Research map / Image generation / Autoregressive and latent generation

Fixed-Point Masked Generative Modeling

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Accelerating autoregressive image generation · 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.

Proposes Fixed-Point Masked Generative Models (FP-MGMs) with CoFRe, a training-to-inference recipe combining a weight-shared fixed-point denoiser, a cross-step consistency loss, and three-state reuse of solver states, reducing cost while improving low-budget generation on OWT and ImageNette.

Problem

Masked generative models need full bidirectional transformer passes at every denoising step, which makes training costly and degrades quality under low sampling compute budgets.

Approach

Replace the middle layers of a masked denoiser with a weight-shared fixed-point block whose solver iteration count sets effective depth per step, add a cross-step hidden-state consistency loss, and warm-start the solver with token-type-dependent reuse (visible, still-masked, newly revealed).

Key results

• On OWT at budget 96, CoFRe reaches generative perplexity 101.8 vs 193.1 for MDLM+SDTT, with 38.8% fewer parameters for FP-MDLM relative to MDLM.
• On ImageNette, CoFRe improves FID over MaskGIT-Large at every budget (e.g., 174.1 to 96.7 at budget 48) while cutting training time from 17h46m to 9h08m and VRAM from 72.45 to 35.74 GiB.
• Converting a pretrained MDLM into FP-MDLM with 40k adaptation steps lowers generative perplexity at budget 96 from 298.6 to 286.4 with initialization and 3SR.

Surprising findings

• Full reuse of fixed-point solutions can hurt generation quality at larger budgets for the base FP-MDLM, and reuse only becomes beneficial after adaptation.
• Wall-clock gains do not always follow from transformer-block forward pass reductions because fixed-point solvers add control-flow overhead.

Limitations

• Experiments limited to OWT-scale language modeling and ImageNette image generation; no evidence at larger scale or multimodal settings.
• Three-state reuse is designed for monotonic masked decoding and may not transfer to samplers that remask tokens.
• Best FP-MDLM results require a short consistency post-training stage whose duration must be tuned to avoid entropy collapse.

Experimental setup

Datasets: OpenWebText ImageNette

Benchmarks: OpenWebText ImageNette

Metrics: generative perplexity unigram entropy FID IS training time VRAM

Baselines: MDLM + SDTT MaskGIT-Large MDLM

Builds on: MDLM MaskGIT Fixed-Point Diffusion Models Stochastic Jacobian-Free Backpropagation SDTT

Closest papers by meaning

• Plug-and-Play Context Feature Reuse for Efficient Masked Generation (NeurIPS 2025 · similarity 0.90)
• Unlocking the Capabilities of Masked Generative Models for Image Synthesis via Self-Guidance (NeurIPS 2024 · similarity 0.88)
• Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · similarity 0.88)
• Co-GRPO: Co-Optimized Group Relative Policy Optimization for Masked Image Generation (NeurIPS 2026 · similarity 0.88)
• Efficient Generative Modeling with Residual Vector Quantization-Based Tokens (ICML 2025 · similarity 0.87)
• Masked Generative Nested Transformers with Decode Time Scaling (ICML 2025 · similarity 0.87)