NeurIPS 2026

VAR-Q: Tuning-free KV Cache Quantization for Visual Autoregressive Image and Video Generation

Boxun Xu1,†Jiaji Lu1,†Yuxuan Yin1Zihu Wang1Ziyue Liu1Yu Wang1Zirui Liu2Peng Li1

1 University of California, Santa Barbara2 University of Minnesota

† Equal contribution · Correspondence: boxunxu@ucsb.edu

Paper Code BibTeX

VAR-Q quantizes the key/value cache of visual autoregressive models to 2–8 bits with no calibration data and no retraining. It assigns one scale factor per feature channel and per generation block (a scale or a frame), following the measured magnitude structure of the cache. Across five model families and nine checkpoints, 4-bit KV caches reduce cache memory by 72–75% at generation quality close to BF16; Infinity-8B keeps its composition at INT2.

Next-scale models
VAR · Infinity · InfinityStar
Next-frame models
Self-Forcing · LongLive
KV-cache precision
2–8 bit, packed into INT32 containers
Calibration data · retraining
None · none

Paired generations: BF16 vs. VAR-Q

Same prompt and seed in every pair; only the KV cache is quantized. Each pair plays in sync when scrolled into view.

Next-scale

InfinityStar 720p · next-scale video · 4-bit KV cache

10 prompts · same seed · KV cache 62.7 → 16.0 GB (−74%)
Next-frame

Self-Forcing · next-frame video, 5 s · 4-bit KV cache

10 prompts · same seed · KV cache 67.5 → 16.9 GB (−75%)
Next-frame

LongLive · long-horizon video generation, 30 s · 4-bit KV cache

6 prompts · same seed · KV cache 64.3 → 16.1 GB (−75%)

LongLive extends the next-frame setting to 30-second sequences; compare motion and scene consistency over the full clip.

Same model, prompt and seed within each pair; only the KV cache is quantized. KV memory is the paper's measurement at the evaluation batch size (3 clips for InfinityStar 720p, 12 for Self-Forcing, 20 for LongLive).

Infinity-8B INT2 image comparisons ↗

Next-block prediction: one view of next-scale and next-frame generation

Visual autoregression is next-block prediction. A block is the set of visual tokens emitted at one decoding step: a resolution scale in next-scale models (VAR, Infinity, InfinityStar), a frame in next-frame models (Self-Forcing, LongLive). Every block attends to the KV cache of all blocks before it.

Seen this way, both paradigms leave behind the same object: a KV cache indexed by feature channel and generation block. Its magnitudes share a channel × block structure: a small set of channels carries large values, and their magnitudes shift from block to block.

Why it matters for quantization. VAR-Q assigns one scale factor per channel within each block, reusing the block boundaries the generation loop already provides. No outlier detection, no calibration data, no retraining.

A common representationNext-scale / next-frame → channel × block → packed KV
Next-scale and next-frame decoding feed a common channel-by-block KV cache. One scale factor covers each channel within a block; low-bit values are packed into INT32 containers.
Conceptual schematic of the grouping rule in §3.4. Block sizes and colors are illustrative; measured cache magnitudes appear below.

Why do massive values emerge in KV caches of visual autoregressive models?

While §3.2 establishes where massive outliers concentrate, this section investigates why they emerge. We attribute these structural anomalies along the channel and block axes to specific operations within the K and V projection chains (§3.3). Appendix A separates the roles of K and V across modalities.

01 0.53–0.62

K Persistence@10 at p = 1%, a hard architectural core shared across inputs; V sits at 0.07–0.14.

Observation · K outliers are architectural and V outliers are largely input-driven 02 55–63×

random recall of final K outliers from the WK → γK → RoPE chain; WV alone yields 5–6×.

Insight · A compounding architectural effect on K-cache channel outliers 03 +0.41

VBench Total Quality from quantizing the post-RoPE cache instead of the pre-RoPE cache, at the same 27.5 GB (Table 7).

Insight · Block-axis dynamics generated by RoPE periodicity 04 61×

ΔK/ΔV damage ratio on LongLive-30s at p = 0.5%; Infinity-2B sits at 0.79–0.87×.

Observation · Video models are more sensitive to K outliers, image models to V outliers

01 · Observation · Cross-prompt persistence

K outliers are architectural and V outliers are largely input-driven

Whether an outlier is architectural or input-driven can be tested empirically through cross-prompt persistence: an architectural outlier should persist across prompts, while an input-driven outlier should drift. We sample N prompts across models and define outliers as the top-p channels by magnitude. Persistence@N is the fraction of a single prompt’s outlier channels that remain outliers under all N prompts (random ≈ 0).

Persistence at 10 for K versus V: 0.54 versus 0.14 on Self-Forcing, 0.53 versus 0.12 on LongLive, 0.62 versus 0.07 on InfinityStar.
Figure 5. Outlier persistence across different prompts.

