MaskGIL: MAR Image Generation Framework

• The paper introduces MaskGIL, a MAR image-generation system that iteratively refines masked tokens using a bidirectional LLaMA backbone, reducing decoding from 256 to 8 steps.
• The methodology replaces causal with bidirectional attention and incorporates 2D RoPE, enabling efficient context use and competitive FID performance.
• MaskGIL scales from 111M to 1.4B parameters, achieving near state-of-the-art quality with versatile applications including text-conditioned and speech-to-image generation.

MaskGIL is a Masked AutoRegressive (MAR) image-generation framework built atop an enhanced bidirectional LLaMA backbone. It refines MAR architecture by replacing causal attention with bidirectional attention and incorporating 2D RoPE, and it is positioned as a system for efficient and scalable image generation. Scaled from 111M to 1.4B parameters, it achieves a FID score of 3.71 on ImageNet $256\times256$ while requiring only 8 inference steps compared to the 256 steps of autoregressive (AR) models; the same framework is also extended to text-driven generation, hybrid AR acceleration, and real-time speech-to-image conversion (Xin et al., 17 Jul 2025).

1. Generative formulation

MaskGIL is based on the MAR objective rather than the standard left-to-right AR factorization. For a tokenized image sequence $x=(x_1,\dots,x_N)$, possibly conditioned on context $c$, the AR baseline is written as

$p(x\mid c)=\prod_{i=1}^N p(x_i\mid x_{<i},c).$

MAR generation instead predicts subsets of tokens in parallel over $T$ refinement iterations. At iteration $t$, a subset of positions $M^{(t)}\subset\{1,\dots,N\}$ is masked and predicted conditioned on the complement $\overline{M^{(t)}}$. Training samples a random mask $M$ and applies cross-entropy only on masked tokens: $$\mathcal{L}(\theta) = -\mathbb{E}_{x,c}\;\mathbb{E}_{M}\; \sum_{i\in M} \log p_\theta\bigl(x_i \mid x_{\setminus M}, c\bigr).$$

This objective is analogous to BERT-style masked prediction but coupled to a generative decode. In the context of image token generation, the central distinction from AR is that MaskGIL does not commit to a single causal ordering of the entire sequence. Instead, it repeatedly reconstructs masked subsets conditioned on the currently visible tokens, which is the mechanism that enables fixed-step parallel decoding (Xin et al., 17 Jul 2025).

2. Bidirectional LLaMA backbone

The architectural core of MaskGIL is an improved bidirectional LLaMA. The causal Transformer in LLaMA is replaced with bidirectional attention so that visible context can be leveraged both forward and backward. The attention mask is specified so that

$x=(x_1,\dots,x_N)$0

The intended effect is to expose each prediction to the currently observed token configuration rather than to a purely prefix-conditioned context.

MaskGIL also incorporates 2D Rotary Position Embeddings (RoPE). Each token at spatial coordinate $x=(x_1,\dots,x_N)$1 receives a sinusoidal embedding $x=(x_1,\dots,x_N)$2, added to the query and key vectors in every head. For the query and key representations, the rotary transform is applied as

$x=(x_1,\dots,x_N)$3

The paper states that this preserves 2D translation equivariance across the grid of image tokens.

Within the overall MAR framework, these modifications are significant because they make the LLaMA-style decoder compatible with iterative masked refinement on image-token lattices rather than with only one-directional next-token prediction. A plausible implication is that the design attempts to recover the context utilization advantages of bidirectional masked modeling while retaining a tractable iterative generative procedure.

3. Tokenization and the 8-step decoding procedure

MaskGIL is paired with an explicit tokenizer study. Four VQ-VAE-style tokenizers, all downsampling by $x=(x_1,\dots,x_N)$4, were compared on ImageNet $x=(x_1,\dots,x_N)$5: MaskGIT-VQ with codebook $x=(x_1,\dots,x_N)$6, Chameleon-VQ with $x=(x_1,\dots,x_N)$7, LlamaGen-VQ with $x=(x_1,\dots,x_N)$8, and Open-MAGVIT2-VQ with $x=(x_1,\dots,x_N)$9. The evaluation criteria were reconstruction rFID, codebook usage, and end-to-end generative FID/IS under both AR and MAR models. The reported rFID values were 10.79 for MaskGIT-VQ, 8.34 for Chameleon-VQ, 4.54 for LlamaGen-VQ, and 4.03 for Open-MAGVIT2-VQ. LlamaGen and Open-MAGVIT2 reached 100% codebook usage, and LlamaGen-VQ consistently yielded the best FID/IS trade-off; accordingly, LlamaGen-VQ was adopted for subsequent MaskGIL models (Xin et al., 17 Jul 2025).

Inference is performed with a fixed $c$0 parallel refinement schedule rather than with token-by-token decoding. Starting from a blank canvas $c$1 with all $c$2 tokens masked, the model predicts logits for all masked positions, samples token values, computes a confidence score $c$3, determines the number of positions to remask as $c$4, and then remasks the lowest-confidence positions. The schedule function $c$5 may take root, linear, cosine, square, or arccos forms. Eight steps were reported to balance quality and speed, while additional steps yielded diminishing returns.

This decoding regime is the operational core of MaskGIL’s efficiency claim. Rather than committing to one irreversible token at each step, it repeatedly updates uncertain positions, which is the mechanism by which MAR seeks to narrow the quality gap with AR without reintroducing 256 sequential iterations.

