FLUX Multi-Reference Character Workflow Discussion

Roshpatoich · reddit · 2026-07-15

The author is trying to build a "3-in-1 character generation workflow" for FLUX in ComfyUI. The goal isn't a face swap, but enabling native identity and pose control during the diffusion generation phase.

Three Target Modes

Text-only: Standard text-to-image.
Text + Face Reference: Preserves identity, facial features, and realism while letting text dictate clothing, pose, and background.
Text + Face + Body Reference:
The face image dictates identity, skin texture, and style;
The body image provides only body type, silhouette, clothing shape, and pose;
The author intends to use the body reference as an "anatomical blueprint" rather than a style reference.

Current Status & Issues

The author originally planned to use ComfyUIPuLIDFluxll for identity preservation. However, after upgrading to the latest ComfyUI Portable, they encountered API compatibility errors like transformeroptions, attnmask, and timestepzeroindex. They suspect ComfyUI's internal FLUX interface has changed, and the nodes haven't caught up.

Points to Confirm

Is PuLID Flux LL still the best approach?
Is anyone maintaining it and adapting it for the latest ComfyUI?
Are there better alternative architectures now, such as Flux Redux, IPAdapter FaceID, PuLID, or multi-node combos?

Original post →

FLUX Multi-Reference Character Workflow Discussion

More from coding & agent