Medical video generation / training-free acceleration

MedCache: Training-free spatially aware caching for medical video generation.

Ufaq Khan Umair Nawaz Sathira Silva Numan Saeed Muhammad Umer Sheikh Muhammad Bilal Junaid Qadir Lena Maier-Hein Yutong Xie Muhammad Haris Khan

MedCache accelerates rectified-flow medical video generation by estimating cache-reuse risk directly in latent space. The cache protects localized structures such as tool-tissue contact, anatomical interfaces, fluid boundaries, and motion-sensitive details without probing the generator.

arXiv Code Coming soon Method

Motivation

Medical video caching needs spatial risk, not only global drift.

Lightweight caches are fast but collapse latent change into one scalar. Probe-based caches recover richer signals but execute part of the network before deciding whether to skip. MedCache targets the third design point: spatially selective reuse from cheap latent operations.

Spatial asymmetry

Small regions can dominate clinical relevance.

Tool boundaries, tissue deformation, mucosal texture, and ultrasound-like interfaces can be decisive even when the global frame is stable.

Probe-free decision

The cache decision should not spend the speedup.

MedCache reads boundary, high-frequency, and relative-motion cues directly from the latent, avoiding fixed transformer-block probe cost.

Training-free use

No retraining or architecture changes.

The method is inserted at inference time and evaluated on released WAN2.1, Cosmos-Predict-2.5, Cosmos-H-Surgical, and HunyuanVideo settings.

Method

MedCache adds a spatially aware cache-decision layer.

The framework has three components: a latent-derived spatial risk map, risk-weighted accumulated drift, and a lazy high-frequency consistency guard. A prompt-derived domain prior scales the threshold but does not replace the latent decision signal.

MedCache framework diagram from the paper

Main framework. A prompt-derived prior sets the effective threshold, a latent-derived spatial risk map reweights drift estimation, and a lazy high-frequency guard is evaluated only in borderline cases.

Latent collapse

{\overline{x}}_{h, w} = \frac{1}{C T} \sum_{c, t} x_{c, t, h, w}

Spatial risk map

r = {clip}_{[0, 1]} (α_{S} \tilde{S} + α_{F} \tilde{F} + α_{M} \tilde{M})

Risk-weighted accumulated drift

δ_{n}^{MC} = \frac{E_{h, w} [(1 + β r_{h, w}) | {\overline{x}}_{n} - {\overline{x}}_{n - 2} |]}{1 + β E_{h, w} [r_{h, w}]}

Lazy guard

ρ_{n} = \frac{\sum_{(h, w) \in Ω} {(Δ * {\hat{y}}_{n})}^{2}}{\sum_{(h, w) \in Ω} {(Δ * y_{n - 2})}^{2} + ε}, compute if ρ_{n} > 1.50

Mechanism in practice

First identify where reuse is risky. Then decide when it is safe.

Generated medical frame followed by structural, high-frequency, motion, and combined latent risk maps — 01 / Spatial readout **Reuse risk is localized, not uniform.**
MedCache composes structural, high-frequency, and relative-motion evidence directly in latent space. The combined response concentrates around the instrument-tissue interaction without probing the generator.

SStructure

FFrequency

MMotion

Accumulated drift and MedCache skip or compute decisions across denoising steps — 02 / Temporal decision **Safe reuse emerges after warm-up.**
Risk-weighted drift controls the denoising trajectory. On this representative Image2World rollout, MedCache settles into a stable skip/compute rhythm while preserving forced computation at the trajectory boundaries.

**21 / 35** steps skipped 0 guard recomputes

Qualitative Evidence

Local structure is where cache failures become visible.

The paper includes Image2World and Text2World qualitative comparisons. These figures show the same conditioning or prompt across baselines, highlighting local degradation near active surgical regions and temporal coherence across rollout frames.

Main qualitative Image2World comparison. Given the same input image and prompt, WorldCache introduces local artifacts around the active tool-tissue interface, while MedCache preserves the corresponding region more coherently over time.

Supplementary Image2World qualitative comparisons — **Additional Image2World comparisons.** Red boxes mark representative WorldCache artifacts; green boxes mark corresponding MedCache regions.

Supplementary Text2World qualitative comparisons — **Additional Text2World comparisons.** MedCache maintains temporally coherent rollouts close to the uncached baseline while avoiding stronger texture degradation.

Complete Results

Main Text2World and Image2World tables.

Metrics follow the paper: Image Quality (IQ), Subject Consistency (SC), Background Consistency (BC), Motion Smoothness (MS), Total, latency, and speedup relative to the uncached baseline.

