Qwen-Image-2.0 Technical Report

Data infrastructure

• A fine-grained captioning framework and multi-stage, multi-resolution data pipeline combine filtered corpora, editing pairs, synthetic data, and curated high-resolution samples.An automated data flywheel uses evaluation signals and user feedback to identify failure modes and guide iterative refinement.Source: §1 · p. 5

Architecture and training

• The model couples a Qwen3-VL encoder with an MMDiT backbone and a 16×-downsampling VAE, then trains through pretraining, fine-tuning, and RLHF stages.The VAE uses residual autoencoding, enlarged latent channels, and semantic alignment loss; training includes a resolution curriculum and task-specific rewards optimized with diffusion RL using GRPO.Source: §1 · p. 5

Text rendering and multilingual typography

• Prompts of up to 1K tokens enable text-dense outputs, while multilingual rendering improves character accuracy and supports complex typography.The report names slides, posters, and infographics as target outputs.Source: §1 · p. 6

High-resolution photorealism

• Native 2K generation produces finer textures, more coherent lighting, and more realistic materials across portraits, natural scenes, and architecture.The model supports photorealistic generation at high resolution.Source: §1 · p. 6

Artistic expression

• Across diverse aesthetic settings, the model maintains robust quality and reduces fluctuations between artistic styles.This addresses quality variation across styles.Source: §1 · p. 6

Instruction following

• The model more precisely follows complex, composition-heavy prompts through stronger semantic understanding.This targets prompts involving multiple entities, spatial constraints, and compositional logic.Source: §1 · pp. 5–6

Unified generation and editing

• A single model supports both text-to-image generation and instruction-based image editing within a unified architecture and training paradigm.The framework is intended to bring multimodal understanding and high-fidelity generation together without separate pipelines or quality trade-offs.Source: §1 · pp. 5–6

Inference efficiency

• Joint optimization of architecture and training strategy makes inference faster while preserving visual quality for interactive creative workflows.The report positions this efficiency improvement for latency-sensitive and resource-limited deployment.Source: §1 · pp. 5–6

2.1 Data Collection

• The collection pipeline combines broad-domain text-image pairs with single- and multi-image editing data, covering diverse content and increasingly complex edit behaviors.Generation data spans realistic photography, graphic design, artistic content, synthetic imagery, and layout-sensitive assets; editing includes attribute changes, text editing, reference-based generation, subject consistency, and compositional merging.Source: §2.1 · pp. 6–7

2.2 Data Annotation

• Four task-specific caption schemes describe general visual content, dense text and layouts, image-related knowledge, and structured entities, attributes, and relations.General captions include scene context, spatial relationships, and text semantics; the other schemes target text-centric materials, supplementary world knowledge, and complex diagrams or graphs.Source: §2.2 · p. 7

2.3 Multi-Stage Training Data Strategy

• Training progresses through six data stages, expanding from filtered 256p generation to joint generation and editing at up to 2048p, then high-resolution supervised fine-tuning.The stages add editing data at 256p, synthetic data at 512p, and filtered 1024p and 2048p samples; final SFT applies stricter distribution filtering to refine high-resolution data quality.Source: §2.3 · pp. 7–8

2.4 Closed-loop Data Flywheel System

• Failure cases are assigned to reinforcement-learning, pre-training data-compensation, or prompt-engineering tracks according to their attributed cause.The pre-training track retrieves and augments prompts or instruction-image pairs with human review, while prompt engineering refines inputs when capability exists but instruction understanding or formulation is inadequate.Source: §2.4 · p. 9
• A data flywheel collects evaluation, mined-failure, and user-feedback signals, routes cases for targeted remediation, and feeds updated models back into evaluation.The loop limits manual intervention to critical data filtering and uses error attribution, retrieval, and iterative updates to support resource-efficient optimization and data reliability.Source: §2.4 · pp. 8–10

3 Architecture

• The architecture couples Qwen3-VL for semantic conditioning, a VAE for image-latent encoding and decoding, and an MMDiT for conditioned latent-space denoising.Qwen3-VL is the implemented MLLM condition encoder (Bai et al., 2025a).Source: §3 · p. 10, Figure 8 · p. 10

3.1 Variational AutoEncoder

• A 16×-compression VAE reduces diffusion training costs, while residual shortcuts and 64 latent channels address reconstruction losses from aggressive compression.The f16c64 design retains the same total channel bottleneck as the f8c16 baseline; text-rich image training targets reconstruction of dense text across English and Chinese.Source: §3.1 · pp. 10–11
• At 16× compression, Qwen-Image-2.0-VAE achieves state-of-the-art reconstruction across all reported metrics on general-domain and text-rich evaluations.The evaluations use PSNR and SSIM on ImageNet-1k validation images at 256×256 and an in-house text-rich corpus. The supplied passages do not provide metric values.Source: Table 1 · p. 11, §3.1 · p. 11
• The VAE uses reconstruction, perceptual, and semantic-alignment losses, with strong early alignment gradually relaxed; adversarial loss is omitted to improve training stability.Semantic alignment follows VA-VAE and aligns latents to representations across domains, aspect ratios, and resolutions.Source: §3.1 · p. 11

3.2 Multi-modal Diffusion Transformer

• The MMDiT jointly processes text and image tokens for T2I and TI2I, including interleaved multi-image inputs, using VAE latents and Qwen3-VL representations.MSRoPE supplies unified positional encoding; the model also uses multiplicative modulation and SwiGLU MLPs to address excessive activations and neuron saturation during joint training.Source: §3.2 · pp. 11–12

