Wan-Image: Unified Visual Generation System

Updated 5 July 2026

Wan-Image is a unified visual generation system that integrates semantic planning and precise pixel synthesis for professional image workflows.
It employs a four-channel VAE for RGBA modeling, ensuring high-fidelity text rendering, clear typography, and accurate transparent edges.
The system leverages a multi-stage data pipeline and refiners to deliver interactive editing, identity preservation, and coherent image-series generation.

Wan-Image is a unified visual generation system from Alibaba Group intended for image generation workflows that require exact controllability, complex typography rendering, strict identity preservation, localized editing, structured multi-subject composition, and reliable multi-image consistency. It combines semantic reasoning and precise pixel generation in a single system through an MLLM-based Planner, a DiT-based Visualizer, a four-channel VAE for RGBA latent modeling, a Prompt Enhancer, and an Image Refiner, with the stated goal of serving as an intelligent design assistant for real-world content creation (Mao et al., 21 Apr 2026).

1. Position within the Wan model family

Wan-Image belongs to the broader Wan ecosystem, but it is not interchangeable with the video-centric Wan suite or with Wan-Weaver. The earlier Wan foundation suite is framed primarily around video generation, with joint image-video training, a DiT backbone, flow matching, a novel 3D causal Wan-VAE, and downstream branches such as Wan-I2V for image-to-video (Wan et al., 26 Mar 2025). By contrast, Wan-Image is explicitly framed as a unified visual generation system for professional image-centric workflows, including typography-heavy generation, interactive editing, image-series generation, and native alpha-channel synthesis (Mao et al., 21 Apr 2026).

A common source of terminological confusion is the proximity between Wan-Image and Wan-Weaver. Wan-Weaver addresses interleaved text-image generation through a planner-plus-visualizer decomposition, dense prompts enclosed by <imagine> ... </imagine>, and image insertion markers such as <BOI> (Xing et al., 26 Mar 2026). Wan-Image also uses Planner-triggered image insertion and resumes generation after the Visualizer outputs an image, but its paper frames the system more broadly as a production-oriented visual generator rather than as a dedicated interleaved-generation model (Mao et al., 21 Apr 2026). This suggests that Wan-Image and Wan-Weaver share architectural motifs within the same ecosystem while targeting different operating regimes.

The distinction from Wan-I2V and its extensions is similarly important. Wan-I2V, PUSA, and Wan-Move are centered on image-to-video generation and motion control, whereas Wan-Image is centered on still-image synthesis, editing, series generation, and layout-sensitive production tasks (Wan et al., 26 Mar 2025, Liu et al., 22 Jul 2025, Chu et al., 9 Dec 2025). In that sense, Wan-Image is best understood as the image-specialized, productivity-oriented branch of a larger Wan research program.

2. System architecture and token-level interaction

Wan-Image adopts a unified architecture that conceptually separates cognition from synthesis while keeping them in one model family. The system combines an MLLM-based Planner, a DiT-based Visualizer trained with rectified flow, a four-channel VAE, a Prompt Enhancer, and an Image Refiner (Mao et al., 21 Apr 2026).

Component	Function	Notable implementation detail
Planner	Semantic understanding, task routing, prompt expansion, resolution inference, chain-of-thought style planning	MLLM-based
Visualizer	High-fidelity image generation	DiT trained with rectified flow
VAE	Latent representation for RGBA images	Four-channel design
Prompt Enhancer	Rewrites sparse requests into structured prompts	Supports T2I, I2I, T2S, TI2S
Image Refiner	Reconstructs and sharpens outputs at 2K–4K	Repair and high-fidelity modes

The architecture is described as a unified Transformer framework in which dedicated Transformer experts share attention within blocks. The understanding branch uses a standard decoder-only Transformer, while the generation branch uses a DiT trained with rectified flow. The generation branch is initialized from the well-trained understanding branch to preserve representation consistency, then separately optimized (Mao et al., 21 Apr 2026).

Attention is differentiated by token type. Causal attention is used among text, ViT, and VAE tokens to preserve autoregressive behavior. Bidirectional attention is used inside the visual tokens of each image. For image-series generation, bidirectional attention is extended across all images in the same series to improve identity and style coherence. For editing tasks, clean VAE tokens encoded from the input image are injected into the DiT as additional conditional inputs. Both clean and noised VAE tokens receive timestep embeddings and modulate each DiT block through AdaLN, and the timestep for clean VAE tokens is set to zero so that they act as deterministic conditions preserving structure and content (Mao et al., 21 Apr 2026).

Positional encoding is also split by branch. In the understanding branch, image pixels are encoded by ViT and jointly position-encoded with text using MRoPE. In the generation branch, a 3D-RoPE scheme separates temporal dimension $T$ and spatial dimensions $H, W$ , and a fixed positional offset is inserted along the $T$ -axis between text and image segments as a semantic boundary. Contextual representations from understanding are integrated into the generation branch to form a unified positional frame (Mao et al., 21 Apr 2026).

