Research line · dominant
In-context editing
Edit by putting the reference image into the model's own token sequence and letting attention handle it, rather than through adapters or inversion.
What defines membership
Editing needs no special mechanism: a reference image is just more context for a sufficiently capable generator.
How the line developed
Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.
Before
OminiControl · 2024-11landmark
Predates ACE++ by about two months and demonstrates the same sequence-concatenation mechanism for general DiT control (subject-driven generation, spatial control), not editing specifically; no confirmed citation links it directly to ACE++ or Kontext, so it is recorded here as a precursor rather than reassigning this line's origin.
The idea
ACE++ · 2025-01strong-followup
Unifies creation and editing under a long-context conditioning unit, implemented cheaply as a LoRA on FLUX.1.
Evidence
ICEdit · 2025-04strong-followup
State-of-the-art instruction editing with roughly 0.1% of the data and 1% of the trainable parameters prior editors needed.
Improvement
Step1X-Edit · 2025-04core
Multimodal LLM parses the instruction, DiT executes it; also the source of GEdit-Bench.
At scale
FLUX.1 Kontext · 2025-06landmark
Concatenates reference tokens onto the generation sequence in one rectified-flow model, and becomes the open reference point for the approach.
Seedream 4.0 · 2025-09core
Industrial-scale generation and editing in a single system.
Combination
OmniGen2 · 2025-06core
Editing as one capability of a unified understanding-and-generation model with separate text and image decoding paths.
What it gets right
- Architecturally simple: no inversion, no adapters, no per-task heads
- Handles local edits and global reference-driven generation with one mechanism
- Very data-efficient when built on a strong pretrained generator
Where it is weak
- Preserving untouched regions exactly is still imperfect
- Referring-expression grounding fails when several similar objects are present
- Depends on a strong base generator, so it inherits that model's biases