Source: raw/Claude_Makes_Insane_Videos_Now_It_Can_Replace_Seedance.md
Creator: AI Samson (channel confirmed from YouTube metadata) · URL: https://www.youtube.com/watch?v=tPiEPzy2JKU · Platform: YouTube · Duration: 23:10 · Uploaded: 2026-11-17 · Transcript fetched: 2026-11-17 · Sponsor: Higgsfield (disclosed; the reskin segment)
A creator’s method for making narrative animation in an ordinary Claude chat with Fable 5.2: Claude writes a program that draws every frame, renders it and returns an MP4, with no video model involved. His case for it is control, repeatability and character consistency, not price. He then pairs it with a video model: the coded animation becomes a video reference that Seedance 2.5 reskins at 4K. Despite the title, his own answer to “will this replace AI video generation” is no. All quality judgments, timings and costs below are one creator’s.
Key Takeaways
- What code buys you. Every detail is addressable, so you can change a character’s dress and location and keep her otherwise identical. The same code renders the same video every time, where a video model returns a different clip per run. There is no length cap; he says most video models stop at about 30 seconds.
- No GPU, no video subscription, no coding. He says every film in the video was rendered on a small cloud machine with two processor cores and no graphics card.
- The pipeline Claude runs: write an animation engine, render a few stills and fix weak ones (up to three rounds), render every frame in a hidden Chrome session, synthesize music and effects in code, then join frames and sound into an MP4 with FFmpeg.
- The engine is the asset. Once Claude has built an engine for a style, the same code makes new scenes, characters and stories. He calls this a reusable animation pipeline for a series or a channel with one visual identity.
- Fable 5.2 gave him the best results and costs the most. Opus 5.5 is much cheaper but “for many styles, it doesn’t create as good output.”
- It is not always cheaper than a video model. His cheapest Opus 5.5 videos cost less than “pretty much any AI video model”; his most detailed Fable 5.2 anime cost more than Seedance 2.5. He spent more than $400 on about 50 to 60 videos.
- It is slow. Building the approach for a new style can take up to an hour, before rendering.
- Use the hybrid for polish. Give the coded animation to Seedance 2.5 as a video reference: Claude has fixed timing, camera moves and layout, and the video model adds texture, lighting, grain and smoother motion.
- Where it fails: photoreal live action, realistic facial acting, and complex motion such as fights or specific dance moves.
How to run it
- Open a new Claude chat on Fable 5.2. He says it works on other models and platforms but Fable 5.2 was best in his tests.
- Paste a style prompt and add your idea. His prompts (a file linked in his description, not fetched here) define the whole style and leave a slot for the idea. His example idea: a boat on Lake Como seen from a “platform view” (as captioned), a small character in a red headscarf, the boat moving slowly across the scene.
- Let Claude build and render. Rendering can run in Claude’s cloud machine or locally. Local is faster but ties up your computer.
- Iterate in words. Ask it to extend the film or change a detail.
- Start a new chat for each new piece. He does this to hold cost down, because Claude otherwise rereads the whole conversation.
Making your own style
- Find a reference with a strong look that is not overly detailed (he browses Pinterest).
- Do not use the artist’s image directly. He generates his own variation in ChatGPT’s image model (he calls it “images 2.6”) and gives that to Claude.
- His reference prompt takes variables for length, aspect ratio and any specific action.
- Storyboard grid. Ask the image model for a 2x2 grid of the character moving through a story (close-up, medium shot, hands on a coffee, a dog at her feet), paste the grid into a new Claude chat, and ask for an animation that leads from frame to frame in order.
Four ways Claude animates, by style
| Method | How it works | His examples |
|---|---|---|
| Flat layers | Flat shapes moved on stacked layers | Marker films, pastel, woodblock |
| Painted background plus drawn characters | Still background layers; characters built from separately animated shapes | Anime |
| 3D scene | A full 3D scene with a camera, lit and filmed with depth of field | 3D brick style |
| Image puppet | A finished illustration bent through small masked areas | His “afternoon light” film |
He also shows a documentary-style dinosaur film where Claude built the scene in Blender and rendered it to video.
Styles he says suit short-form channels: short anime, quirky 3D for comedy or explainers (physics, space, maths, history, engineering), stick figures, woodblock, Japanese painterly, paper collage, children’s illustration, chalk pastel and pen drawing.
