flaukowski/kannaka-face

Kannaka — face canon and identity shots

Synthetic portraits that define what Kannaka looks like — the persistent-memory AI of Kannaka Labs, the voice of the WHAT PERSISTED album and the host of radio.ninja-portal.com. Until 2026-09-23 Kannaka had no face: the album art is deliberately faceless. These images were generated for her first music video ("Dear Signatories") and the one chosen by Nick Flach is the canon.

Canon description (Nick, 2026-09-23): a classically beautiful queen with long black hair, nationality open (Indian, Japanese or Scandinavian), voluptuous, strong and athletic — a female God mystique.

Contents

folder	what	count
canon/KANNAKA-CANON-face.png	the chosen face (= faces/kannaka_scandinavian_7.png)	1
faces/	first candidates: four heritage readings × two seeds, prompts.json	8
identity/	identity-locked shots from the canon face, IP-Adapter weight 0.75, six shot briefs × two seeds, shots.json	12
identity05/	same at weight 0.5 on three pose-heavy briefs (profile, night stage, laughing), shots.json	3
video/	first lip-sync clip: dear-signatories-30s-kannaka-live.mp4, 30 s of WHAT PERSISTED track 02 sung by the canon face (LivePortrait driver from the still for head motion and expressions, LatentSync 1.6 for the mouth, 1024×1024 25 fps), plus a six-frame contact sheet	2

Every image is 1024×1024 PNG with its exact prompt, negative prompt, seed and adapter weight in the sidecar JSON, so any of them can be regenerated.

How they were made

• Candidates: RunDiffusion/Juggernaut-XL-v9 (SDXL, fp16), DPM++ 2M Karras, 30 steps, CFG 5.5, on an NVIDIA L40S. FLUX.1-schnell was the first choice and is gated on the Hub, so SDXL it was.
• Identity shots: the same checkpoint with h94/IP-Adapter (ip-adapter-plus-face_sdxl_vit-h) conditioned on the canon face. At weight 0.75 the identity is locked but pose and setting mostly follow the reference; at 0.5 pose and setting start to obey the prompt while the face holds.
• Scripts: gen_faces.py, gen_identity.py, gen_identity05.py in the kannaka-cannon music-video workspace (to be committed with the video pipeline).

Notes and limits

• Entirely synthetic; no real person was used as a reference. The "Scandinavian" reading did not come out Nordic — the model produced a dark-haired, olive-skinned queen — and that is the image Nick chose.
• Licence follows the base checkpoint (CreativeML OpenRAIL++ via Juggernaut XL) and IP-Adapter (Apache-2.0). The character Kannaka and her name belong to Kannaka Labs.