FLUX Kontext edges rivals on facial identity in local editing model shootout

A side-by-side test of four open-weight image editing models—FLUX Klein 4B, Klein 9B, FLUX Kontext, and Qwen—shows FLUX Kontext maintaining facial identity more reliably than its peers, though it struggles to follow spatial composition prompts. The comparison, shared by a practitioner running local inference workflows, highlights a common tradeoff in diffusion-based editing: models that preserve subject identity often sacrifice prompt adherence.

The tester reported that FLUX Kontext kept facial features consistent across edits but couldn't reliably execute a "selfie shot" framing instruction. Qwen produced what the user described as "plastic skin," while FLUX Klein 9B "barely maintains identity most of the times." Klein 4B was tested but not singled out for specific issues, suggesting middling performance on both axes.

Key findings

1. 01FLUX Kontext wins on identity preservation — facial features stay recognizable across edits, the primary goal for character-consistent workflows.
2. 02Prompt adherence lags behind — Kontext fails to parse composition instructions like "selfie shot," limiting its use in scenarios where framing matters as much as the subject.
3. 03Qwen's texture artifacts persist — the "plastic skin" issue suggests over-smoothing or a training bias toward synthetic-looking renders, a known problem in some Qwen vision checkpoints.
4. 04Klein 9B underperforms on identity — despite its larger parameter count, it loses facial consistency more often than Kontext, raising questions about architecture choices or fine-tuning strategy.
5. 05 — practitioners needing both identity consistency and prompt adherence may need to chain models or fall back to inpainting workflows that isolate the subject from the composition.