FAQ

LLM stands for Large Language Model — the same category of model as Claude or GPT. In Bridge, an LLM reads the client note, extracts the actual intent, splits it into action points, and later writes the structured prompt that drives image generation.

Grounding DINO is an object-detection model. Given a frame and a plain-language label like "the mountain in the background," it locates that region and returns a bounding box. It's the step that turns a written description into a specific area of the image.

SAM2 stands for Segment Anything Model 2. It takes the rough bounding box from Grounding DINO and converts it into a pixel-precise mask — the exact boundary the edit will be confined to.

FLUX.1-Fill-dev is a diffusion model — the same family of model behind tools like Midjourney or Stable Diffusion — specifically tuned for inpainting. It generates new pixel content only inside the masked region, leaving the rest of the frame untouched.

ControlNet is a conditioning method that guides a diffusion model's output to follow a structural reference — like depth or edge data — rather than generating freely. It's planned for Bridge as a consistency pass across sequential frames, but isn't built yet; the current pipeline is validated for single-frame visual briefs.

Several, each doing a distinct, narrow job: a language model for understanding the note, vision models (Grounding DINO, SAM2) for locating and masking the edit, and a diffusion model (FLUX.1-Fill-dev) for generating the actual pixels. Bridge's job is orchestrating that chain reliably, not relying on one model to do everything.

No. Bridge produces a visual brief for the supervisor to review — it doesn't approve, publish, or send anything to an artist without a person signing off first.

Bridge is built around non-destructive editing — every output is the original frame with only the specified region changed, not a reimagined version of the shot. Anything outside the masked area is untouched.

It's generated from your studio's own historical production data, broken down by department, so it reflects how your team actually works rather than generic industry averages.

The supervisor sees the extracted action points and the generated brief before anything goes further — if either is off, they can adjust or annotate directly rather than having it propagate to an artist.

It's built to detect the difference between a global grade (no masking needed) and a localized edit (which needs detection and masking) automatically, and route accordingly. Notes that are pure approvals or filler ("this shot is working perfectly") are recognized and don't generate spurious action points.