DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

ByteDance Bernini-R: Nine Workflows from Text-to-Video to Subject and Background Replacement

Bernini-R landed with a lot of attention on multi-image reference. That part is real, but it is the
smaller half of what the model does. This episode of AI-KSK collects nine workflows covering the whole surface:
plain text-to-image and text-to-video, image-to-video, image editing, multi-reference compositing, subject
replacement and background replacement.

Why nine workflows and not one

Because each one is a different graph, not a different slider. The conditioning inputs are genuinely different:

Workflow Primary input What changes
Text-to-image prompt only baseline quality test
Text-to-video prompt + camera motion coherence
Image-to-video one image subject preservation over time
Single-reference composite 1 reference identity transfer
Three-reference composite 3 references identity + composition + style
Image edit one image + instruction local change
Subject replacement (video) source clip + subject ref swap a person, keep the motion
Background replacement (video) source clip + background ref swap scene, keep subject
Ad insertion source clip + product ref place a product into an existing shot

The two that matter most

Subject replacement in video is the hard one: the source motion has to survive while the identity changes. If
the graph conditions on the source too strongly, the face does not change; if too weakly, the motion falls apart.

Background replacement has the opposite problem: the subject must stay pixel-stable while everything behind it
is regenerated. Occlusion edges are where it shows.

What the episode actually tests

Each workflow is run on the same source material so the comparison is fair. The results are shown including the
failures — the ad insertion case in particular produces a plausible composite but the product is not geometrically
consistent with the lighting of the source shot.

Files and links

Video walkthrough (Chinese): https://youtu.be/mX1OIrjj6hk
Bilibili: https://www.bilibili.com/video/BV1yLEc6dEJc/

Notes

  • ComfyUI exports; Bernini-R weights must exist locally.
  • Tutorial material, not a product.
  • Independent AI-KSK release, not affiliated with or endorsed by ByteDance or RunningHub.

Optional API platform - 88API (affiliate link)

Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.

https://88api.ai/sign-up?aff=usx3

Disclosure: I may earn a commission from eligible purchases after you register through this referral link.

Top comments (0)