4. Scaling behavior and quantitative performance

The paper reports four class-conditional MaskGIL variants, all evaluated with 8 inference steps and CFG guidance applied. MaskGIL-B has 111M parameters and reaches FID 5.64 and IS 229.96. MaskGIL-L has 343M parameters and reaches FID 4.01 and IS 281.11. MaskGIL-XL has 775M parameters and reaches FID 3.90 and IS 296.25. MaskGIL-XXL has 1.4B parameters and reaches FID 3.71 and IS 303.47 (Xin et al., 17 Jul 2025).

Training stability at this scale is attributed to Query-Key Normalization and Post-Norm, with the paper stating that training remains stable up to 1.4B. At the same time, the reported marginal FID gains beyond 775M–1.4B suggest dataset size constraints on scaling. In other words, the scaling curve is favorable but not unbounded within the reported training regime.

The main comparative point is that MaskGIL is presented as narrowing the longstanding gap between MAR and AR quality. Earlier MAR systems had often been treated as faster but weaker alternatives. The reported results complicate that view: MaskGIL-XXL reaches FID 3.71 in 8 steps, whereas the paper lists LlamaGen-XXL, an AR baseline with 1.4B parameters, at FID 2.34 with 256 steps. This does not eliminate the AR advantage in absolute FID, but it materially changes the quality-efficiency frontier.

5. Relation to prior AR and MAR image generators

The comparative baseline set in the paper includes both AR and MAR systems. Among AR baselines using 256 steps, RQTransformer with 3.8B parameters is reported at FID 7.55, and LlamaGen-XXL with 1.4B parameters is reported at FID 2.34. Among MAR baselines using 8–20 steps, MaskGIT with 227M parameters and 8 steps is reported at FID 6.18, and MAGE with 230M parameters and 20 steps is reported at FID 6.93 (Xin et al., 17 Jul 2025).

Against these baselines, MaskGIL-L at 343M parameters achieves FID 4.01 in 8 steps, improving on MaskGIT’s 6.18 under the same step count. At the high end, MaskGIL-XXL at 1.4B achieves FID 3.71, which the paper describes as approaching AR quality at 8× faster decoding. The abstract also states that MaskGIL achieves FID 3.71, matching state-of-the-art AR models in the ImageNet $c$6 benchmark, while requiring only 8 inference steps compared to the 256 steps of AR models.

A common misconception in this area is that MAR models are intrinsically limited to materially lower image quality than AR models. The results reported for MaskGIL do not remove all quality differences, but they directly challenge that categorical characterization. The paper further notes a minor diversity–fidelity trade-off in qualitative comparison, while describing MaskGIL images as exhibiting crisp detail and diversity comparable to AR models.

6. Text-conditioned generation and broader applications

MaskGIL is also extended to text-driven generation with a 775M-parameter model. The text-conditional architecture uses the same bidirectional LLaMA backbone, prepends a learnable text-embedding prefix, and applies classifier-free guidance by dropping text with probability 0.1. Training uses cross-entropy on masked tokens, AdamW $c$7, batch size 256, and various image resolutions, including $c$8, $c$9, and up to $p(x\mid c)=\prod_{i=1}^N p(x_i\mid x_{<i},c).$0. The conditioning modality is free-form English text (Xin et al., 17 Jul 2025).

On the GenEval benchmark, the reported overall scores are 0.50 for SDv2.1 at 0.9B parameters, 0.48 for PixArt-$p(x\mid c)=\prod_{i=1}^N p(x_i\mid x_{<i},c).$1 at 0.6B, 0.32 for LlamaGen (AR) at 0.8B, 0.56 for Lumina-mGPT at 7.0B, and 0.49 for MaskGIL at 0.7B. The sample results are described as photorealistic and as exhibiting high text alignment at multiple resolutions.

Beyond standalone generation, the framework is used for hybrid AR+MAR inference. The procedure first runs an AR model for $p(x\mid c)=\prod_{i=1}^N p(x_i\mid x_{<i},c).$2 tokens, after which MaskGIL completes the remaining $p(x\mid c)=\prod_{i=1}^N p(x_i\mid x_{<i},c).$3 fraction in 8 steps. On ImageNet $p(x\mid c)=\prod_{i=1}^N p(x_i\mid x_{<i},c).$4 with a LlamaGen-L backbone, the paper reports the following trade-off: 100% AR uses 256 sample steps and reaches FID 3.80; 75% AR uses 200 steps and reaches FID 3.96 with +21.9% speedup; 50% AR uses 136 steps and reaches FID 4.23 with +46.9% speedup; 25% AR uses 72 steps and reaches FID 4.68 with +71.9% speedup; and 0% AR corresponds to 8 steps with FID 5.64 and +96.9% speedup. This suggests a tunable continuum between AR fidelity and MAR efficiency.

The reported real-time speech-to-image pipeline proceeds from audio to a Whisper transcriber/translator, then to an English text prompt, and finally to MaskGIL. The tested languages are English, Chinese, Portuguese, French, and Spanish, with 100 samples each. Evaluation uses CLIPScore, where Image CLIPScore shows cross-modal consistency at approximately 0.97–1.00 and Text CLIPScore at approximately 30–31 indicates preserved semantic alignment. Within the scope of the reported experiments, this positions MaskGIL not only as an image generator but as a general masked-refinement engine for multimodal generation workflows.