Wan-Image: Pushing the Boundaries of Generative Visual Intelligence

TL;DR

Wan-Image is Alibaba Tongyi Lab’s technical report for Wan2.7-Image — the system previously pre-announced via Wan2.7-Image — unified model for image generation and editing (Alibaba Wan announcement). It’s a unified visual generation stack with an MLLM Planner + DiT Visualizer + 4-channel RGBA VAE that targets professional image production (ultra-long text rendering, 1:8 aspect ratios, multi-subject identity preservation, palette-guided generation via HEX codes, up-to-12 coherent image series, alpha-channel output, 4K). Human eval places Wan-Image above Seedream 5.0 Lite and GPT Image 1.5 overall and at parity with Nano Banana Pro on challenging tasks like text rendering and photorealism. The architecture is a unified Transformer where understanding and generation experts share attention within each block; the generation branch is rectified-flow DiT, with causal attention between text/ViT/VAE tokens and bidirectional attention inside each image (and across images in a series).

Key claims

Unified Transformer architecture: a single decoder-only stack with dedicated understanding and generation experts that share attention within each block, trained in two decoupled stages — understanding first, then the generation branch initialized from the well-trained understanding attention weights [§3.1].
Generation branch is a DiT under Rectified Flow, with LayerNorm + SiLU + QK-Norm for training stability [§3.1].
Structured attention mask: causal attention among text, ViT, and VAE tokens; bidirectional attention within the visual tokens of each image; bidirectional attention extended across images in a series to enforce identity/style consistency across the up-to-12 outputs [§3.1, Attention and modulation mechanism].
Image editing conditioning: clean VAE tokens of the input image are fed to the DiT as additional conditional inputs, with timestep explicitly set to zero (Teacher-Forcing style) so they act as deterministic conditions while noised tokens carry the denoising signal [§3.1].
3D-RoPE in the generation branch with a fixed positional offset between text and image segments as a semantic boundary; MRoPE in the understanding branch; the two branches share a unified positional reference frame so generation visual tokens can attend over a shared spatio-temporal context [§3.1, Positional encoding].
Novel 4-channel VAE jointly modelling RGB content and alpha transparency, with a hybrid reconstruction loss constraining visible content, alpha structure, and transparent-boundary quality — explicitly designed to avoid the black/white fringes, halos, and background leakage that RGB-only VAEs produce on PNG / e-commerce / game assets [§3.2].
VAE trained in three progressive stages (low-res → mixed-resolution → mixed-resolution + 4-channel GAN discriminator) [§3.2, Training strategy].
Data engine combines a multi-modal retrieval system (image-to-image, multi-image, text-to-image, image–text hybrid, batch retrieval) with cluster-based diversity re-ranking and a 5-dimension operator stack (feature extraction, aesthetic quality, AIGC detection, low-level info-entropy/JPEG-bpp/edge-variance, overall quality including watermark and greasy-texture detection) [§2.2.1].
Structured 5-class image taxonomy (photorealistic / non-photorealistic / text-centric / charts / multi-image compositions) with 25 dynamic dimension-specific annotation attributes — captions come in three forms: raw web metadata, MLLM-generated natural-language at multiple granularities, and structured JSON [§2.2.2].
Stage-wise multi-stage training (PT → CT → SFT) with dynamic resampling: text generation and identity preservation are upweighted as longer-tail tasks; SFT retains only data passing rigorous quality + aesthetic filtering [§2.2.3].
SigLIP-2 semantic embeddings plus low-level metadata are used for parallel feature extraction and same-source deduplication; >20-dim per-sample scoring with stage-wise dynamic thresholding [§2.2.3].
Reinforcement learning data is curated specifically for human aesthetic preferences, identity preservation, and complex instruction following [§1, §2.2.3].
Headline human-eval result: Wan-Image surpasses Seedream 5.0 Lite and GPT Image 1.5 in overall performance and reaches parity with Nano Banana Pro on challenging text-to-image tasks (text rendering, photorealism) [§1, Fig. 1].
Productivity tasks reach ~80% pass rate: interactive editing and image-series generation [§1].

Method

The system has four logical modules: a Prompt Enhancer (LLM that disambiguates user intent into a precise prompt), a Planner (MLLM for semantic reasoning and task routing), a Visualizer (DiT under Rectified Flow), and an Image Refiner (pushes output to 4K). Architecturally these collapse into a unified Transformer where understanding and generation experts share attention within each block.

