Character consistency and iteration

last edited

Character consistency, iteration, and production pipelines now depend on reference-image limits that differ by tool. Several tools named in older guides have been retired or superseded, so this page compares current documented figures and keeps a ComfyUI pipeline section.

Reference images by tool

FLUX.2 accepts up to 8 reference images through the API. The playground allows up to 10. FLUX 3 Image arrived in October 2026. It accepts up to 10 references and adds bounding-box placement and 4K output.

Nano Banana 2.1 reached general availability in October 2026. The date was the 6th. It takes up to 14 references. It keeps up to 4 characters consistent.

The Midjourney default model changed on June 11, 2026. The new default was V8.1. Omni-reference (`--oref`) from V7 stays available while the V8 version is in training. Public testing of the Edit Model began on August 27, 2026. It takes up to 4 image references.

Iteration and editing

On February 13, 2026, OpenAI retired its older chat model in its chat product. The retired model was GPT-4o. The current OpenAI image release shipped on September 8, 2026. Its API offers a faster model and a more precise model. OpenAI highlights consistent details across multiple edits.

The Sora app closed on April 26, 2026. The API closed on September 24, 2026. The character feature is gone.

Production pipelines

ComfyUI is the node-graph tool used for local pipelines. It added official FLUX 3 Klein support in release 0.43.0. The date was about October 27, 2026. Klein is a 5.6B open-weights model. It supports low-rank adapter training.

Related coverage: AI Art Techniques & Workflows — Organized Reference, Tool Adoption Patterns recent, Platform Migration Patterns (2024-2026), and Kohya Trainer Edge Cases & Advanced Settings.

Sources