At N = 10, p = 1%, K Persistence@10 reaches 0.53–0.62, a hard architectural core shared across inputs; V sits at 0.07–0.14, a 3.8×–8.5× gap (Fig. 5). V’s near-vanishing persistence indicates input activations dominate V outlier identity.

Persistence at N for N from 1 to 10. K curves (solid) decay gently to about 0.55 to 0.62; V curves (dashed) drop to about 0.3 at N equals 2 and continue down to 0.07 to 0.14.
Figure 13. Outlier persistence across different amounts of prompts (Appendix B).

The shapes of the K and V curves differ qualitatively, not just in magnitude. V (dashed) collapses on the first added prompt: even a single second prompt removes about two-thirds of the outlier channels (Persistence@2 = 0.31 to 0.35 across the three backbones), and the curve continues to decay toward 0.07 to 0.14. K (solid) instead bends gently, retaining 0.77 to 0.82 at N = 2 and asymptoting at 0.53 to 0.62. The plateau shape on K is the signature of an architectural core: most outlier channels survive every additional prompt, consistent with their being selected by the static WK, γK, and RoPE-pair chain. The crash on V is the signature of input-driven outliers: each new prompt is a fresh sample of token amplitudes, and the V channel identity tracks those samples.

Observation 01 — K outliers are architectural and V outliers are largely input-driven. K Persistence@10 reaches 0.53–0.62, a hard architectural core shared across inputs; V sits at 0.07–0.14. Value caches exhibit the same two-axis structure with weaker per-axis attribution, sufficient to justify the same grouping rule without a separate mechanism.

02 · Insight · The channel axis

A compounding architectural effect on K-cache channel outliers

We attribute these structural anomalies along the channel and block axes to specific operations within the K and V projection chains. Formally, we decompose the generation of Kt and Vt as follows:

Kt = RoPEt( RMSNorm(xtWK⊤) ⊙ γK )Vt = xtWV⊤Equation 2 · RMSNorm-K and RoPE are applied where the model defines them.

The architectural core of K identified above arises from three multiplicative sources along the projection chain (Eq. 2): (i) heavy-tailed row norms of WK seed an initial outlier baseline; (ii) where RMSNorm-K is present, the channel gain γK amplifies this baseline; and (iii) RoPE rotates each (2i, 2i+1) pair, redistributing energy between the two channels while preserving the pair’s ℓ2 norm.

WK

Seed

Heavy-tailed row norms of WK seed an initial outlier baseline.

γK

Amplify

Where RMSNorm-K is present, the channel gain γK amplifies this baseline.

RoPE

Redistribute

RoPE rotates each (2i, 2i+1) pair, redistributing energy between the two channels while preserving the pair’s ℓ2 norm.

We quantify this compounding by progressively adding components and measuring the recall of final K outliers (Fig. 6). At p = 1%, WK alone recovers 35–42% (35–42× random); adding γK lifts recall further; the full chain reaches 55–63× random. In contrast, WV alone yields 5–6× random recall on every model, an order of magnitude weaker than the K chain.

Recall of final K outliers versus top-K percent on Self-Forcing, LongLive and InfinityStar as W K, gamma K and RoPE scattering are added; the W V control stays near the random baseline. Full-chain recall at the top 1 percent is 57, 55 and 63 times random.
Figure 6. Recall of K outliers as projection-chain components are progressively added; WV shown as control.
Insight 02 — A compounding architectural effect on K-cache channel outliers. Heavy-tailed WK row norms seed an outlier baseline, post-projection RMSNorm gain γK amplifies it where present, and RoPE pair rotation redistributes energy at the pair level. The full chain reaches 55–63× random recall of the final K outliers; WV alone yields 5–6×.

03 · Insight · The block axis

Block-axis dynamics generated by RoPE periodicity

A class of K block-axis outliers is generated by RoPE alone, independent of input dynamics. RoPE rotates each channel pair (2i, 2i+1) at frequency ωi by an angle (bωi) along the pair’s RoPE axis, where b is the block index. Within block b, every token shares the same rotation, and per-token magnitudes are

|K2i(b, t)| = Mt |cos(bωi + φt)||K2i+1(b, t)| = Mt |sin(bωi + φt)|Equation 3 · Mt is the pre-RoPE pair amplitude at token t and φt its initial angle.

Two consequences follow and are visible in Fig. 7. (i) Pair envelope is approximately block-invariant. RoPE preserves the per-token pair ℓ2 exactly, and in the massive-value regime this carries over to the pair of channel maxes (Appendix C). The two channels move in anti-phase within each block. (ii) Fast pairs dominate the block axis. The total angle a pair traverses across the block sequence determines its swing: slow pairs barely rotate and stay flat, fast pairs complete multiple cycles and produce block bands.