Training is staged. The understanding branch is trained first as a standard decoder-only Transformer. Then the generation branch is initialized by copying the understanding-branch attention weights and adding new randomly-initialized modules (the DiT-specific layers and the 4-channel VAE pipeline). This decoupled-initialization recipe places Wan-Image in the Decoupled slot of the UniG2U taxonomy from UniG2U-Bench: Do Unified Models Advance Multimodal Understanding?, with the twist that the generation branch inherits the understanding branch’s attention parameters rather than being a separate frozen / from-scratch module.

The attention design is the key architectural lever. Within a single image, visual tokens see each other bidirectionally (standard DiT). Across images in a series, bidirectional attention is also enabled — this is how the model maintains identity and style consistency across up to 12 outputs in one generation call. Between text / ViT / VAE token streams, attention is causal, preserving the autoregressive structure needed for the planning side of the system. Editing conditioning is handled by encoding the input image with the VAE and feeding the clean (timestep-zero) tokens alongside noised tokens to AdaLN-modulated DiT blocks, so structural content is held fixed while edits are denoised.

The 4-channel VAE is the second architectural lever. Standard RGB-only VAEs fail on PNG / RGBA assets because they cannot disentangle background color from transparency — the encoder leaks background into foreground colors near alpha boundaries. Wan-Image’s 4-channel VAE encodes the alpha channel jointly with RGB, trained with a hybrid loss that separately constrains visible-content reconstruction, alpha structure, and transparent-boundary quality. This is what enables native alpha-channel generation — the model can emit a transparent-background PNG directly rather than relying on background-removal post-processing.

Results

The headline evaluation is human preference (Figure 1, comprehensive). Against Seedream 5.0 Lite and GPT Image 1.5, Wan-Image wins overall; against Nano Banana Pro, it reaches parity on the most challenging text-to-image scenarios — text rendering and photorealism specifically — and trails on the overall composite. On the productivity side (interactive editing, image-series generation), pass rate is ~80%, described as sufficient to fulfill real-world workflow requirements.

The paper does not report standard quantitative benchmarks (GenEval, ImgEdit, OneIG, OCR/text-rendering benchmarks like AnyText) in the fetched portion of the report — the comparative claims are human-preference, in-house, against named closed competitors. Note that Seedream 5.0 Lite, GPT Image 1.5, and Nano Banana Pro are all closed-API-only baselines, which is consistent with this being a positioning report against the closed frontier rather than a head-to-head with open peers like HunyuanImage 3.0 (HunyuanImage 3.0 Technical Report) or Z-Image (Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer).

Why it’s interesting

Wan-Image is the technical report behind the Wan2.7-Image — unified model for image generation and editing (Alibaba Wan announcement) announcement that Luma already filed — the prediction in that page (“Worth re-filing once the unrolled thread or a technical report lands so the claims can be checked against benchmarks”) is now resolved. The architecture sits in the Unified Multimodal Models Decoupled slot but with a novel twist: the generation branch’s attention is initialized from the trained understanding branch, which is a middle-ground between fully E2E (BAGEL, UniPic2-Metaquery-9B) and fully decoupled (DuoGen’s frozen-MLLM-plus-DiT recipe in DuoGen: Towards General Purpose Interleaved Multimodal Generation). The 4-channel RGBA VAE is the most concrete architectural-differentiation lever — most filed image-generation work (HunyuanImage 3.0, Z-Image, FLUX.2, Qwen-Image-2.0) uses RGB-only VAEs and ships background-removal as a separate post-process; Wan-Image bakes alpha into the latent space. The “1:8 aspect ratio + 12-image-series consistency + 3K-token text rendering” capability surface is also the most aggressive productized claim filed so far for a single open-image-foundation candidate.

The positioning is professional design / e-commerce / typography rather than general aesthetic generation — a deliberate departure from the Introducing Recraft V4: Design Taste Meets Image Generation and Introducing MAI-Image-2: for limitless creativity aesthetic-first framings, and closer to the productivity-tool framing of Introducing MAI-Image-2: for limitless creativity / Adobe / Photoshop than to research-pure UMM work. Whether weights and license will be released openly is not confirmed in the report fetched at filing — the abstract describes a web service at create.wan.video/generate/image/generate?model=wan2.7-pro, suggesting at least the Pro tier may be API-gated.