3.3 Prompt Enhancer

• The Prompt Enhancer rewrites prompts of varying specificity into structured, detail-rich instructions to convey intended layouts, object relations, visual hierarchy, and composition.For generation, training data pairs degraded colloquial prompts with detailed annotations and inverse-reasoning traces; for editing, an MLLM summarizes annotations into concise prompts.Source: §3.3 · p. 12
• Initialized from Qwen3.5-9B (Team, 2026), the Prompt Enhancer is trained with SFT followed by GRPO reinforcement learning using visual-consistency, aesthetic, and textual-constraint rewards.A frozen image generator evaluates candidate rewrites during RL, aligning prompt enhancement with downstream image quality.Source: §3.3 · p. 12
• Prompt enhancement consistently improves generation quality, prompt following, and reasoning performance compared with using original captions.Figure 9 provides a qualitative comparison of T2I results from original and prompt-enhanced captions.Source: §3.3 · p. 12, Figure 9 · p. 13

4.1 Multistage Training

• Training progresses through 700K-step pre-training, 250K-step continual pre-training, and approximately 10K-step supervised fine-tuning, with changing resolutions, data mixtures, and filtering.Pre-training uses a 9:1 T2I/TI2I mixture at 1 × 10−4 learning rate; continual pre-training uses 7:3 at 2 × 10−5 while increasing resolution to 512–2048; fine-tuning uses 1 × 10−5 with diverse, strictly filtered and manually curated data.Source: §4.1 · p. 14

4.2 Reinforcement Learning with Human Feedback

• Task-specific composite rewards target T2I aesthetics, image-text alignment, and portrait quality, and TI2I instruction following and visual consistency.The criteria include composition, illumination, texture, prompt fidelity, anatomical plausibility, execution of edits, and preservation of unmodified regions.Source: §4.2 · pp. 14, 16
• The RLHF pipeline uses adapted GRPO (Liu et al., 2026; Wang et al., 2025; Zheng et al., 2025), applying CFG in rollout sampling but excluding the unconditional branch from policy optimization.This hybrid strategy supports high-quality reward candidates while reducing optimization overhead; prompt distributions and reward weights are adjusted during training.Source: §4.2 · p. 16
• RLHF yields consistent qualitative gains in T2I and editing, improving texture fidelity, realism, and visual consistency over the base model.Figure 10 compares Qwen-Image-2.0-Base and Qwen-Image-2.0-RL across portraits, landscapes, posters, and natural scenes.Source: Figure 10 · p. 15, §4.2 · p. 16

4.3 Few-step Distillation

• Few-step distillation uses DMD, comparing student and teacher conditional scores on noisy interpolations of student-generated samples.The student prediction may be a final clean sample or an intermediate clean-state prediction; the student score is estimated by an auxiliary fake-score model trained on student samples.Source: §4.3 · pp. 16–17
• 4-NFE Qwen-Image-2.0-Distillation is visually comparable to its 40-step teacher across diverse prompts, preserving appearance, composition, and semantic alignment at lower inference cost.The student is distilled from Qwen-Image-2.0-Base and retains perceptual quality and prompt-following capability.Source: §4.3 · p. 17, Figure 11 · p. 17

5.1 LMArena Benchmark Evaluation

• 1168 ELO places Qwen-Image-2.0 at #9 globally and #1 among Chinese models, where it outperforms Nano Banana.LMArena compares model outputs anonymously for the same prompt and uses ELO rankings to measure user preference.Source: §5.1 · p. 17, Figure 12 · p. 18

5.2 Qualitative Results on Text-to-image Generation

• Qwen-Image-2.0 uniquely combines character-level accuracy, correct spatial binding of text, and coherent composition in the text-rendering examples.The comparisons include Chinese text-rendering examples where other models produce errors, omit or duplicate text, or place text incorrectly.Source: §5.2 · p. 18, Figure 13 · p. 20
• Across two portrait examples, Qwen-Image-2.0 preserves sign text and photorealistic lighting while following scene constraints such as motion blur and decal placement.The comparisons note failures in text, material realism, and scene instructions among the other models.Source: §5.2 · p. 19

5.3 Qualitative Results on Image Editing

• Qwen-Image-2.0 renders complex Chinese poems with accurate characters, canonical line order, and coherent vertical layouts in image-editing examples.The longer example contains a 40-character poem; baseline models reorder couplets, fragment the poem, or render it barely legibly.Source: §5.3 · p. 19, Figure 16 · p. 23
• In single-image and multi-image edits, Qwen-Image-2.0 preserves subject identity and fine details while accurately carrying out object-placement and scene-composition instructions.Examples include preserving a cat’s appearance while adding objects and a hat, and retaining a painter’s facial identity and clothing pattern in a composed scene.Source: §5.3 · p. 19, Figure 17 · p. 24

• Qwen-Image-2.0 supports both text-to-image generation and instruction-based image editing within one framework.Source: §6 · p. 27
• Its multimodal encoder, efficient MMDiT backbone, and high-compression VAE address challenges in long-text rendering, multilingual typography, photorealism, instruction following, and inference efficiency.Source: §6 · p. 27
• The model is presented as a foundation for future research and practical deployment of general-purpose image-generation systems.Source: §6 · p. 27