I Published 7 Preprints in One Day to Claim Every Open Niche in Mamba-2 Image/Video Generation

The SSD-Series: a complete research roadmap for replacing Transformers in diffusion models — from generation to editing to super-resolution.

Last week, I published DiM-2, a technical report proposing Mamba-2 as a diffusion backbone for image and video generation. The response was encouraging — but it also made me realize something.

The entire Mamba-2 visual generation landscape is nearly empty.

I did a systematic literature survey. As of February 2026, only ~8 papers use Mamba-2 for any kind of visual generation. Meanwhile, Mamba-1 has 30+ papers. And some of the most production-critical applications — ControlNet, consistency distillation, portrait animation, instruction editing, super-resolution — have zero Mamba-2 papers. Not one.

So I wrote six more.

Today, the SSD-Series has 7 papers, all publicly available on Zenodo with DOIs. Each one targets a specific gap where Mamba-2 has never been applied. Each proposes novel architectural components that exploit the unique properties of Mamba-2’s Structured State Space Duality.

Here’s the full series.

The Gap Map

Before I explain each paper, here’s what the landscape looks like:

Application Mamba-1 Papers Mamba-2 Papers SSD-Series Image/video generation DiM, DiMSUM, ZigMa, … DiM-2, LinGen DiM-2, SSD-Flow Consistency distillation 0 0 SSD-CM ControlNet / spatial control ControlMambaIR (restoration only) 0 SSD-Control Portrait animation ACTalker (Mamba-1) 0 SSD-Portrait Super-resolution 15+ (MambaIR, MambaIRv2, …) 0 SSD-SR Instruction-based editing 0 0 SSD-Edit Flow matching (image) M4V (video only) LinGen (video only) SSD-Flow

Every “0” in the Mamba-2 column is a paper in the SSD-Series.

Paper 1: DiM-2 — The Foundation

**DOI: 10.5281/zenodo.18689888**

The base architecture. DiM-2 replaces Transformer blocks in diffusion models with Mamba-2 SSD layers, introducing:

• Dual-Axis SSD Scanner (DASS): Separate spatial (S-SSD) and temporal (T-SSD) branches for image and video
• Semi-Separable Conditioning (SSC): Diffusion timestep directly modulates the SSD decay matrix A, not just an additive bias — the SSM equivalent of adaptive layer normalization

DiM-2 serves as the backbone for every other paper in the series.

Paper 2: SSD-CM — Few-Step Generation

**DOI: 10.5281/zenodo.18690580**

The gap: Mamba-2 diffusion models still need 50–250 sampling steps. Consistency models can reduce this to 1–4 steps. But nobody has combined them — not for Mamba-1, not for Mamba-2. LinGen explicitly called this out as future work.

The insight: Consistency distillation on recurrent architectures needs special treatment. If the student’s hidden state dynamics diverge from the teacher’s, the recurrence becomes unstable — even when the output loss converges.

Key components:

• SSD-Aware Consistency Training (SSD-CT): Applies distillation loss in both output space and hidden state space
• Timestep-Adaptive State Initialization (TASI): Carries hidden state across denoising steps instead of zero-initializing — exploiting the fact that consecutive denoising steps process similar inputs

Theoretical speedup: Linear complexity per step × 1–4 steps = 50–200× over DiT + DDIM.

Paper 3: SSD-Control — ControlNet for Mamba-2

**DOI: 10.5281/zenodo.18690631**

The gap: ControlNet is perhaps the most-used component in production AIGC workflows (especially in ComfyUI). It assumes convolutional or attention-based blocks. Pure Mamba-2 backbones have no ControlNet equivalent.

Key components:

• Zero-Projection SSD Adapter (ZPSA): The SSM analogue of ControlNet’s zero-convolution — a trainable copy of SSD layers connected to the frozen backbone via zero-initialized linear projections
• Condition-Injected State Modulation (CISM): Encodes depth/pose/edge maps directly into the SSD B and C matrices — controlling what information enters and exits the state based on spatial conditions

Supports depth, pose, canny edge, segmentation, and temporal video conditions. Maintains O(LN) complexity throughout.

Paper 4: SSD-Portrait — Talking Head Animation

**DOI: 10.5281/zenodo.18690678**

The gap: Portrait animation (LivePortrait, Hallo, EchoMimic) is bottlenecked by Transformer temporal attention — limited to ~10 seconds before quality degrades. ACTalker (ICCV 2025) proved Mamba-1 works for talking heads. DiM-Gestor proved Mamba-2 works for body gesture animation. But nobody has done Mamba-2 for portrait animation.

Key components:

• Facial Landmark SSD (FL-SSD): Processes driving signals (audio, landmarks) as a state-space trajectory — facial motion is inherently sequential and causal, making SSM a natural fit
• Appearance-Motion Decoupled SSD (AMD-SSD): Separate S-SSD (identity) and T-SSD (motion) branches to prevent identity leakage
• Audio-Conditioned Decay Gating (ACDG): Modulates the temporal decay matrix A with audio features — slow decay for sustained vowels (slow mouth change), fast decay for plosive consonants (/p/, /b/, /t/) where the mouth transitions rapidly

Enables minute-length portrait animation with constant per-frame cost.

Paper 5: SSD-SR — Super-Resolution

**DOI: 10.5281/zenodo.18697475**

The gap: The most striking gap in the landscape. Over 15 Mamba-1 SR papers exist (MambaIR, MambaIRv2, DVMSR, Hi-Mamba, MambaCSR, VSRM, MambaSR, MambaOSR, MambaLiteSR…). Zero use Mamba-2. MambaIRv2 even references SSD theory but still builds on Mamba-1.

