Edit the Bits, Diff the Codes: Bitwise Residual Editing for Visual Autoregressive Models

2026-06-11

AI Summary

Existing vision autoregressive models struggle to simultaneously achieve precise edits and preserve background content in text-guided image editing, primarily due to underutilization of bit-level prediction and multi-scale residual code structures. This work proposes BitResEdit, a training-free editing method that, for the first time, jointly leverages bit-level Bernoulli prediction heads and additive multi-scale residual code fields to enable synergistic decision guidance and code composition during inference. By integrating bit-level negative guidance (BitEdit), Bernoulli-KL confidence-region projection, local masked injection of residual codes (ResEdit), and multi-scale reconstruction, the approach significantly improves CLIP scores in edited regions by +1.07 on PIE-Bench using the Infinity-2B model, while maintaining strong background preservation.

Abstract

Text-guided image editing with visual autoregressive (VAR) generators requires controlling both what the model samples and where the sampled change is written back into the image code. Existing VAR editors mainly operate on token streams, features, or flat next-token logits, leaving two native structures of bitwise-residual VAR models underused: the per-bit Bernoulli prediction head and the additive multi-scale residual code field from which the image is assembled. We propose BitResEdit, a training-free editor for bitwise-residual VAR generators such as Infinity. BitEdit performs source-negative guidance by tilting the post-CFG per-bit log-odds along a source--target contrast computed on a shared edited prefix, then projects each update into a closed-form Bernoulli-KL trust region around the clean CFG sampler. ResEdit converts the sampled bits into per-scale continuous-code residuals, gates them with a localization mask, and re-injects them through the generator's native sum-of-scales. Together they couple decision-time bit guidance with combination-time code composition, so masked-out latent features are preserved exactly by code arithmetic while localized, scale-aware edits are applied inside the target region. On PIE-Bench with Infinity-2B, BitResEdit attains the strongest text alignment among same-backbone VAR editors, improving CLIP on the edited region by +1.07 over the strongest prior editor while keeping background preservation competitive with it. Ablations show BitEdit and ResEdit play complementary roles in target alignment and background preservation.

Problem

• visual autoregressive models
• text-guided image editing
• bitwise residual
• code composition
• localization

Innovation

• bitwise residual editing
• visual autoregressive models
• training-free editing
• multi-scale residual codes
• Bernoulli-KL trust region