CoANeRV: One Feed-Forward Pass, One Shared Decoder, No Per-Video Optimisation

What happened

CoANeRV, published on arXiv on 17 August 2026 under cs.CV, encodes a video into a fixed set of compact tokens in a single feed-forward pass and reconstructs arbitrary spatio-temporal coordinates through one shared decoder. The paper's central claim: this amortized pattern improves reconstruction quality over both per-video neural-field optimisation and prior feed-forward NeRV baselines while cutting peak attention memory.

Context

Neural video representation sits between classical codecs and implicit neural representations. The standard formulation fits a small network per video — sequential, expensive, and awkward beyond a handful of clips. INR methods avoid per-instance training but trade coordinate-level fidelity for a one-shot encode. CoANeRV arrives in a subfield that has been iterating on exactly this trade-off: keep the continuous-query flexibility of a neural field without paying the per-video optimisation cost. It positions itself as the next step in the token-conditioned NeRV line, where a learned token set replaces per-video weights.

How it works

CoANeRV takes a video and produces a compact set of tokens in one forward pass of a feed-forward encoder. A single shared decoder then maps arbitrary spatio-temporal coordinate queries back to pixel values, so the same decoder serves every video — no per-clip fine-tuning, no per-clip weight generation.

Two design choices set it apart from earlier token-conditioned NeRV work. Query-to-token alignment uses axis-adaptive positional encoding combined with temperature-modulated cross-attention, letting the decoder resolve a given (x, y, t) query against the token set without a bespoke per-video attention map. Coordinate queries are processed in blocks rather than all at once, capping peak attention memory and making high-resolution reconstruction tractable.

The paper reports consistent reconstruction-quality gains over prior feed-forward NeRV baselines and INR baselines across diverse video datasets, and lower peak memory than attention-based coordinate decoders. No dataset names, PSNR or SSIM figures, or token counts appear in the abstract.

Our read

The interesting question is not whether CoANeRV beats its baselines — the abstract says it does. The interesting question is what "amortized encoding without per-video optimisation" buys a production pipeline.

Honestly, not much yet. CoANeRV compresses a clip into tokens and reconstructs it. It does not generate new frames or condition a diffusion model. A ComfyUI graph running a generation model has no slot where a token-based NeRV decoder replaces a component. The paper's "efficient amortized encoding" is an efficiency argument inside a research subfield, not a pipeline change.

Where the architecture does matter is in the pattern it validates. Single-feed-forward-pass encoder, shared decoder, compact token set — that is the shape a video-to-video conditioning signal would take: compress a reference clip into tokens, feed them as context, let the generation model do the creative work. The block-wise coordinate querying is also a borrowable trick: cap attention memory by processing queries in chunks rather than all at once.

The abstract offers no number that would change a procurement decision. "Consistently improves" is qualitative. Without a PSNR table, a token-budget figure, or a VRAM ceiling, the practical takeaway is to read the full paper, note the architecture, and file it under representation subfield.

What this changes

Nothing on Monday. The paper names no dataset, gives no PSNR or SSIM figure, states no minimum GPU VRAM, ships no ComfyUI node, and links no pretrained weights with a usable licence. The code URL is redacted in the abstract. If your pipeline generates video through a diffusion or autoregressive model, CoANeRV does not slot in.

The one concrete action: if you are building or evaluating a video-to-video conditioning path — compressing a reference clip into a compact signal for a downstream model — the token-encode-then-shared-decode pattern and the block-wise attention trick are the two architectural ideas worth pulling from the full paper. Read it as a reference, not a dependency.

License

The sources do not state a licence. The abstract confirms code is available but the repository URL is redacted, and no licence identifier appears in the provided text. Check the code repository before building anything commercial on it.

Key takeaways

• CoANeRV encodes a video into compact tokens in one feed-forward pass and reconstructs it through a single shared decoder, eliminating per-video optimisation and per-video weight generation.
• Axis-adaptive positional encoding and temperature-modulated cross-attention align spatio-temporal queries to tokens; block-wise querying caps peak attention memory so high-resolution reconstruction stays tractable.
• The method is a representation and reconstruction technique, not a generative model — it does not produce new frames or condition a diffusion pipeline.
• No quantitative metrics, dataset names, GPU requirements, or licence appear in the abstract; the full paper and code repository are needed to evaluate practical utility.
• The token-encode-then-shared-decode architecture is a pattern worth tracking for future video-to-video conditioning workflows, but there is nothing to integrate into a production graph today.

Sources

1. CoANeRV: Coordinate-Aware Token-Space Neural Video Representation — tier 1