Setting	Model	Method	IQ	SC	BC	MS	Total	Latency	SpeedUp
Small-scale models -- Text2World
T2W	Cosmos-Predict-2.5 2B	Baseline	0.751	0.959	0.961	0.991	0.916	271.00	1.00x
T2W	Cosmos-Predict-2.5 2B	FasterCache	0.383	0.917	0.943	0.936	0.795	102.00	2.65x
T2W	Cosmos-Predict-2.5 2B	EasyCache	0.691	0.878	0.908	0.975	0.863	198.00	1.37x
T2W	Cosmos-Predict-2.5 2B	DiCache	0.725	0.948	0.957	0.977	0.902	205.00	1.32x
T2W	Cosmos-Predict-2.5 2B	WorldCache	0.701	0.932	0.943	0.987	0.891	128.00	2.12x
T2W	Cosmos-Predict-2.5 2B	MedCache	0.747	0.918	0.942	0.990	0.899	108.00	2.51x
T2W	Cosmos-H-Surgical 2B	Baseline	0.625	0.906	0.934	0.982	0.862	272.00	1.00x
T2W	Cosmos-H-Surgical 2B	FasterCache	0.348	0.893	0.971	0.962	0.794	130.00	2.09x
T2W	Cosmos-H-Surgical 2B	EasyCache	0.462	0.774	0.899	0.933	0.767	185.00	1.47x
T2W	Cosmos-H-Surgical 2B	DiCache	0.445	0.868	0.930	0.981	0.806	250.00	1.09x
T2W	Cosmos-H-Surgical 2B	WorldCache	0.467	0.877	0.929	0.984	0.814	140.00	1.94x
T2W	Cosmos-H-Surgical 2B	MedCache	0.549	0.884	0.928	0.986	0.837	95.00	2.86x
Large-scale models -- Text2World
T2W	WAN2.1 14B	Baseline	0.713	0.947	0.968	0.990	0.904	663.52	1.00x
T2W	WAN2.1 14B	EasyCache	0.713	0.933	0.963	0.987	0.898	560.62	1.18x
T2W	WAN2.1 14B	FasterCache	0.711	0.929	0.961	0.986	0.897	345.58	1.92x
T2W	WAN2.1 14B	DiCache	0.713	0.933	0.960	0.987	0.898	562.77	1.18x
T2W	WAN2.1 14B	WorldCache	0.713	0.933	0.960	0.987	0.898	561.13	1.18x
T2W	WAN2.1 14B	MedCache	0.704	0.925	0.949	0.987	0.891	183.09	3.62x
T2W	Cosmos-Predict-2.5 14B	Baseline	0.726	0.929	0.941	0.987	0.896	767.97	1.00x
T2W	Cosmos-Predict-2.5 14B	FasterCache	0.478	0.916	0.936	0.886	0.804	442.36	1.74x
T2W	Cosmos-Predict-2.5 14B	EasyCache	0.706	0.928	0.944	0.986	0.891	434.60	1.78x
T2W	Cosmos-Predict-2.5 14B	DiCache	0.723	0.928	0.942	0.985	0.895	700.64	1.10x
T2W	Cosmos-Predict-2.5 14B	WorldCache	0.715	0.916	0.937	0.984	0.888	365.05	2.10x
T2W	Cosmos-Predict-2.5 14B	MedCache	0.725	0.923	0.944	0.987	0.895	331.40	2.32x

Setting	Model	Method	IQ	SC	BC	MS	Total	Latency	SpeedUp
Small-scale models -- Image2World
I2W	Cosmos-Predict-2.5 2B	Baseline	0.640	0.718	0.850	0.959	0.792	52.14	1.00x
I2W	Cosmos-Predict-2.5 2B	FasterCache	0.377	0.510	0.708	0.920	0.629	29.24	1.78x
I2W	Cosmos-Predict-2.5 2B	EasyCache	0.620	0.720	0.851	0.960	0.788	43.59	1.20x
I2W	Cosmos-Predict-2.5 2B	DiCache	0.633	0.714	0.845	0.959	0.788	36.61	1.42x
I2W	Cosmos-Predict-2.5 2B	WorldCache	0.581	0.717	0.848	0.964	0.778	27.50	1.90x
I2W	Cosmos-Predict-2.5 2B	MedCache	0.628	0.723	0.862	0.968	0.795	22.91	2.28x
I2W	Cosmos-H-Surgical 2B	Baseline	0.604	0.881	0.948	0.968	0.850	48.65	1.00x
I2W	Cosmos-H-Surgical 2B	FasterCache	0.414	0.514	0.720	0.901	0.637	23.57	2.06x
I2W	Cosmos-H-Surgical 2B	EasyCache	0.590	0.882	0.947	0.970	0.847	31.45	1.55x
I2W	Cosmos-H-Surgical 2B	DiCache	0.577	0.861	0.934	0.965	0.834	29.25	1.66x
I2W	Cosmos-H-Surgical 2B	WorldCache	0.587	0.864	0.933	0.973	0.839	26.12	1.86x
I2W	Cosmos-H-Surgical 2B	MedCache	0.592	0.884	0.950	0.973	0.850	21.05	2.31x
Large-scale models -- Image2World
I2W	WAN2.1 14B	Baseline	0.719	0.876	0.956	0.987	0.885	524.44	1.00x
I2W	WAN2.1 14B	FasterCache	0.718	0.877	0.956	0.987	0.884	247.14	2.12x
I2W	WAN2.1 14B	EasyCache	0.709	0.880	0.955	0.987	0.883	181.03	2.90x
I2W	WAN2.1 14B	DiCache	0.719	0.876	0.957	0.987	0.885	257.32	2.04x
I2W	WAN2.1 14B	WorldCache	0.717	0.876	0.956	0.986	0.884	259.49	2.02x
I2W	WAN2.1 14B	MedCache	0.717	0.876	0.953	0.985	0.883	138.97	3.77x
I2W	Cosmos-Predict-2.5 14B	Baseline	0.677	0.865	0.937	0.980	0.865	207.26	1.00x
I2W	Cosmos-Predict-2.5 14B	FasterCache	0.474	0.512	0.722	0.886	0.649	130.27	1.59x
I2W	Cosmos-Predict-2.5 14B	EasyCache	0.671	0.870	0.944	0.981	0.867	102.85	2.02x
I2W	Cosmos-Predict-2.5 14B	DiCache	0.671	0.860	0.931	0.980	0.861	162.43	1.28x
I2W	Cosmos-Predict-2.5 14B	WorldCache	0.655	0.825	0.911	0.980	0.843	91.74	2.26x
I2W	Cosmos-Predict-2.5 14B	MedCache	0.678	0.872	0.945	0.984	0.870	76.80	2.70x