Self-Forcing: RoPE-attributable outliers concentrate in the lowest-index fast-rotating channel pairs; within a pair the two channel maxima swap magnitude across frames while the pair envelope stays approximately flat.
Figure 7. (a) On Self-Forcing, block-axis outliers concentrate on the lowest-index (fastest-rotating) pairs. (b) Within such a pair, |c0| and |c1| swap energy across blocks while pair ℓ2 stays approximately flat.

Appendix CWhy the pair envelope stays flat: approximate ℓ2 invariance of channel maxima under RoPE

Within one block of T tokens, let the pair (ct, dt) at token t have pre-RoPE amplitude Mt = √(ct2 + dt2), with order statistics M(1) ≥ M(2) ≥ ⋯ ≥ M(T). After RoPE rotation by the shared block angle θ, write at = Mt cos(θ + φt) and bt = Mt sin(θ + φt), where φt is the per-token pre-RoPE angle. The pair of channel maxima satisfies

M(1)2 ≤ ‖a‖∞2 + ‖b‖∞2 ≤ M(1)2 + M(2)2.(4)

Lower bound. Let t⋆ = arg maxt Mt. Per-token RoPE invariance gives at⋆2 + bt⋆2 = M(1)2. Since ‖a‖∞ ≥ |at⋆| and ‖b‖∞ ≥ |bt⋆|, the sum is at least M(1)2.

Upper bound. Let ta = arg maxt |at| and tb = arg maxt |bt|. Per token, |at|, |bt| ≤ Mt, so ‖a‖∞2 ≤ Mta2 and ‖b‖∞2 ≤ Mtb2. If ta = tb, the sum equals Mta2 ≤ M(1)2. If ta ≠ tb, at least one of them is not t⋆, so min(Mta, Mtb) ≤ M(2) and the sum is at most M(1)2 + M(2)2. ∎

Massive-value regime. When pre-RoPE amplitudes are heavy-tailed within the block, M(2)2 ≪ M(1)2 and the two bounds collapse: the channel-max pair ℓ2 is approximately invariant under RoPE rotation. This is the mathematical basis for the flat envelope in Figure 7(b), while the individual maxima ‖a‖∞ and ‖b‖∞ trade magnitude as θ advances from block to block.

Pre-RoPE vs. post-RoPE quantization (Appendix F). Post-RoPE quantization consistently outperforms pre-RoPE on the metrics most related to temporal behavior and overall generation quality: Subject Consistency, Background Consistency, Dynamic Degree, Aesthetic Quality, and Total Quality (83.34 vs. 82.93 on Self-Forcing at INT6, same 27.50 GB KV cache, Table 7). This supports quantizing the final post-RoPE cache, which is directly consumed by attention, rather than the pre-RoPE cache. We therefore use post-RoPE quantization as the default.

Insight 03 — Block-axis dynamics generated by RoPE periodicity. Position-varying RoPE projects channel magnitudes through a sinusoidal modulation, especially at high-frequency channels in the block index, producing the cross-block shifts directly from the rotation schedule rather than from any change in input. The block boundaries come for free from the generation loop itself, without detection, calibration, or retraining.

04 · Observation · Asymmetric roles of K and V across modalities

Video models are more sensitive to K outliers, while image models are more sensitive to V outliers

The joint result of §3.1 masks a per-tensor split: video models are more sensitive to K outliers, while image models are more sensitive to V outliers. Mechanistically, K determines which past tokens are attended to and V determines what content is retrieved. The split therefore suggests that video generation places a heavier demand on precise temporal addressing into the cache, whereas image generation tolerates coarser addressing but depends more on the fidelity of the content retrieved at each cached position.

Image AR is V-dominant: on Infinity-2B at p = 5%, V-only ablation drives ΔLPIPS to 0.20 against 0.08 for K-only, and the LAION aesthetic score drops by 1.3 under V-only against near-zero under K-only. Video AR reverses the direction: on LongLive-30s at p = 5%, K-only ablation drives ΔLPIPSmid to 0.40 against 0.14 for V-only, and the same direction holds on InfinityStar-720p. The image-side V-dominance is consistent across both Infinity backbones and the video-side K-dominance across both video backbones, indicating that the asymmetry tracks the modality rather than per-model idiosyncrasy.

Three-by-three grid of ablation curves. Top row, image models: the blue V-only curves show the larger damage. Middle row, video models: the red K-only curves show the larger damage. Bottom row, LLaMA-3-8B: red K-only dominates strongly.
Figure 10. K-only (red) versus V-only (blue) outlier ablation. Top row: image AR backbones on perceptual and aesthetic metrics. Middle row: video AR backbones on perceptual and VBench-MUSIQ metrics. Bottom row: LLaMA-3-8B-Instruct on language benchmarks (PPL on log scale). All metrics are oriented so that higher values indicate worse generation. The dominant color flips between image and video AR.
Left, Infinity-2B images at a 5 percent mask: V-only masking collapses outputs into texture while K-only masking looks like the baseline. Right, LongLive-1.3B frames over 30 seconds at a 0.5 percent mask: K-only masking collapses frames into solid colour while V-only masking keeps the scene layout.
Figure 11. Generation under K-only, V-only, and joint K+V outlier masking. (a) Infinity-2B at p = 5%: V-only masking collapses outputs into texture or symmetric patterns, while K-only masking leaves outputs visually close to the baseline. (b) LongLive-1.3B at p = 0.5%: K-only masking collapses frames into solid color and chroma stripes, while V-only masking preserves scene layout with degraded fine detail. Inlier masking matches the baseline in both cases.