Key components:

• SSD Residual Groups (SSD-RG): Drop-in replacement of Mamba-1 blocks with SSD-native layers, state dimension N = 128–256 vs. Mamba-1’s N = 16
• Frequency-Aware State Routing (FASR): Low-frequency components → SSD channels with slow decay (long-range structure). High-frequency components → SSD channels with fast decay (sharp local detail). This frequency-adaptive processing is only possible because Mamba-2’s decay matrix A is large enough to encode meaningful per-channel behavior
• Diffusion-Guided SSD-SR (DG-SSD-SR): The first diffusion-based Mamba SR model, combining perceptual quality of diffusion SR with linear complexity

Paper 6: SSD-Edit — Instruction-Based Editing

**DOI: 10.5281/zenodo.18697566**

The gap: Instruction editing (InstructPix2Pix, MagicBrush, MGIE) — zero Mamba papers of any kind. FluencyVE used Mamba-1 for video style transfer, but not instruction-following editing.

The unique insight: SSM hidden states enable something Transformers fundamentally can’t do — automatic edit localization by comparing hidden states.

Key components:

• Source-Instruction Dual-Stream SSD (SI-SSD): Parallel SSD streams for source preservation and edit generation, with state-level cross-injection (no cross-attention needed)
• Selective Edit Masking via State Divergence (SEMD): Where the instruction-conditioned hidden state diverges from the source state → that’s the edit region. No explicit mask needed. This is a qualitative advantage of recurrent architectures, not just a speed improvement
• Temporal Edit Propagation (TEP): For video editing, T-SSD’s recurrent state naturally propagates edits forward through time — persistent edits get slow decay, transient edits get fast decay

Paper 7: SSD-Flow — Flow Matching for Image Generation

**DOI: 10.5281/zenodo.18697643**

The gap: Flow matching (Stable Diffusion 3, Flux, SiT) has become the dominant training paradigm. LinGen combined Mamba-2 + flow matching for video. But no pure Mamba-2 + flow matching image generation model exists — the most fundamental and widely benchmarked setting.

Positioning: SSD-Flow is to DiM-2 what SiT is to DiT — the flow matching upgrade.

Key components:

• Velocity-Conditioned SSD (V-SSD): Adapts SSC conditioning from noise prediction to velocity field prediction, with learned output scaling matched to velocity magnitude across the ODE trajectory
• Adaptive Timestep Routing (ATR): Learned per-layer gates that allocate more compute to high-curvature regions of the flow trajectory (where structure emerges from noise) and less to low-curvature regions
• Logit-Normal State Scheduling (LNSS): Adapts the SSD decay matrix across the flow timestep — short memory at t≈0 (pure noise), long memory at t≈0.5 (structure forming), medium at t≈1 (local refinement)

The Bigger Picture

These 7 papers form an ecosystem:

SSD-Flow (backbone)
  └── SSD-CM (few-step inference)
  └── SSD-Control (spatial conditioning)
  └── SSD-Portrait (talking head animation)
  └── SSD-SR (super-resolution)
  └── SSD-Edit (instruction editing)
  └── DiM-2 (DDPM-based alternative backbone)

Every component is designed to compose with the others. SSD-Portrait + SSD-CM = real-time talking heads. SSD-SR + SSD-CM = few-step super-resolution. SSD-Control + SSD-Edit = controlled instruction editing.

Why Theory-First Preprints?

I should be transparent: these are architectural proposals with theoretical motivation, not experimental results papers. Every paper includes a detailed experimental protocol — the specific benchmarks, baselines, metrics, and ablations needed to validate the ideas — but the experiments haven’t been run yet.

I published them for three reasons:

1. Timestamped priority. Mamba-2 visual generation is accelerating. Having DOI-stamped proposals on Zenodo establishes when these ideas were formalized.
2. Inviting collaboration. I have an RTX 5090 and production AIGC experience. What I don’t have is a 64-GPU cluster. If these ideas resonate with a lab that does, I’d love to collaborate on the experiments.
3. The ideas are the hard part. Identifying that SSD hidden state divergence can localize edits, or that audio features should modulate the decay matrix for lip sync, or that consistency distillation needs state-space regularization — these are architectural insights that don’t require experiments to be useful to the community.

All Papers

Paper Title DOI DiM-2 Exploiting SSD in Mamba-2 for Unified Image and Video Diffusion 10.5281/zenodo.18689888 SSD-CM Few-Step Generation via Consistency Distillation on Mamba-2 10.5281/zenodo.18690580 SSD-Control Spatially Controllable Generation with Mamba-2 10.5281/zenodo.18690631 SSD-Portrait Linear-Time Portrait Animation via Mamba-2 10.5281/zenodo.18690678 SSD-SR Upgrading SR from Mamba-1 to Mamba-2 with SSD 10.5281/zenodo.18697475 SSD-Edit Instruction-Based Editing with Mamba-2 Diffusion 10.5281/zenodo.18697566 SSD-Flow Flow Matching Meets Mamba-2 for Image Generation 10.5281/zenodo.18697643

All papers are CC BY 4.0. If you’re working on Mamba-2 for visual generation, I’d love to hear from you.

Hiroki Abe — Independent Researcher, Tokyo Contact: genelab999@gmail.com