Updated editing model with native LoRA integration, geometric reasoning for industrial design, multi-person consistency with relational lighting, and auxiliary construction line generation.
A workable 20B-parameter dense image generator from Alibaba. Treat the modality benchmarks above as the leading indicator of fit — composite scoring across modalities is still maturing.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
No benchmark data available for this model yet.
See which devices can run this model and at what quality level.
Alibaba's Qwen-Image-Edit-2511 is a 20-billion parameter dense image editing model built on the MMDiT (Multi-Modal Diffusion Transformer) architecture. This is the second iteration of their dedicated editing specialist, following September 2024's Qwen-Image-Edit-2509, and it addresses the core failure mode of most image editing models: consistency degradation across edits.
The model operates as a text-conditioned image-to-image pipeline. You provide one or more input images plus a text instruction, and the model generates an edited output that preserves structural elements, identity, and spatial relationships. It uses the diffusers library and requires the QwenImageEditPlusPipeline to run.
What distinguishes this release from its predecessor is the explicit targeting of production-grade editing workflows. The key improvements center on four areas: character identity preservation across multiple edits, multi-person group consistency, native integration of community LoRAs directly into the base weights, and geometric reasoning for industrial design tasks. The model also introduces auxiliary construction line generation, which is relevant for design and engineering workflows.
Licensed under Apache 2.0, it competes with other open-weights editing models at similar parameter counts, though the 20B dense architecture puts it above most consumer-grade editing models in terms of raw capacity.
Qwen-Image-Edit-2511 uses a dense 20B parameter MMDiT architecture. Unlike Mixture-of-Experts (MoE) models where only a subset of parameters activate per forward pass, dense architectures use all 20B parameters for every inference step. This has direct implications for hardware requirements: you need enough VRAM to hold the full model weights, not just a fraction of them.
The model processes images through a diffusion pipeline with classifier-free guidance (CFG) support. The inference configuration from the official code uses 40 steps with a true_cfg_scale of 4.0 and guidance_scale of 1.0, which is worth noting for performance tuning. The true_cfg_scale parameter controls how strongly the model adheres to the editing instruction versus the input image structure.
The architecture supports multi-image input, enabling tasks like merging two separate portraits into a coherent group photo. This is not a simple image compositing operation — the model fuses identity features from both inputs while generating a new coherent scene with appropriate lighting and spatial relationships.
Context length is not specified, which is typical for diffusion models that operate on latent image representations rather than token sequences. The primary memory constraint is the image resolution and the number of diffusion steps, not a token limit.
Character consistency editing. This is the headline capability. The model can take a portrait and apply imaginative edits — changing clothing, altering background, modifying style — while preserving facial identity and visual characteristics. In practice, this means you can iterate on a subject through multiple edits without the face drifting into someone else.
Multi-person group editing. The model accepts two separate input images and can merge them into a single coherent scene. This goes beyond simple compositing: it handles relational lighting, spatial positioning, and interaction between subjects. For example, you can input two individual portraits and generate a group photo where both subjects are naturally lit as if they were in the same environment.
Industrial design with geometric reasoning. The model can generate auxiliary construction lines and maintain geometric consistency in product design tasks. This is relevant for engineers and designers working on product visualization, where maintaining precise spatial relationships and perspective is critical.
Built-in LoRA integration. Selected community-developed LoRAs have been integrated directly into the base model weights. This means effects like lighting enhancement, specific artistic styles, or material rendering are available without downloading separate adapter files or performing additional tuning. This reduces friction in production pipelines where you might otherwise need to manage multiple LoRA checkpoints.
Text rendering. The model supports complex text rendering in both Chinese and English, which is relevant for applications like advertisement generation, poster design, or any workflow requiring legible text within the edited image.
This is a 20B dense model. You need to be realistic about hardware.
Minimum VRAM requirements by quantization:
Recommended quantization for most users: Q4_K_M. The quality loss at 4-bit is minimal for image generation tasks, and the VRAM savings are substantial. If you have a 24GB card and want maximum quality, INT8 is viable but expect slower inference.
Expected performance:
Quickest way to get started: Use the diffusers pipeline as shown in the official code. Install from the GitHub source (pip install git+https://github.com/huggingface/diffusers) to ensure you have the QwenImageEditPlusPipeline class. The model is available on Hugging Face at Qwen/Qwen-Image-Edit-2511.
vs. FLUX.1-dev (12B, rectified flow): FLUX is smaller and faster, but it is a general-purpose text-to-image model, not a dedicated editing specialist. Qwen-Image-Edit-2511 has explicit architectural support for editing tasks — multi-image input, consistency preservation, instruction-guided edits — that FLUX cannot match without external tooling. If you need general generation, FLUX is more practical. If your workload is primarily editing, Qwen is the better choice.
vs. SDXL (2.6B, latent diffusion): SDXL is significantly smaller and runs on much less hardware (8GB VRAM is sufficient). However, it lacks the capacity for high-fidelity identity preservation and multi-person consistency that Qwen's 20B parameters enable. For simple edits on a single subject, SDXL with ControlNet can be sufficient. For production editing requiring consistency across multiple iterations and subjects, Qwen justifies the hardware cost.
Tradeoff to consider: Qwen-Image-Edit-2511 requires substantially more VRAM than any consumer-friendly alternative. If you are running on a 12GB or 16GB card, Q4_K_M quantization is your only option, and even then you may need to reduce image resolution. The model's strength is quality and consistency, not speed or accessibility. Choose it when edit fidelity matters more than inference cost.
View on Hugging Face