The same protocol on LLaMA-3-8B-Instruct behaves like video AR in direction but more sharply in magnitude: at p = 0.5%, K-only ablation increases WikiText-2 perplexity from 8.4 to 82.4 (∼10×), while V-only ablation leaves it at 11.5 (1.4×); MMLU and GSM8K accuracy degrade in the same direction. Figure 12 summarizes the damage ratio ΔK/ΔV at p = 0.5%: image AR sits at or below 1×, video AR ranges from 2.5× to 61×, and LLaMA-3-8B reaches 320× on perplexity.

Log-scale bar chart of the K-over-V damage ratio at a 0.5 percent prune: Infinity image models at 0.79 to 1 times, video models at 2.5 to 61 times, LLaMA-3-8B at 6.9 to 320 times.
Figure 12. ΔK/ΔV damage ratio at p = 0.5% (log scale). Bars below the dashed K = V line are V-dominant; bars above are K-dominant. Image AR sits at or below 1×, video AR ranges from 2.5× to 61×, and LLaMA-3-8B reaches 320× on perplexity.
Observation 04 — Video models are more sensitive to K outliers, while image models are more sensitive to V outliers. The LLM observation aligns with prior reports of K-side outlier dominance in LLM KV caches, placing video AR closer to the LLM regime while leaving image AR as the singular setting in which V outliers carry the bulk of the load.

Channel × block KV-cache quantization

A channel × block view of the KV cache
explains where quantization groups belong.

01

Quantization sensitivity is concentrated in outlier channels

Masking just the top 0.5% of channels in LongLive collapses generation. Masking ten times as many inlier channels leaves quality largely intact.

02

Magnitude structure spans channels and generation blocks

Large magnitudes persist along feature channels and shift across generation blocks. Grouping along only one axis misses part of the structure.

03

Scale- and frame-aligned quantization

A scale or frame defines a block. VAR-Q reuses these boundaries, with no outlier detection, calibration data, or retraining.

Measured K and V cache magnitudes across five visual autoregressive models show persistent channel stripes and shifts across blocks.
The measured structure. Median absolute K and V values across five visual AR models. Color scales are per-panel. Figure 4 of the paper.

Explore the grouping

VAR-Q separates feature channels and generation blocks, so each group follows both axes of the cache.

Illustrative grouping diagram. Colors are schematic, not measured activations.

Feature channels →
Generation blocks →
Block 1Block 2Block 3
Low magnitudeHigh magnitude

Table 1 · Aligning both axes reduces quantization error.

Attention-output error falls from 32.32% to 10.13% when moving from per-tensor to channel × block grouping. The joint rule has the lowest error in every reported column.

Table 1 · Infinity-2B · INT4 · relative errors (%)
Grouping K error ↓ V error ↓ Attention error ↓ Output error ↓
Per-tensor 24.01 43.54 29.20 32.32
Per-block 21.59 32.93 26.36 22.78
Per-channel 12.11 18.07 16.17 13.13
Channel × block 10.04 15.28 13.38 10.13

Relative Frobenius error: εX = ‖X − X′‖F / ‖X‖F. Joint grouping gives the lowest error in all four columns. Source ↗

Ablation: channel × block grouping

Removing the channel axis sharply degrades generation. Keeping channels but removing block boundaries also loses quality—on both a next-scale and a next-frame model.

Grouping ablation · INT4 KV cache
Grouping VAR-d30
FID ↓		 Self-Forcing
VBench Total ↑
Channel × block 1.97 80.99
Without block axis 2.83 79.38
Without channel axis 12.07 67.22

Metrics are compared within each column. Full ablation →

Tables 5–6 · Grouping and boundary-alignment ablations

Table 5 · INT4 · VAR-d30 (VAR) and Self-Forcing (SF)
Grouping VAR FID ↓ IS ↑ PSNR ↑ SSIM ↑ LPIPS ↓ VAR KV (GB) ↓ SF Subject ↑ Dynamic ↑ Aesthetic ↑ Total ↑ SF KV (GB) ↓
Without block axis 2.83 292.13 12.07 0.257 0.538 9.45 88.49 69.00 54.51 79.38 16.90
Without channel axis 12.07 159.52 10.42 0.198 0.645 9.34 79.17 27.00 33.86 67.22 16.87
Per-token & per-head 2.21 292.12 11.97 0.261 0.545 10.51 90.18 67.00 54.96 79.46 17.93
Per-token 7.13 203.50 10.51 0.208 0.635 9.38 78.55 25.50 39.51 67.72 16.96
VAR-Q (full) 1.97 323.67 13.54 0.312 0.437 10.44 91.28 75.50 57.40 80.99 16.95

Same number and sizes of groups, shifted by half a block: misalignment lowers Self-Forcing quality by 0.45 points in this controlled experiment.

