Ansatz: Neural 4K Talking-Head Video Reconstruction

A technical report from Inverity

Author: Brandon Cade, Founder, Inverity.

Abstract

Conventional video codecs compress a talking head the same way they compress anything else: by storing a more efficient version of the pixels. We present Ansatz, a generative reconstruction codec that takes a different path for this content. Rather than storing the pixels, it transmits a reference frame and a compact per-frame motion representation, and regenerates the face on the other side. On a mechanically sampled population of 206 single-face 4K talking-head clips, Ansatz is a median ~65% smaller than AV1 and 49% smaller than x265 at faithful reconstruction, beats AV1 on 97% of clips, and reaches up to roughly ~98% smaller than AV1 on the clips where its representation applies most fully. The reconstruction method is proprietary and is not described here; the contribution of this report is the measured result, the honest distribution behind it, and the safety mechanism that keeps it production-viable.

Abstract

Conventional video codecs compress a talking head the same way they compress anything else: by storing a more efficient version of the pixels. We present Ansatz, a generative reconstruction codec that takes a different path for this content. Rather than storing the pixels, it transmits a reference frame and a compact per-frame motion representation, and regenerates the face on the other side. On a mechanically sampled population of 206 single-face 4K talking-head clips, with the reference frame compressed by Epiron, our neural still codec, Ansatz is a median 65% smaller than AV1 and 60% smaller than x265 at faithful reconstruction, beats AV1 on essentially every clip, and reaches up to roughly 98% smaller than AV1 on the clips where its representation applies most fully. The reconstruction method is proprietary and is not described here; the contribution of this report is the measured result, the honest distribution behind it, and the safety mechanism that keeps it production-viable.

1. What is claimed, and what is not

We are precise about scope because the result is a targeted one, not a universal one.

Claimed: on single-face talking-head video at 4K, generative reconstruction achieves a compression margin over the best conventional codecs (x265, AV1) that conventional-codec tuning cannot reach, at a fidelity a viewer accepts. This is a decisive win in a specific cell of content, not an average across all video.

Not claimed: Ansatz is not a general-purpose 4K codec; this is the talking-head case only. In its current form it saves storage and origin bandwidth rather than last-mile user bandwidth, because generative reconstruction runs on a GPU at decode. And the reported facial detail is bounded by the reconstruction model rather than being native-4K across the face.

2. Why a different representation wins here

A talking head is one of the most structured things on the internet: a mostly static background and a face that moves in a constrained, learnable way. Conventional codecs do not exploit that structure; they encode the pixels regardless of what the pixels depict. Ansatz does exploit it. It sends the identity and appearance once, as a reference, then per frame sends only a compact description of how the face has moved and changed, and reconstructs each frame from the two. The per-frame cost drops from the size of a compressed image to the size of a small motion representation. The savings margin that produces is structural, it comes from representing the content differently, which is why no amount of conventional-codec tuning reaches it.

3. Method of evaluation

3.1 Population, not best case

The corpus is 206 single-face 4K talking-head clips harvested and filtered by a mechanical face gate: clips were kept only at 4K height with a face present in the majority of sampled frames at a minimum in-frame size. Of 358 candidates, roughly 56% were rejected before evaluation. Critically, the population is whatever cleared the mechanical gate, not a hand-picked set. We report the distribution, not a best-case mean.

3.2 Baselines and metrics

Baselines are x265 and SVT-AV1 at fixed, near-transparent quality settings, a customer-shippable bar. Savings are the byte reduction of Ansatz output against each baseline over the same frames.

For quality we report decoded identity and landmark alignment, the codec's own acceptance metrics for a face reconstruction, and LPIPS as the perceptual metric. We deliberately do not use PSNR, SSIM, or VMAF here, and the reason matters: those metrics assume pixel correspondence between output and original. A generative codec regenerates the face rather than reproducing pixels, so pixel-correspondence metrics read catastrophically low on output that is visually faithful.

LPIPS measures perceptual similarity without requiring pixel alignment, which is the correct ruler for reconstructed content. We verified this during bring-up and dropped the pixel-correspondence metrics from the population run as structurally inapplicable.

4. Results

On the final population (n=198 after decode-error exclusions), against the AVIF-referenced measurement:

Ansatz beats AV1 on essentially every clip (197/198) at faithful reconstruction. The p90 against AV1 reaches roughly 98%, which is the basis for the up-to figure: on the clips where the representation applies most fully, the file is a fraction of the AV1 size.

Epiron matters here because the reference frame is a meaningful part of the total cost, and Epiron compresses it materially smaller than the best conventional still codec. At equal reference quality, the conventional still baseline needs roughly 2.2 times the bytes of Epiron. That efficiency flows straight through to the totals above, most on high-coverage clips where the reference is the dominant cost.

4.1 The tradeoff, quantified

The size of the win scales with how much of a clip the codec reconstructs generatively versus falling back to conventional coding. The more the content suits reconstruction, the larger the win, up toward the high-90s percent against AV1 on the best-fit clips, at a modest and still-good perceptual cost (LPIPS held well under the acceptance bar). Clips less suited to it (multiple faces, extreme turn-away) fall back to conventional coding: a smaller win, but pixel-sharp. This is the design working as intended, win decisively where the representation applies, degrade safely where it does not.

4.2 A note on precision

The Epiron figures above are the final end-to-end measurement, with the Epiron reference fully in the reconstruction loop, so identity and perceptual scores are measured against the true Epiron-decoded reference. These are the numbers that matter: Epiron is the production reference, and this is the complete result, not an estimate.

For reference, a deliberately conservative measurement using a high-quality conventional still proxy instead of Epiron produces a lower floor (median around 57% against AV1, up to roughly 86% on high-coverage clips). Epiron is the production reference and the numbers that matter; the proxy floor is included only to show the result holds well even under a pessimistic reference assumption.

5. The safety mechanism

A generative codec that can rebuild a face can rebuild it wrong, and unlike conventional compression, a generative failure can look clean while being subtly false. Ansatz is built around that risk. Every reconstructed segment is checked against the original on identity and alignment before it is accepted. Segments that pass ship as reconstruction.

Segments that fail are not shipped as questionable reconstructions; the codec falls back to conventional coding for those segments and sends real pixels. The result is a hybrid that captures the large savings where reconstruction is verifiably faithful and stays pixel-accurate where it is not. This verify-and-fallback floor is what makes the approach production-viable rather than a demo.

6. Limitations

Stated plainly, as they bound the claims.

• Scope is single-face talking-head at 4K. This is a content-specific result, not a general codec.
• Storage and origin bandwidth, not last-mile. Generative decode currently requires a GPU, so v1 saves storage and origin egress, not bytes to the end user.
• Baselines are fixed-quality, not iso-quality anchored. Matched-quality anchoring is a planned follow-up.
• Facial detail is bounded by the reconstruction model. Background is native 4K; native-4K facial sharpness would require an additional stage.

7. Conclusion

For talking-head video, storing the pixels is not the only option, and it is not the most efficient one. By transmitting a reference and a compact motion representation and reconstructing the face, Ansatz is a median ~65% smaller than AV1 across a mechanically sampled 4K population, up to roughly ~98% smaller where its representation fits best, at a fidelity viewers accept, with a verify-and-fallback floor that keeps it honest.

The margin is structural, not incremental, because it comes from representing the content differently rather than coding the pixels better. The method is proprietary; the result, and the honest distribution behind it, is what we report here.

For correspondence regarding evaluation or independent verification, contact the author.