DEV Community

AI-KSK
AI-KSK

Posted on

Bernini-Diffusers Full Build: MLLM Semantic Planning and a Chinese WebUI for Video Generation and Editing

#comfyui #ai #tutorial

This is not a light Bernini-R variant: it is the Bernini-Diffusers full build, with MLLM semantic planning and a Chinese WebUI preserved end to end.

Evidence note: no workflow JSON was shipped and node parameters were not parsed; the build contents below come from the episode description and the mirror page.

Why the "full build" matters

Bernini's community reputation is built on two things: semantic planning before generation (an MLLM decides what the shot should contain) and an editing/driving UI that non-technical users can operate. Lightweight ports often drop one of them. This mirror keeps both — the MLLM planning stage and the Chinese WebUI.

What you get

  • MLLM semantic planning — describe the shot in natural language; the model plans semantics before the video graph runs.
  • Chinese WebUI — the operation surface is localised, so the workflow is not a node-graph exercise.
  • Generation and editing in one build — the same environment handles both.
  • Diffusers backend — the Python-side pipeline rather than a minimal single-file inference script.

Practical use

  1. Deploy the mirror: https://www.compshare.cn/images/JwzEWFfKjGMY?referral_code=A7fTQUYY9lCDFJwA6MMvoq
  2. Open the WebUI.
  3. Write the shot description; let the MLLM plan.
  4. Generate, then use the editing route for revisions.

Limits

  • This is a heavy build; it expects a real GPU allocation.
  • MLLM planning adds latency before the video graph starts.
  • Bernini is a community model family; licensing follows the upstream release.
  • No JSON/parameters parsed for this episode.

Episode resources

Top comments (0)