Table 6 · Boundary alignment ablation
Boundary placement Subject ↑ Dynamic ↑ Aesthetic ↑ Quality ↑
Misaligned 90.80 73.50 56.75 80.46
Aligned 91.23 74.50 57.41 80.91

Table 6 is a separate alignment experiment; its aligned score is not substituted for Table 5’s full-method score. Source ↗

Quantization, bit packing, and retrieval

VAR-Q quantizes the cache per channel and block, then packs low-bit values into INT32 storage. Before attention, cached values are unpacked and dequantized to BF16. Packing turns reduced precision into real KV-memory savings.

Read the algorithm
VAR-Q pipeline: quantize K and V, pack into INT32 with scale factors, store in HBM, then unpack for attention.
Figure 8 · Algorithm and system support

Examples

Infinity-8B at a 2-bit KV cache: BF16, FlexGen, KIVI and VAR-Q from the same prompt and seed. Click an image to enlarge. Paired BF16 vs. VAR-Q videos are in the comparisons at the top of the page.

Infinity-8B · next-scale · 1024×1024 · 2-bit KV cache

At 2 bits the KV cache shrinks by 87%; VAR-Q keeps the BF16 subject and layout while KIVI and FlexGen drift or break. KV memory is the paper's measurement at the evaluation batch size.

Key results

Quality and KV memory as reported in the paper, with batch size and precision stated next to every number.

Next-scale · 720p · INT379.15% less KV memory

InfinityStar: 80.59 VBench vs. 82.33 BF16.

Next-frame · 30 s · INT659.25% less KV memory

LongLive: 80.94 VBench vs. 80.92 BF16.

Next-frame · 5 s · INT474.88% less KV memory

Self-Forcing: 80.99 VBench vs. 82.94 BF16.

Video generation with low-bit KV caches (Tables 2–4)

Download results ↧

Compare KV memory and VBench Total Quality at the same precision and batch size. BF16 provides the reference for each model.

Next-scale video · INT3

InfinityStar · 720p

8B · 720p · batch 3
Method KV · GB ↓ VBench ↑
BF16 62.72 82.33
FlexGen 13.56 79.95
KIVI 13.81 75.46
VAR-Q 13.08 80.59
Full metrics and precisions ↗

Next-frame video · INT6

Self-Forcing · 5 s

1.3B · 5-second clips · batch 12
Method KV · GB ↓ VBench ↑
BF16 67.49 82.94
FlexGen 27.95 82.98
KIVI 28.22 82.85
VAR-Q 27.50 83.34
Full metrics and precisions ↗

Next-frame video · INT6

LongLive · 30 s

1.3B · 30-second clips · batch 20
Method KV · GB ↓ VBench ↑
BF16 64.28 80.92
FlexGen 26.62 80.60
KIVI 26.88 80.65
VAR-Q 26.19 80.94
Full metrics and precisions ↗

Four-bit KV caches across all nine evaluated checkpoints

Five model families, nine checkpoints: VAR-d20/d24/d30, Infinity-2B/8B, InfinityStar 480p/720p, Self-Forcing and LongLive. VAR-Q retains competitive generation quality while reducing KV memory by 72–75%.

Tables 2–4 · VAR-Q INT4 vs. BF16 · KV memory at the batch sizes reported in the paper
Model / setting Quality metric BF16 VAR-Q · INT4 KV (GB) · BF16 → VAR-Q KV saved
VAR-d20Next-scale FID ↓ 4.11 3.44 16.60 → 4.64 72.06%
VAR-d24Next-scale FID ↓ 2.97 2.52 23.91 → 6.68 72.06%
VAR-d30Next-scale FID ↓ 2.07 1.97 37.35 → 10.44 72.06%
Infinity-2BNext-scale GenEval Overall ↑ 69.61 69.35 41.10 → 10.38 74.75%
Infinity-8BNext-scale GenEval Overall ↑ 83.54 83.67 67.43 → 17.03 74.75%
InfinityStar · 720pNext-scale Total Quality ↑ 82.33 81.66 62.72 → 16.00 74.49%
InfinityStar · 480pNext-scale Total Quality ↑ 81.30 81.32 55.87 → 14.60 73.86%
Self-ForcingNext-frame Total Quality ↑ 82.94 80.99 67.49 → 16.95 74.88%
LongLiveNext-frame Total Quality ↑ 80.92 79.60 64.28 → 16.15 74.88%

FID ↓ for VAR; GenEval Overall ↑ for Infinity; VBench Total Quality ↑ for video models. Full metrics, precisions, and baselines are reproduced below.