3. RGBA latent modeling and high-fidelity reconstruction

A central technical contribution is the four-channel VAE for RGBA image generation. The paper argues that conventional VAEs struggle with small-font text, dense layouts, fine textures, transparent regions, and crisp boundaries. Wan-Image’s VAE explicitly models alpha transparency, making it suitable for layered design assets, e-commerce graphics, and other production-oriented PNG content (Mao et al., 21 Apr 2026).

The VAE uses a $2 \times 2$ patchify input scheme, an $8 \times 8$ spatial downsampling backbone, overall $16 \times 16$ compression, and multi-stage residual blocks in encoder and decoder. It is trained with a hybrid reconstruction loss that jointly constrains visible-content reconstruction, alpha structure recovery, and transparent-boundary quality. Training is progressive: a $256\times256$ stage, a mixed-resolution stage, and a mixed-resolution stage plus a 4-channel GAN discriminator (Mao et al., 21 Apr 2026).

The reported reconstruction metrics position the VAE as a substantial subsystem rather than an implementation detail. On 5,000 RGB images at 512p, the Wan-Image VAE achieves PSNR 35.429, SSIM 0.958, and LPIPS 0.010, outperforming FLUX.1 VAE, SD3.5 VAE, Qwen-Image-Layered VAE, HunyuanImage-3.0 VAE, and Wan2.2 VAE. On 500 RGBA images at 1080p, it achieves PSNR 38.274, SSIM 0.970, and LPIPS 0.018, again leading the compared RGBA baselines. The qualitative claim attached to these results is that it preserves sharper glyph boundaries and cleaner transparent edges (Mao et al., 21 Apr 2026).

These results clarify why Wan-Image emphasizes typography, transparency, and design-oriented outputs. The VAE is not merely a compression layer for diffusion; it is a task-aligned latent interface designed for image regimes in which edge acuity, alpha fidelity, and layout integrity are first-order constraints.

4. Data pipeline, annotation system, and training recipe

Wan-Image attributes a substantial part of its capability to a large, carefully structured data pipeline. On the understanding side, the authors construct general understanding data, text-proxy data, and a user-aligned query set. General understanding data consists of filtered and balanced text-only and image-text QA data, with low-quality answers regenerated and filtered using LLMs and MLLMs; explicit reasoning traces enclosed in > ... are generated to create think-augmented supervision. Text-proxy data uses dense visual prompts for generation-oriented tasks, with generated image slots marked with <BOI> and paired with descriptions in <imagine> ... </imagine>. The user-aligned query set annotates raw queries by task category, difficulty, whether numerical or logical reasoning is needed, and whether the query is purely VQA-like, followed by balanced sampling (Mao et al., 21 Apr 2026).

The generation corpus spans T2I, I2I, T2S, TI2S, and interleaved generation tasks. The T2I subset includes long text and extreme aspect ratios up to 1:8. The I2I subset includes image-to-image editing and reference-based generation with up to 9 input images. The T2S and TI2S subsets cover image-series generation with up to 12 output images (Mao et al., 21 Apr 2026).

Retrieval and cleaning are implemented as a closed-loop multi-modal retrieval system: retrieval, annotation, iterative retrieval, and cleaning. The infrastructure supports image-to-image search, multi-image search, text-to-image search, image-text hybrid search, and batch retrieval, with cluster-based diversity reranking to improve coverage while keeping relevance high. Filtering operators are defined across five dimensions: image feature extraction, aesthetic quality, AI-generated content detection, low-level information evaluation, and overall image quality assessment. Example metrics include compression artifact ratio, edge pixel variance, bits per pixel, and artificiality score (Mao et al., 21 Apr 2026).

The annotation pipeline is structured by taxonomy. Images are organized into five top-level categories: photorealistic images, non-photorealistic images, text-centric images, charts, and multi-image compositions. The system supports about 25 fine-grained attributes, including global semantics, human subjects, objects, background, layout, and composition. Structured JSON captions are updated over time to correct missing or inaccurate fields, then rewritten into natural language captions using an MLLM (Mao et al., 21 Apr 2026).

Training is stage-wise. Understanding training proceeds through understanding continual pre-training, multi-task supervised fine-tuning, and on-policy multi-teacher distillation. Generation training has three stages: PT with 13.27T tokens, 713K steps, resolutions $192/320/640$, and T2I:I2I = 7:3; CT with 8.85T tokens, 223K steps, resolutions $512\text{–}2048$ , and T2I:I2I:T2S:TI2S = 7:2:0.5:0.5; and SFT with 13K steps, about 0.62T tokens, high-resolution $512\text{–}2048$ , and more curated data. Optimization uses AdamW, weight decay 0.02, gradient norm clipping 0.5, unconditional dropout 0.05, and learning rates of $H, W$ 0 in PT and CT and $H, W$ 1 in SFT (Mao et al., 21 Apr 2026).

