Research map / Image generation / Editing and style transfer

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

NeurIPS 2026 · Image generation › Editing and style transfer · approach: Instruction-based diffusion editing · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Training-free text-guided image editor for bitwise-residual VAR models (Infinity-2B) that tilts per-bit Bernoulli log-odds with a source-negative, KL-bounded guidance and writes mask-gated code-space residuals into the multi-scale sum.

Problem

Text-guided image editing with visual autoregressive generators lacks methods that exploit the per-bit Bernoulli prediction head and additive multi-scale residual code structure; existing VAR editors operate on tokens, features, or flat logits and preserve background poorly or align text poorly.

Approach

Two components: BitEdit applies source-negative contrast tilting in per-bit log-odds before sampling, projected onto a per-bit Bernoulli-KL trust region around the CFG anchor; ResEdit replaces token reassembly with a mask-gated continuous-code residual against the cached source code, injected via Infinity's sum-of-scales.

Key results

• On PIE-Bench, BitResEdit reaches CLIP-Whole 27.13 and CLIP-Edited 24.54, the best among Infinity-2B editors (+1.07 CLIP-Edited over EditInfinity per the paper).
• SSIM 91.71 is within 0.42 of EditInfinity (92.12) while being training-free.
• Editing one image takes 10.41 s on 4 A100s versus about 467 s for per-image-tuned EditInfinity (~45x faster).

Surprising findings

• The KL trust-region projection is essentially inactive at the default radius, and removing it changes metrics within seed noise.
• Replacing the GT mask with a tight bounding box collapses background preservation (PSNR 19.68) even though it slightly raises CLIP scores.

Limitations

• Requires a user-supplied localization mask; mask quality directly bounds preservation and edit quality.
• Weaker on object deletion (lowest CLIP-Edited among edit types), since the subtractive prompt gives no positive content to fill the mask.

Experimental setup

Datasets: PIE-Bench

Benchmarks: PIE-Bench

Metrics: PSNR LPIPS MSE SSIM CLIP-Whole CLIP-Edited

Baselines: AREdit VAREdit-8B EditInfinity MLN FlowChef VAGS RewardFlow FlowEdit Prompt-to-Prompt MasaCtrl NTI PnP-Inv

Builds on: Infinity AREdit classifier-free guidance Binary Spherical Quantizer FlowEdit

Closest papers by meaning

• Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.88)
• FlexVAR: Flexible Visual Autoregressive Modeling without Residual Prediction (NeurIPS 2025 · similarity 0.87)
• EditAR: Unified Conditional Generation with Autoregressive Models (CVPR 2025 · similarity 0.87)
• NEP: Autoregressive Image Editing via Next Editing Token Prediction (NeurIPS 2025 · similarity 0.87)
• Visual Autoregressive Modeling for Image Super-Resolution (ICML 2025 · similarity 0.86)
• The Promise of RL for Autoregressive Image Editing (NeurIPS 2025 · similarity 0.86)