Complete tables

Every precision, metric and baseline (BF16, FlexGen, KIVI) from Tables 2–4. Open a table to expand it.

Table 3Next-scale video
Source table ↗

At 720p with INT3 caches, VAR-Q scores 80.59 on VBench Total Quality, compared with 79.95 for FlexGen and 75.46 for KIVI, while reducing KV memory by 79.15% relative to BF16.

VAR-Q at a glance · VBench Total Quality and KV memory
Model KV precision Total Quality ↑ Δ vs. BF16 KV (GB) ↓ KV reduction
InfinityStar 720p INT4 81.66 -0.67 16.00 74.49%
InfinityStar 720p INT3 80.59 -1.74 13.08 79.15%
InfinityStar 480p INT4 81.32 +0.02 14.60 73.86%
InfinityStar 480p INT3 80.98 -0.32 12.03 78.47%
Table 3. Next-scale video generation with InfinityStar.
Method KV precision VBench ↑ KV cache
Subject Consistency Background Consistency Temporal Flickering Motion Smoothness Dynamic Degree Aesthetic Quality Imaging Quality Total Quality Memory (GB) ↓ Reduction (%) ↑
InfinityStar · Next-scale video · 720p / 8B · 720p · batch 3
Baseline BF16 93.69 94.62 97.71 98.86 67.50 58.49 66.79 82.33 62.72 —
FlexGen INT4 93.83 94.76 97.83 98.91 63.00 55.50 64.80 80.11 16.48 73.72%
KIVI INT4 92.01 94.61 97.27 98.70 68.00 56.12 63.60 81.24 16.73 73.33%
VAR-Q INT4 92.65 94.09 97.37 98.59 74.50 56.87 64.35 81.66 16.00 74.49%
FlexGen INT3 90.12 92.68 97.84 96.91 62.00 56.01 60.48 79.95 13.56 78.37%
KIVI INT3 90.45 90.42 97.65 96.79 74.00 50.97 61.92 75.46 13.81 77.99%
VAR-Q INT3 91.41 93.56 97.25 98.37 75.00 55.55 61.75 80.59 13.08 79.15%
InfinityStar · Next-scale video · 480p / 8B · 480p · batch 6
Baseline BF16 93.01 94.53 97.25 98.68 66.00 58.13 64.01 81.30 55.87 —
FlexGen INT4 90.59 94.39 95.16 98.64 66.50 56.48 62.14 78.94 15.03 73.10%
KIVI INT4 92.44 94.42 96.65 97.79 64.00 56.97 62.92 80.46 15.25 72.71%
VAR-Q INT4 92.75 94.51 97.21 98.67 68.50 57.84 63.62 81.32 14.60 73.86%
FlexGen INT3 91.89 93.91 97.24 98.62 64.00 57.54 63.69 80.64 12.45 77.71%
KIVI INT3 93.01 94.46 97.33 98.71 64.00 57.79 63.56 78.05 12.67 77.33%
VAR-Q INT3 92.86 94.30 97.26 98.64 65.50 57.52 63.37 80.98 12.03 78.47%

Bold and underline mark the best and second-best quality scores among quantized methods at the same precision, within each model. BF16 is the reference. ↑ Higher is better; ↓ lower is better. Memory reduction is relative to BF16 at the listed batch size.

Table 4Next-frame video
Source table ↗

At INT6, VAR-Q stays close to BF16 on VBench Total Quality for both 5-second and 30-second videos. INT4 reduces KV memory by 74.88%, with quality drops of 1.95 and 1.32 points, respectively.