When to use which
| Need | His pick |
|---|---|
| Explainers, stylized animation, motion graphics, precise control | Coded animation in Claude |
| Realistic human movement, complex action, cinematic live action | A video model such as Seedance |
| Long narratives that must stay consistent “across hours” | Coded animation |
| Professional finish on a stylized piece | Both: code for structure and movement, video model for final look |
The Seedance 2.5 reskin (sponsored segment)
Vendor and sponsor claims; he uses Seedance 2.5 inside Higgsfield.
- Setup. Connect the Higgsfield MCP so Fable 5.2 can call video generations from the chat.
- Steps. Download the coded animation, drag it into a new chat as a video reference, paste his reskin prompt, name the video model, and match the output length to the original.
- Why 4K, in his words. Claude’s output is clean flat colour and hard lines; Seedance adds brush strokes, film grain and fine line work that lower resolutions smudge. He prefers generating at 4K to upscaling, because an upscaler has to invent detail, and 4K leaves room to crop to a close-up or a vertical cut.
- Length and sound. He says Seedance 2.5 renders up to 30 seconds in one generation with sound generated at the same time, which covers for the coded soundtrack he calls “a little bit lackluster.”
- Hit rate. With the reference fixing composition, he says you get the look you want in fewer generations than prompting the video model from scratch.
- Results he shows: a flat-shadow portrait, a brick-style scene and a classic illustrated style, each smoother and more detailed after the reskin.
- A voice model can be called the same way for voiceover.
Cost notes
- Cost depends on the model and on how often Claude rereads the conversation.
- Reuse the engine for each new film in a style.
- Try a cheaper model. He shows the same video in Opus 5.5 and Fable 5.2.
- Ask Claude to make it cheaper. It can run fewer self-checks.
- He compares finished cost with a video model’s cost per attempt: a video model rarely gives the wanted take first time, so its real cost is several generations.
- He says you can experiment on Claude’s free tier. No per-video dollar table is in the transcript; he refers to one on screen.
How this fits the wiki
- Claude Motion Graphics — Three Methods records a creator’s forecast that AI video will be built in code first and re-rendered with a video model at the end. This video is a worked version of that idea.
- Futurepedia’s code-drawn explainers use the same chat-only route for motion graphics and report Sonnet matching Opus when the style prompt carries example code. Here the subject is character animation, and the creator finds the model tier matters. Different content, and both are single-creator judgments.
- The wiki’s August notes put Seedance 2.5’s practical output at about 10 seconds at 720, with 30 seconds as a costly option (the August 2026 video-model wave). A 4K Seedance 2.5 on Higgsfield is new in this source and rests on a sponsored segment.
Try It
- Pick one flat style (marker, paper collage or woodblock) and ask Fable 5.2 in a new chat to build a 15-second scene with one character and one camera move. Ask it to render stills first and show them.
- Change one detail in words (outfit or location) and re-render to confirm the character holds.
- Ask Claude to save the engine code, then start a new chat, paste the engine and request a second scene in the same style.
- Run the same prompt on Opus 5.5 and compare quality against usage.
- If the result needs more polish and you have Higgsfield connected, pass the MP4 as a video reference to Seedance 2.5 at matching length. Confirm the credit cost before generating.
Open Questions
- The creator and the prompt file. Neither is in the transcript; the style and reskin prompts were not fetched.
- Actual costs. $400 for 50 to 60 videos is the only figure; per-style costs are in an on-screen table.
- Seedance 2.5 at 4K. Availability, credit cost and the 30-second limit on Higgsfield are sponsor-segment claims with no first-party source here.
- Example credits. He shows a robot anime by a poster the captions render as “Chassloo” on X and a CodePen version; neither was checked.
- How the chat sandbox renders. He describes a hidden Chrome session and FFmpeg on a two-core cloud machine; plan and usage limits for long renders are not covered.
Related
- Claude Motion Graphics — Three Methods — video model vs code vs After Effects.
- Code-Drawn Explainer Videos in Claude Chat (Futurepedia) — the chat-only code route for explainers, with style prompts.
- Higgsfield MCP — calling image and video models from inside Claude.
- Claude + Higgsfield Motion Graphics — six Seedance 2.5 builds through the same MCP.
- Video-to-Video Trigger Editing — using an existing video to drive a generated one.
- The August 2026 Video-Model Wave — Seedance 2.5 background.
- Claude Dashboards and Motion — Anthropic’s hosted code-to-video artifact type.
- Claude Fable 5.2 — the model used, and its pricing against Opus 5.5.