers script, assets, storyboards, video and voice; external model calls may add cost.
Start here: Use one scene to check whether character assets, storyboard and video tasks preserve the same decisions before estimating external model calls.
Which errors occurred, and how were they found and corrected?
Product entry · Code animation beta · Text/data, code animation, revisions, MP4
Continue editable code animation by changing words, figures and timing before export. Compare revision routes with HyperFrames and Manus; this is not native footage generation or a fixed Opus model.
Start here: Eligible Team/Enterprise users can try a short explainer from their own report, check figures and one revision, then export. Enterprise must enable access; other plans are not included.
Maker's original process and storyboard choice · Concept, approved stills and shot generation
The maker shares a five-beat structure, full prompt, discarded pour still, replacement ice-drop still and final film. This was an uncommissioned brand spec ad; the stills show a changed shot design, not a same-frame prompt A/B test.
Start here: Compare the pour and ice stills, then watch beat four of the final film. If a hand action repeatedly fails, reconsider the shot design before more rerenders.
Maker's process and frame comparison · Generation, review and targeted revision
Five first-take and five selected-cut frames from the same climax shot, plus the reason for revision, prompt, final clip file and frame review. The full first take is not public.
Start here: Compare the two sets of original Shot 05 frames and the maker's reason for revision before viewing the selected cut; stills only establish what is visible in those frames.
Official downloadable workflow and guide · Shot input and generation
The official guide links workflow JSON, nodes and model files so you can inspect the inputs of an image-to-video run.
Start here: Open the image-to-video workflow JSON and check its reference image, model files and nodes before running a batch.
How are shots selected, trimmed, voiced and delivered?
Maker's short recording demo and open-source editor · Recording, transcript, cuts and undoable edits
The maker's short demo strikes out a false start in a recording's transcript and shows red cuts and blue zoom regions on the same timeline. Current code resolves each word's unique ID to its source-audio times and adds a trim; in-app chat and the editing MCP share the executor, with undo in the open project. Automatic zooms instead follow recorded clicks without a model call. The making-of dates to September 27, and v2.0.0 was released October 3. The October 13 source changes add inserting other video assets, duplicating clips, replanning zooms from recorded clicks and per-clip cropping. October 14 code fixes the MCP GIF-export failure caused by passing a video-quality parameter. October 16 source fixes cover cursor and clock progression while holding the last frame, whole-segment copying and Intel Mac texture synchronization; replay the actual project to check the result. Version 2.1.0 was released October 17, followed by formal v2.1.1 early October 18 Beijing time. Its source fixes clipped corners on tilted zoom cards under a filled format, along with blur-mask mapping and motion blur in that scene. The guide retains whole-turn rewind and project-targeted MCP operations. Replay, inspect masking and export to verify the fixes; BitShovel has not installed the new release.
Start here: Watch the homepage's Describe the edit demo and AI editing guide, then use a short recording you may use to check a cut, playback and undo. Windows, macOS 13+ and Linux are supported, with system recording permissions; first transcription downloads approximately 264 MB for Whisper and under 1 MB for voice detection. GPU-assisted English or French word timing separately downloads about 110 MB or 350 MB. Built-in chat needs your model key. For an external agent using the in-app HTTP MCP server, separately enable its initially disabled MCP edits switch. Its 2.1.0 guide allows naming another project; whole-turn Rewind restores project, conversation and agent state together. Confirm the target project before editing. Ordinary recording, manual editing and export need no model key. The source reviewed October 15 routes transcription through ffmpeg and reports a missing dependency clearly. On macOS 15.2+, naming a window through the in-app recording MCP separately requires Screen Recording permission. If system-audio setup fails, recording can continue with a warning but without system sound; a completed take does not establish complete audio. When using the latest source tools, read the project back first: neighbouring clips from the same asset merge when their source ranges meet and crops match. Set range and crop together when changing framing; the new fix prevents a range-first merge from losing the crop. Use trim ranges for pauses; asking to split a clip does not create two independent objects in this model. The latest guide also clarifies dropping: video on the clip row appends to the end; audio enters through the timeline’s Add audio menu. There is no split-and-insert dialog, and the agent’s insert action is a separate entry. For MCP GIF export, use GIF size and frame-rate settings; video quality belongs to MP4. A later development fix read October 20, 1e917109, lets a custom zoom scale avoid a conflict with the prefilled depth; omitting both still defaults to depth 3. Formal release remains 2.1.2 and we did not install it. The later development commit read October 20 downloads transcription and alignment models through Electron’s network stack to honor system proxies. This is the 3c9bd5d source fix; formal release remains 2.1.2 and we have not installed it. The development guide read October 19 adds a stdio demo path: an agent starts the app with local-user rights, without the MCP edits switch. New tools load in a new agent session, and word cuts still need transcription. When the homepage badge cannot copy its prompt, it now opens the relevant guide section for manual selection. These are current-source/documentation conditions; the formal release is now 2.1.2, published October 19. Its change includes stdio MCP startup in anonymous usage counts, not evidence of improved edits; we have not installed it. The d8bf2d8 development guide checked October 22 fixes Windows Store discovery: before an execution alias ships, use app/Openscreen.exe under its Get-AppxPackage install location and re-register the versioned path after updates; avoid a PowerShell stdio wrapper. If tools are absent just after MCP setup, retain the task and load them in a new agent session. This is documentation; formal release remains 2.1.2, without our installation test. The 2159f42 development fix checked October 25 records voiceover against the uncut timeline and applies cut mapping at edited playback. A short project with one cut can check recording, replay and undo separately. Formal release remains 2.1.2; this was source review without a recording trial. Replay and undo once to check the cuts, framing and zooms.
Open desktop-editor entry · Early release · Media, agent/manual edits, checkpoints, rendered review
An independent desktop app begun from a HyperFrames snapshot, with local project files, a manual timeline and per-prompt rollback checkpoints. Agents call the chosen model provider with prompts, transcripts and sampled frames, and asset research uses the web. Local projects do not imply fully offline or zero-cost work. Render QA is the maker's feature description, not our replication.
Start here: Read the v0.5.4 requirements: Apple Silicon Mac or Windows x64, Chrome, FFmpeg and your own model-provider connection. Builds retain OS signing limitations. A suggested trial uses your own media for one edit and rollback; BitShovel has not installed or tested the editor.
Who pays, and are costs and receipts documented?
Maker's failure account and shared project · Storyboard, keyframes, generation and budget
The maker shares a 14-shot storyboard, per-shot prompts and revised keyframes, with reported character drift, unwanted text and iteration costs. The stated total is about $120, while rounded line items add to about $125. The repository has no complete film or failed clips to verify motion or market results.
Start here: Inspect prompts and revised keyframes for shots 4, 6, 10 and 12 before reading the budget account; set a limit for continuity fixes and retries.