VAR-Q at a glance · VBench Total Quality and KV memory
Model KV precision Total Quality ↑ Δ vs. BF16 KV (GB) ↓ KV reduction
Self-Forcing INT6 83.34 +0.40 27.50 59.25%
Self-Forcing INT4 80.99 -1.95 16.95 74.88%
LongLive INT6 80.94 +0.02 26.19 59.25%
LongLive INT4 79.60 -1.32 16.15 74.88%
Table 4. Next-frame video generation with Self-Forcing and LongLive.
Method KV precision VBench ↑ Claude Opus 4.7 · VLM judge ↑ (1–5) KV cache
Subject Consistency Background Consistency Motion Smoothness Dynamic Degree Aesthetic Quality Total Quality Visual Quality Dynamics Text Alignment Memory (GB) ↓ Reduction (%) ↑
Self-Forcing · Next-frame video / 1.3B · 5-second clips · batch 12
Baseline BF16 93.57 94.07 98.43 75.00 60.07 82.94 3.81 3.10 3.62 67.49 —
FlexGen INT6 93.52 94.11 98.44 76.00 60.02 82.98 3.79 3.15 3.73 27.95 58.59%
KIVI INT6 93.58 94.12 98.49 73.00 60.07 82.85 3.82 3.14 3.63 28.22 58.18%
VAR-Q INT6 93.96 94.76 98.49 76.50 61.25 83.34 3.89 3.17 3.62 27.50 59.25%
FlexGen INT4 86.39 85.84 95.40 73.00 57.39 75.91 3.60 3.40 3.75 17.40 74.22%
KIVI INT4 90.53 91.86 98.29 74.50 57.64 80.33 3.54 3.44 3.76 17.68 73.81%
VAR-Q INT4 91.28 92.55 98.35 75.50 57.40 80.99 3.55 3.50 3.73 16.95 74.88%
LongLive · Next-frame video / 1.3B · 30-second clips · batch 20
Baseline BF16 91.44 93.83 98.94 44.50 60.74 80.92 3.75 3.25 3.69 64.28 —
FlexGen INT6 91.33 93.49 98.95 46.50 60.29 80.60 3.78 3.33 3.78 26.62 58.59%
KIVI INT6 91.14 93.45 98.94 43.50 60.36 80.65 3.73 3.27 3.70 26.88 58.18%
VAR-Q INT6 91.26 93.57 98.93 47.50 60.17 80.94 3.84 3.31 3.71 26.19 59.25%
FlexGen INT4 86.72 90.19 95.83 45.00 55.16 76.59 3.48 3.48 3.77 16.57 74.22%
KIVI INT4 85.83 92.09 97.80 49.50 56.59 78.95 3.48 3.55 3.79 16.84 73.81%
VAR-Q INT4 88.39 91.97 98.80 47.50 57.11 79.60 3.54 3.57 3.80 16.15 74.88%

Bold and underline mark the best and second-best quality scores among quantized methods at the same precision, within each model. BF16 is the reference. ↑ Higher is better; ↓ lower is better. Memory reduction is relative to BF16 at the listed batch size.

VLM evaluation uses Claude Opus 4.7 (Opus4.7-VL in the paper). The judge rates visual quality, dynamics, and text alignment on a 1–5 scale using 12 uniformly sampled frames and the generation prompt. These ratings are separate from VBench scores.

Table 2aClass-conditional image generation
Source table ↗

FID measures the output distribution; PSNR, SSIM, and LPIPS measure fidelity to matched-seed BF16 outputs. Both are needed to judge quantization.

Table 2a. Class-conditional image generation with VAR.
Method KV precision Distribution quality Fidelity to BF16 KV cache
FID ↓ IS ↑ Precision ↑ Recall ↑ PSNR ↑ SSIM ↑ LPIPS ↓ Memory (GB) ↓ Reduction (%) ↑
VAR-d20 · Next-scale image / 600M · ImageNet 256×256 · batch 128
Baseline BF16 4.11 357.32 0.8650 0.5160 — — — 16.60 —
FlexGen INT4 4.03 325.39 0.8326 0.5230 11.84 0.264 0.553 4.29 74.14%
KIVI INT4 3.55 341.52 0.8458 0.5363 11.81 0.256 0.552 4.65 71.97%
VAR-Q INT4 3.44 342.77 0.8502 0.5281 13.22 0.308 0.462 4.64 72.06%
FlexGen INT3 15.33 139.59 0.5944 0.6308 10.35 0.212 0.650 3.46 79.14%
KIVI INT3 6.36 233.84 0.7260 0.5748 9.88 0.187 0.662 3.82 76.97%
VAR-Q INT3 3.77 262.24 0.7706 0.5767 12.24 0.266 0.538 3.81 77.06%
VAR-d24 · Next-scale image / 1.0B · ImageNet 256×256 · batch 128
Baseline BF16 2.97 357.24 0.8444 0.5605 — — — 23.91 —
FlexGen INT4 2.74 325.69 0.8175 0.5816 11.78 0.256 0.560 6.18 74.14%
KIVI INT4 2.56 340.55 0.8295 0.5786 11.92 0.268 0.545 6.70 71.97%
VAR-Q INT4 2.52 341.60 0.8321 0.5718 13.03 0.296 0.479 6.68 72.06%
FlexGen INT3 14.91 144.09 0.6066 0.6424 10.63 0.201 0.643 4.99 79.14%
KIVI INT3 4.29 239.40 0.7271 0.6340 10.88 0.226 0.631 5.51 76.97%
VAR-Q INT3 3.60 258.33 0.7477 0.6130 11.90 0.252 0.558 5.48 77.06%
VAR-d30 · Next-scale image / 2.0B · ImageNet 256×256 · batch 128
Baseline BF16 2.07 335.10 0.8260 0.5812 — — — 37.35 —
FlexGen INT4 2.00 307.92 0.8061 0.6006 12.47 0.277 0.510 9.66 74.14%
KIVI INT4 1.99 321.46 0.8162 0.5974 12.86 0.290 0.484 10.47 71.97%
VAR-Q INT4 1.97 323.67 0.8156 0.5961 13.54 0.312 0.437 10.44 72.06%
FlexGen INT3 9.52 180.57 0.6738 0.6417 11.33 0.232 0.597 7.79 79.14%
KIVI INT3 3.26 264.93 0.7536 0.6189 11.69 0.249 0.573 8.60 76.97%
VAR-Q INT3 3.05 268.16 0.7612 0.6307 12.41 0.270 0.517 8.57 77.06%

Bold and underline mark the best and second-best quality scores among quantized methods at the same precision, within each model. BF16 is the reference. ↑ Higher is better; ↓ lower is better. Memory reduction is relative to BF16 at the listed batch size.

