Podcast Atlas
Source note Episode guide Original audio Topics: Technology, Culture

No.227 对话叫兽易小星&粉墨：人人都能用 AI 拍电影了？

Summary

This 三五环 episode has 刘飞 interviewing 叫兽易小星 and 李晨 of 粉墨 about making the AI short 《女娲之死》 for the B站 AI 创作大赛. The conversation turns AI Video Production Workflow and AI Director-Core Workflow into a concrete film-production account: script development, style choice, storyboards, image generation, model-specific video generation, editing, deleting shots, regenerating, sound, color, and human review all remain part of the work.

The durable contribution is a practitioner boundary for AI filmmaking. The guests argue that Video Models lower production and trial costs, but they do not remove story, taste, performance judgment, audience understanding, or responsibility. AI may absorb more spectacle, low-quality acting, short-drama, and previsualization work, while human creators remain decisive where comedy timing, subtle performance, style control, and original point of view matter.

Key Claims

  • B站 AI 创作大赛 is presented as a large creator event with a 3 million RMB prize pool, 77 awards, and enough participation to reveal current AI-video workflow strengths and limits.
  • 《女娲之死》 used a claymation-like stop-motion style because the material fit the mythic clay/human theme and because the team wanted the result to feel less like a generic AI video.
  • 叫兽易小星 treats AI as another creator tool in the long movement from film cameras to digital cameras, phones, and short-video platforms: creative rights keep moving downward, but strong creators remain scarce.
  • 李晨 says AI video first helped him see whether expensive script ideas could become images at all, especially scenes involving snowfields, grasslands, and bodies emerging from water.
  • 粉墨 is using AI short films as a small-team anthology workflow, closer to a series of speculative shorts than to one-off prompt demos.
  • The guests argue that current AI video has visual spectacle before it has equally strong story expression; once visual access is democratized, audience attention shifts back to creativity, emotion, and theme.
  • AI Director-Core Workflow appears in the collaboration: 易小星 supplied overall idea and script, while 李晨 broke the text into executable visual plans and generation tasks.
  • The episode rejects the idea that AI production workers are merely “抽卡师”; good AI artists still need composition, lighting, scene design, timing, and the ability to tell models what should appear when.
  • 《女娲之死》 reportedly took over half a month of script development and about one month of generation, while a comparable ten-minute stop-motion animation might take around two years traditionally.
  • AI can move a project from 0 to 80% quickly, but the guests say the jump from 80% to 85% or 90% can require disproportionately more regeneration, editing, and professional finishing.
  • AI is useful in preproduction as a research and brainstorming assistant; 易小星 used it to check whether the “humans forget Nüwa” premise had precedents such as Lu Xun’s “补天”.
  • Future live-action film use depends on editability and integration with real footage; 易小星 speculates that if those improve, AI could touch a large share of movie shots within a few years.
  • Film workers are described as watchful rather than simply resistant: public AI use can trigger criticism around laziness, craft, plagiarism, copyright, and labor replacement.
  • The guests expect earlier pressure on low-quality, repetitive, or low-threshold content, including some short-drama and weak acting work, while top human actors retain a higher performance ceiling.
  • Prompt work is becoming easier as models understand natural language better, so the long-term barrier shifts from tool operation to taste, expression, and knowing what one wants.
  • AI-generated comedy remains weak in this account because successful comedy depends on face, motion, pause, breath, rhythm, scene chemistry, and fine-grained editability rather than only text jokes.
  • The concrete workflow used tools including Midjourney, Banana, 可灵, Seedance, ChatGPT, 豆包, Veo, Higgsfield, and Claude Opus, selected according to image, motion, action, physical feedback, or brainstorming need.
  • The guests warn against blindly buying courses or tool subscriptions: direct practice, English ability, source verification, and cross-checking AI research matter more than anxiety-driven consumption.
  • AI can hallucinate facts and lose generation control, so people still need verification, prompt constraints, review, and final responsibility for important work.

Key Quotes

“尽量做得不像 AI 作品” - 易小星 on the style aim behind 《女娲之死》.

“从 0 到 80% 很快” - 李晨 on why final polish can still be expensive.

“不要报班” - both guests’ practical advice to new AI creators.

Connections

Contradictions

  • No settled contradiction found. The episode reinforces existing wiki distinctions between AI as production assistance, AI as lower execution cost, and human judgment as the remaining creative bottleneck.
  • The source title names “粉墨” while the body identifies the speaking guest as director 李晨 from the team; this ingest creates 粉墨 and 李晨 separately to avoid merging a person with a team.
  • Tool performance, monthly spending, production duration, future two-year adoption timing, B站 contest figures, and claims about which roles will be hit first are source-scoped practitioner observations rather than independently verified industry data.