Generative Vision Atlas

Research line · contested

Adapter-based conditioning

Bolt a small trainable side network onto a frozen pretrained diffusion model to inject a new control signal, instead of retraining or restructuring the backbone itself.

What defines membership

A frozen base model already knows how to generate; new conditioning signals (edges, pose, depth, an image prompt, a face identity) can be added cheaply and safely without touching its weights.

The bet

Retraining or restructuring a large pretrained text-to-image model for every new kind of control signal is wasteful and risky. The base model already knows how to generate convincingly; freeze it, and train only a small side network that learns to steer it.

The arc

ControlNet and T2I-Adapter answer this independently, the same month (February 2023), for spatial control (edges, depth, pose, segmentation). ControlNet duplicates the frozen UNet’s encoder blocks and reconnects them through zero-initialized convolutions so training starts as a no-op; T2I-Adapter uses a much smaller external network aligning internal features to the control signal. IP-Adapter moves the same idea from spatial control to image-prompt conditioning via decoupled cross-attention — a second cross-attention branch, with its own key/value projections, that keeps image-prompt conditioning from fighting the base model’s existing text conditioning. InstantID and PuLID specialize the same pattern for face identity, each solving the same core tension — a strong enough conditioning signal tends to leak into and override everything else the base model does — with a different isolation mechanism (semantic/spatial separation for InstantID; contrastive and accurate-ID losses for PuLID).

The challenge to the whole family arrives from inside the DiT era, not from a better adapter. OminiControl shows that a sufficiently capable DiT’s own frozen weights and attention can absorb a condition image directly — reusing the model’s own VAE encoder and transformer blocks, concatenating condition and generation tokens into one sequence — for about 0.1% additional parameters and no separate trained network, matching or beating specialized adapters on their own tasks. This is the opening move of in-context editing, which this atlas records as the line adapters now compete against for image and reference conditioning specifically.

What holds up, what doesn’t

Adapters remain the practical default wherever no established in-context substitute exists yet — most visibly, spatial control signals (edges, depth, pose) still overwhelmingly ship as ControlNet-style modules even in 2026 toolchains, because DiT-native sequence-concatenation was demonstrated for image and reference conditioning first and has not been shown at the same maturity for arbitrary spatial control signals. For image-prompt and identity conditioning specifically, though, the shift documented above is real: OminiControl, then FLUX.1 Kontext and its successors, increasingly do with plain attention what IP-Adapter and InstantID needed a dedicated second attention branch or side network to do. Composability — stacking several adapters on one generation — is a genuine advantage no in-context method has fully replicated, since concatenating many reference images into one sequence has its own quadratic attention cost and grounding failure modes (see line-in-context-editing’s own weaknesses). This is why this line’s status is recorded as contested rather than superseded: real, live disagreement about which approach wins where, not a settled loss.

How the line developed

Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.

The idea

ControlNet · 2023-02landmark

Trainable copy of the frozen UNet's encoder blocks, connected back through zero-initialized convolutions so training starts as a no-op.

T2I-Adapter · 2023-02core

Independently published the same month with a much smaller external adapter network aligning internal features to an external control signal.

Improvement

IP-Adapter · 2023-08landmark

Moves the idea from spatial control to image-prompt conditioning via a decoupled cross-attention branch, so image and text prompts stop competing for one shared cross-attention layer.

InstantID · 2024-01core

Combines an IP-Adapter-style semantic face branch with a ControlNet-style spatial landmark branch (IdentityNet) for zero-shot identity preservation.

PuLID · 2024-04strong-followup

Adds contrastive and accurate-ID losses on a parallel Lightning-T2I branch to further reduce interference with the base model's own behavior.

Limitation

OminiControl · 2024-11landmark

Shows a DiT's own frozen weights and attention absorb condition tokens directly via sequence concatenation, matching adapter-level control for about 0.1% extra parameters and no separate adapter network — the opening move of the shift line-in-context-editing documents.

What it gets right

  • Frozen base model stays untouched and reusable across community checkpoints
  • Composable — multiple adapters (e.g. several ControlNets) can stack on one generation
  • Each adapter is cheap to train relative to retraining or fine-tuning the base model

Where it is weak

  • Every new control type needs its own trained adapter or branch
  • An adapter can fight the base model's existing text conditioning if not carefully isolated — the problem IP-Adapter's decoupled cross-attention was built specifically to fix
  • DiT-native in-context conditioning (line-in-context-editing) increasingly matches or beats adapters with less architecture for image/reference conditioning, leaving adapters' long-run role contested outside of spatial control signals (edges, depth, pose) that have no established in-context equivalent yet

Competing answers