AniSora: Open-source anime video generation model

Original Article ↗ Hacker News Discussion ↗

Model access, safety & format

  • Commenters initially struggled to find “open source” materials; others linked the Hugging Face repo with weights.
  • One checkpoint file is flagged as unsafe by scanners, triggering concern about malware via .pth / pickle-based checkpoints.
  • Several argue this is likely a false positive but advise caution and advocate for safetensors and diffusers formats as industry standards.
  • A diffusers conversion is already available, and at least one web UI (SD.Next dev branch) supports the model.

Quality, artifacts & capabilities

  • Testers report visually impressive results, but with clear temporal artifacts: hair flicker, disappearing details, clothing glitches, and limited actual motion beyond simple pans and limb movements.
  • Some examples are criticized for obvious glitches even in showcase clips.
  • The underlying Wan2.1-14B base leads to questions about frame rate (e.g., whether it’s locked to 16 fps).
  • Paper notes training on 2–8 second clips at 720p; one user wants head‑to‑head comparisons against FramePack for longer 2D sequences.
  • Unclear whether the model can maintain a consistent character across multiple scenes and angles, a known weak spot of current gen.

Naming, branding & web infrastructure

  • The “AniSora” name is widely assumed to be playing off OpenAI’s Sora; others point out “sora” is a common Japanese word/name.
  • Some note OpenAI’s sora.com now redirects to a subdomain, leading to side-discussion about cookies, cross‑domain auth, and ad‑tech.

Copyright, training data & legality

  • Many assume the model is trained on copyrighted anime, manga, webtoons, and Pixiv-style art.
  • Debate centers on whether Bilibili’s distribution licenses imply any right to train models; some say it’s analogous to Crunchyroll releasing a model and being pressured by licensors; others argue “China doesn’t care about licenses.”
  • Broader point: almost all major models (including Western ones) are suspected of using copyrighted material; precedents (e.g., Meta’s book data) are cited.
  • Several note that the legal “right to train” is unresolved; enforcement is seen as effectively “pay to play” favoring large firms.

Impact on artists, copyright theory & “what is art?”

  • Long, nuanced debate compares visual artists to translators: both transform prior works/inputs, but artist outputs are clearly copyrighted; translators’ status and creativity are contested.
  • One side argues:
    • AI training consumes huge corpora of protected work and directly undermines illustrators, especially commercial ones (gacha art, light novel covers, etc.).
    • Mass AI use risks collapsing markets, reducing incentives for new styles, and flooding the web with low-quality “AI slop.”
  • Others respond:
    • All creativity is derivative; humans are also trained on massive “datasets” of lived experience and prior art.
    • Copyright is already messy and overextended; some advocate reducing or even abandoning it, or redefining derivative use for models.
    • The real harm is not exposure to copyrighted work, but models that can reproduce specific works or identifiable individual styles at scale; technical solutions to avoid memorization are proposed.
  • There’s disagreement on whether AI outputs can be “art”:
    • Some insist art requires human intent, expression, and a personal creative journey.
    • Others say if AI-generated work successfully evokes complex impressions and is shaped by human direction, it functionally is art; tools don’t negate artistry.
  • A common prediction: AI devastates the “bottom half” of commercial work (cheap illustration, junior roles, penny‑dreadful novels), while high‑end or deeply personal art remains but becomes more niche and/or luxury‑like.

Anime industry & content ecosystem

  • Several see this as enabling “infinite anime” (fan continuations, AMVs, fan seasons for series like Haruhi or Solo Leveling), and empowering small teams/indies.
  • Others fear an overwhelming influx of low‑effort AI anime, worsening already‑perceived quality decline and making high‑effort shows harder to find.
  • People note that animation has always been cost‑driven: past shifts (xerox vs inking, Flash-era TV, 3D shortcuts) already traded style for efficiency.
  • A Toei Animation report is cited: they plan AI for storyboards, color specification/correction, in‑betweening, and backgrounds—suggesting mainstream studios will adopt AI as a production aid rather than full replacement, at least initially.
  • Some argue audiences already tolerate “sloppy” visuals when writing is strong (e.g., low‑frame shows, simple-looking series), so they may accept mild artifacts for more content; others say jarring AI in‑betweens in a beloved show would be infuriating.

Artist livelihoods, future of work & culture

  • Multiple posts express sympathy for illustrators whose work is being scraped to train models that then undercut their commissions and studio jobs.
  • Comparisons are drawn to translators and musicians: machine assistance and streaming have driven down rates and made full‑time creative careers rarer, even as overall output volume rose.
  • One prominent theme: we risk a world of abundant personalized media but fewer shared cultural touchstones (everyone watching unique AI‑tailored “Frozen‑like” content instead of the same show), weakening art’s social role.
  • Others counter with economic analogies (furniture, custom cars): mass‑produced media and a smaller “hand‑made” sector can coexist, with human‑made work becoming a premium status good, potentially authenticated by cryptographic “100% human” labels.

Legal status of outputs

  • A US Copyright Office bulletin is referenced: generative AI outputs are only copyrightable where a human “determined sufficient expressive elements.”
  • This raises concern that AI‑generated shows might be weakly protected: if courts see them as primarily machine‑authored, anyone could freely copy or remix them, undermining monetization.

Motivations & demand

  • Some ask “who needs this?” and view it as pointless or creepy compared to human animation.
  • Counterarguments:
    • Huge unmet global demand for anime‑style content; East Asian studios can’t meet it at current price points.
    • AI anime could break what’s seen as an “East Asian monopoly” on the style and respond to skewed supply‑demand dynamics (e.g., doujinshi scarcity, scalping).
    • Faster production cycles for sequels and adaptations are seen as a major draw for fans.

User experience & access

  • Several users report the web demo or associated tools:
    • Some say it’s free and works well.
    • Others encounter build failures, errors while consuming credits, or are annoyed by Google login and hidden account requirements for uploads.
    • One alternative site (anisora.ai) is recommended as working smoothly.
  • There’s also curiosity (and expectation) that the model can/will be used for hentai or explicit content, given perceived weaker guardrails in some Chinese AI services; no definitive answer is shared.