The Prompt Enhancer is a dedicated subsystem for rewriting short or ambiguous user inputs into richer structured prompts. Two versions are trained. The Non-CoT PE is a low-latency variant producing 400–600 output tokens, roughly one second inference, based on Qwen3-VL-2B. The CoT PE is a higher-quality variant producing 1,500–2,000 output tokens, based on Qwen3-VL-30B-A3B, and explicitly reasons before rewriting. Both support T2I, I2I, T2S, and TI2S (Mao et al., 21 Apr 2026).

For reference-guided tasks, the Prompt Enhancer outputs explicit common and difference fields to separate what should be preserved from what should be edited. This design is tightly coupled to Wan-Image’s editing and identity-preservation objectives, since it gives the Planner a formal mechanism for decomposing invariants from editable attributes. The PE is trained with SFT and RL, with RL using GRPO and reward dimensions such as intent fidelity, descriptive richness, and linguistic accuracy (Mao et al., 21 Apr 2026).

The Image Refiner addresses the resolution-pressure problem induced by longer token sequences. Because longer input and output sequences reduce resolution under fixed token budgets, the refiner boosts outputs to 2K–4K and operates in two modes: repair mode, which fixes text errors, artifacts, and local defects, and high-fidelity mode, which preserves identity and structure while improving details. It is trained on degraded images produced by blur, noise, compression, and diffusion-noising pipelines, and is further optimized with distillation and reinforcement learning (Mao et al., 21 Apr 2026).

Post-SFT alignment is performed through Cascade RL, a cascaded domain-wise framework combining DPO, ReFL, DenseGRPO, and ReFMA. The RL dataset is disjoint from SFT prompts to avoid reward leakage, and a reward benchmark is built to prevent overfitting of the reward model. The reported improvement in win score over the SFT baseline ranges from 15% to 50%, depending on scenario (Mao et al., 21 Apr 2026).

6. Capabilities, evaluation, and interpretation

Wan-Image is presented as supporting ultra-long text rendering, hyper-diverse portrait generation, palette-guided generation with explicit hex-code palettes and proportions, multi-subject identity-preserving generation, logical image series generation, multi-modal interactive editing, true alpha-channel generation, and high-efficiency 4K synthesis. It also supports extreme aspect ratio adaptation up to 1:8, and image-series generation can produce coherent sequences of up to 12 images (Mao et al., 21 Apr 2026).

The system’s interleaved-generation behavior is notable because the Planner can produce text and trigger image insertion points with <BOI> and <imagine> tokens, after which the Visualizer outputs the image and generation resumes. This places Wan-Image in partial conceptual alignment with Wan-Weaver’s interleaved framework, although the latter is explicitly benchmarked as an interleaved text-image generator and evaluates long-range textual coherence and visual consistency in that specific setting (Mao et al., 21 Apr 2026, Xing et al., 26 Mar 2026).

On understanding benchmarks, the model is evaluated on MMMU, MMStar, MathVista, HalluBench, MMBench, OCRBench, and AI2D, as well as the text-understanding benchmarks AIME2025, GPQA, HLE, and LCBV6. Ours Instruct averages 72.5 on image understanding and 27.7 on text understanding, while Ours Thinking improves to 76.3 on image understanding and 28.9 on text understanding. The paper states that the thinking mode surpasses the baseline by 5.4 points on image understanding and that Wan-Image outperforms other unified models while being competitive with dedicated understanding-only models (Mao et al., 21 Apr 2026).

Generation evaluation uses a composite human-evaluation setup. For Text-to-Image and General Editing, the win metric is defined as

$H, W$ 2

Interactive Editing and Series Generation use pass rate. Wan-Image is compared against Seedream 5.0 Lite, GPT Image 1.5, and Nano Banana Pro. The paper states that it substantially surpasses Seedream 5.0 Lite and GPT Image 1.5 overall, reaches parity with Nano Banana Pro in challenging tasks, is especially strong in text rendering and photo realism, and achieves about 80% pass rate in interactive editing and image-series generation (Mao et al., 21 Apr 2026).

The practical interpretation advanced by the paper is that Wan-Image is a production-grade visual intelligence system rather than a pure art generator. Its stated application domains are e-commerce, entertainment, education, and personal productivity. A plausible implication is that the system is organized around constraint satisfaction in visual generation: text fidelity, identity preservation, reference adherence, structured editing, and alpha-aware asset production are treated as primary objectives rather than as secondary improvements over conventional aesthetic generation (Mao et al., 21 Apr 2026).