Table 2bText-to-image generation
Source table ↗

GenEval measures compositional prompt following, while DPG evaluates finer-grained alignment. Infinity-8B retains a DPG Overall score of 83.83 at INT2, using 8.60 GB of KV memory versus 67.43 GB for BF16.

Table 2b. Text-to-image generation with Infinity.
Method KV precision GenEval ↑ DPG ↑ KV cache
Two-Object Position Color Overall Entity Relation Attribute Global Other Overall Memory (GB) ↓ Reduction (%) ↑
Infinity-2B · Next-scale image / Text-to-image · batch 8
Baseline BF16 80.61 25.51 88.03 69.61 86.55 89.51 86.62 83.24 90.14 81.97 41.10 —
VAR-Q INT8 79.34 26.26 87.23 69.06 87.69 89.39 88.70 86.40 85.85 82.21 20.65 49.75%
FlexGen INT4 78.57 24.75 54.50 67.75 87.39 91.56 85.38 82.24 82.04 81.01 10.60 74.22%
KIVI INT4 80.61 24.49 87.23 67.48 88.18 91.13 86.51 83.18 80.81 80.04 10.78 73.76%
VAR-Q INT4 79.34 26.01 86.97 69.35 87.24 89.60 87.43 86.75 90.13 82.02 10.38 74.75%
KIVI INT3 77.30 26.26 86.70 67.56 87.12 92.39 85.66 82.94 79.31 80.76 8.73 78.76%
VAR-Q INT3 78.32 24.24 88.03 68.04 87.74 89.92 87.01 85.13 87.01 81.18 8.32 79.75%
Infinity-8B · Next-scale image / Text-to-image · batch 6
Baseline BF16 95.41 62.12 90.43 83.54 88.27 89.30 91.57 92.83 88.61 85.72 67.43 —
VAR-Q INT8 96.43 62.88 89.89 84.00 90.57 92.57 89.23 87.12 92.68 85.97 33.88 49.75%
FlexGen INT4 95.41 60.89 89.02 82.45 90.75 92.33 89.15 85.04 84.49 84.58 17.38 74.22%
KIVI INT4 95.66 65.40 91.22 83.26 91.13 94.10 88.93 85.67 82.04 85.74 17.69 73.76%
VAR-Q INT4 95.92 63.64 91.49 83.67 90.70 91.99 91.07 87.47 89.23 85.90 17.03 74.75%
KIVI INT3 95.66 63.38 91.49 82.75 87.74 94.33 88.97 84.05 85.71 84.89 14.32 78.76%
VAR-Q INT3 95.41 63.13 90.69 83.28 89.65 88.84 90.82 89.16 91.80 85.89 13.65 79.75%
VAR-Q INT2 85.97 56.57 86.70 74.38 87.80 90.09 89.31 91.22 88.12 83.83 8.60 87.25%

Bold and underline mark the best and second-best quality scores among quantized methods at the same precision, within each model. BF16 is the reference. ↑ Higher is better; ↓ lower is better. Memory reduction is relative to BF16 at the listed batch size.

Code and citation

The runtime hooks, packed low-bit kernels and curated configs are public. Please cite the paper if VAR-Q is useful in your research. Contact the authors ↗

Get the code
git clone https://github.com/lujiaji/VAR-Q.git
cd VAR-Q && git checkout varq_official
pip install -e .
Launchers: VAR · Infinity · InfinityStar · Self-Forcing · LongLive · LiveTalk ↗Curated JSON configs (2–8 bit) ↗Runtime hooks, CUDA kernels & packing ↗

Checkpoints are loaded by each upstream repository; VAR-Q never forks their source.

BibTeX · NeurIPS 2026 
@inproceedings{xu2026varq,
  title = {VAR-Q: Tuning-free KV Cache Quantization
           for Visual Autoregressive Image and Video Generation},
  author = {Boxun Xu and Jiaji Lu and Yuxuan Yin and
            Zihu Wang and Ziyue Liu and Yu Wang and
            Zirui Liu and Peng Li},
  year = {2026},
  booktitle = {Advances in Neural Information Processing Systems},
  url = {https://boxunxu.top/assets/pdf/VAR-Q.pdf}
}