Ablations + Analysis

Component, domain-prior, VBench, and sensitivity evidence.

The ablation map separates speed-driving components from fidelity safeguards and includes transfer to a generic VBench/HunyuanVideo setting.

Component removal: Cosmos-H-Surgical Image2World

Config	Wall	Speedup	Skip	IQ	SC	BC	MS	Total
Vanilla	55.7	1.00x	-	0.610	0.866	0.946	0.952	0.844
EasyCache	40.1	1.39x	-	0.594	0.865	0.943	0.959	0.840
+ risk weighting	22.1	2.52x	64.0%	0.593	0.862	0.943	0.961	0.840
+ domain prior	23.5	2.37x	61.1%	0.593	0.865	0.942	0.960	0.840
+ lazy guard	23.5	2.37x	61.1%	0.612	0.867	0.946	0.960	0.846

VBench transfer: HunyuanVideo

Method	IQ	SC	BC	MS	Total	Latency	SpeedUp
Baseline	0.661	0.960	0.971	0.993	0.896	984.04	1.00x
EasyCache	0.662	0.962	0.972	0.993	0.897	485.68	2.03x
WorldCache	0.637	0.959	0.971	0.993	0.890	334.78	2.93x
MedCache	0.664	0.961	0.974	0.995	0.899	286.54	3.44x

Per-domain skip shift

Domain	Sigma	NoPrior skip	WithPrior skip	Delta
Intervention	0.70	48.6%	58.1%	+9.5 pp
Diagnostic imaging	0.85	59.0%	62.9%	+3.9 pp
Clinical exam	0.95	60.0%	61.9%	+1.9 pp
Rehabilitation	1.00	59.0%	59.0%	0.0 pp
Animation/education	1.25	62.9%	61.9%	-1.0 pp
All mean	N/A	57.9%	60.8%	+2.9 pp

Risk-map composition

Config	Speedup	Skip	wS	(wM,wF)	IQ	SC	BC	MS	Total
Vanilla	1.00x	-	-	-	0.610	0.866	0.946	0.952	0.844
M-only	2.28x	61.4%	0.0	(1.25,0.0)	0.593	0.864	0.943	0.960	0.840
S-only	2.23x	60.0%	1.0	(0.0,0.0)	0.592	0.866	0.944	0.959	0.840
S+M	2.29x	60.0%	1.0	(1.25,0.0)	0.593	0.866	0.944	0.959	0.841
S+F+M	2.24x	60.0%	1.0	(1.25,0.75)	0.601	0.867	0.946	0.960	0.844

Probe overhead analysis — Efficiency auditWhere probe cost consumes acceleration
WorldCache pays a fixed network-probe cost before each reuse decision. MedCache keeps gross and measured speedups closely aligned by using latent arithmetic instead.

Sensitivity of MedCache to tau, beta, and kappa — Sensitivity studyA predictable operating region
Sweeps over tau, beta, and kappa expose smooth, interpretable speed-quality behavior around the selected default.

Pareto speed and quality tradeoff — Pareto analysisOne clear control for the speed-quality frontier
Kappa provides the strongest operating-point control, while beta remains comparatively stable near the default. The highlighted MedCache point sits in the balanced region of the measured frontier.

Resources

Paper, code, and citation.

Citation details are available below. The arXiv record and public implementation will be linked here at release.

arXivCode Coming soonCitation

Citation

BibTeX

@misc{khan2026medcache,
  title  = {MedCache: Training-Free Spatially Aware Caching for Accelerated Medical Video Generation},
  author = {Khan, Ufaq and Nawaz, Umair and Silva, Sathira and Saeed, Numan and Sheikh, Muhammad Umer and Bilal, Muhammad and Qadir, Junaid and Maier-Hein, Lena and Xie, Yutong and Khan, Muhammad Haris},
  year   = {2026},
  note   = {Manuscript under review}
}