Qwen-Image-Edit-2509

'The September 2025 update of Qwen-Image-Edit, and the first to take several reference images at once (best with 1 to 3). Same 20B MMDiT and Qwen2.5-VL-7B encoder, with better identity, product and text consistency than the original edit model.'

weights: Qwen/Qwen-Image-Edit-2509 ↗
org: Qwen (Alibaba)
modality: image
tasks: image-editing · multi-image-editing
license: Apache 2.0
released: 2025-09-22not verified
native output: About 1 MP at the last reference image aspect ratio
total params: 28.85B
model.arch: qwen_image_edit_plus

Components

role	model	params	size	dtype	trained
Transformer	Qwen-Image-Edit-2509 MMDiT (20B) Qwen/Qwen-Image-Edit-2509/transformer diffusers.QwenImageTransformer2DModel Same shape as the Qwen-Image transformer. ai-toolkit loads it from this repo.	20.43B 20,430,401,088	40.86 GB	bf16	yes
Text encoder	Qwen2.5-VL-7B-Instruct Qwen/Qwen-Image-Edit-2509/text_encoder transformers.Qwen2_5_VLForConditionalGeneration Includes the vision tower, which the edit models keep: every reference image is shown to it.	8.29B 8,292,166,656	16.58 GB	bf16	no
Tokenizer	Qwen2 tokenizer Qwen/Qwen-Image-Edit-2509/tokenizer transformers.Qwen2Tokenizer	—	—	—	—
Processor	Qwen2-VL image processor Qwen/Qwen-Image-Edit-2509/processor transformers.Qwen2VLProcessor	—	—	—	—
VAE	Qwen-Image VAE Qwen/Qwen-Image-Edit-2509/vae diffusers.AutoencoderKLQwenImage The same Wan 2.1-style VAE as Qwen-Image. Images go through it as a single frame.	126.9M 126,892,531	254 MB	bf16	no
Training adapter	Accuracy recovery adapter, 3-bitadapter ostris/accuracy_recovery_adapters/qwen_image_edit_2509_torchao_uint3.safetensors Used with qtype "uint3\|ostris/accuracy_recovery_adapters/qwen_image_edit_2509_torchao_uint3.safetensors" (the UI option "3 bit with ARA", and the ai-toolkit 32 GB example config).	148.0M 147,961,856	296 MB	bf16	no
total		28.85B	57.70 GB

Latent space

spatial

8×

channels

patch

2×2

autoencoder: Qwen-Image VAE
pixels per token: 16×16
notes: 8× spatial, 2×2 patches into 64-channel tokens, so one token covers 16×16 pixels. Each reference is VAE-encoded at about 1 MP by default and appended after the target tokens, so every reference adds roughly another 1 MP worth of tokens.

input	latent (c×t×h×w)	tokens
1024×1024	16×1×128×128	4,096	ai-toolkit sample default (target only)
1328×1328	16×1×166×166	6,889	Qwen-Image native 1:1 (target only)

Architecture

Blocks: 60 double-stream (MMDiT) blocks
Hidden size: 3072 (24 heads × 128)
FFN size: 12288 (GELU-tanh), separate image and text FFNs
Text conditioning: Joint attention with Qwen2.5-VL last hidden states (3584-d). Each reference is inserted as "Picture N:" vision tokens before the prompt; template tokens (64) are dropped
Image conditioning: Each reference latent appended to the image tokens as its own RoPE frame (1, 2, 3, ...)
Guidance: No guidance embedding. True CFG with a negative prompt
Objective: Rectified flow, resolution-dependent exponential shift (0.5 to 0.9)
Norm / position: QK RMSNorm, 3D RoPE (16/56/56 axes)

In AI Toolkit

model.arch: qwen_image_edit_plus
UI label: Qwen-Image-Edit-2509 (instruction)
model.name_or_path: Qwen/Qwen-Image-Edit-2509
source: extensions_built_in/diffusion_models/qwen_image/qwen_image_edit_plus.py
extensions_built_in/diffusion_models/qwen_image/qwen_image_pipelines.py
extensions_built_in/diffusion_models/qwen_image/qwen_image.py
toolkit/models/v2/diffusion_models/qwen_image.py
toolkit/models/v2/text_encoders/qwen25_vl.py
toolkit/models/v2/vae/qwen_image.py
extensions_built_in/diffusion_models/ui.tsx
extra UI sections: datasets.multi_control_paths, sample.multi_ctrl_imgs, model.low_vram, model.layer_offloading, model.qie.match_target_res

UI defaults

quantize / quantize_te: true / true
qtype: qfloat8
low_vram: true
train.unload_text_encoder: false (section hidden)
noise_scheduler / sampler: flowmatch / flowmatch
timestep_type: weighted
model_kwargs.match_target_res: false
network.conv: disabled (linear LoRA only)
ARA option: 3 bit with ARA (uint3)

Specifics

Training data: Give up to three reference folders as control_path_1, control_path_2, control_path_3 (or a control_path list). Each target image is paired with the same-named file in every folder, in folder order; that order is Picture 1, 2, 3 in the prompt. Captions are the edit instruction.
References into the text encoder: References keep their own size and aspect ratio in the dataloader. For Qwen2.5-VL each one is resized to a 384×384 pixel area (own aspect, snapped to 32 px). Each sample is encoded separately and the embeddings are right-padded to batch them. With cache_text_embeddings the control paths are part of the cache key.
References into the transformer: Each reference is resized to a 1024×1024 pixel area (own aspect, snapped to 32 px), VAE-encoded, packed and appended after the noisy target tokens. With model_kwargs.match_target_res: true they are sized to the target bucket area instead. References stay clean; only the target tokens go into the loss.
Sampling: Set sample.ctrl_img_1 to ctrl_img_3 (or ctrl_img) per prompt. The pipeline is a custom QwenImageEditPlus pipeline with do_cfg_norm off by default: ai-toolkit found the official CFG renormalization hurts more often than it helps.
Text encoder: Qwen2.5-VL with its vision tower, plus the processor. Never trained. The UI forces unload_text_encoder off. The model refuses a prompt without references, so static prompts (blank, trigger, unconditional) are encoded with a random-noise 224×224 reference instead.
Resolution: Buckets and sample sizes snap to multiples of 32 (8× VAE × 2×2 patch).
Single-file checkpoints: A .safetensors name_or_path falls back to Qwen/Qwen-Image for the transformer config, text encoder and VAE, and that repo has no processor. Set extras_name_or_path to the Edit repo.not verified
Loss: Flow-matching velocity target: noise − latents, on the target tokens only.
LoRA target: QwenImageTransformer2DModel
Saving: LoRA keys use the diffusion_model. prefix, which ComfyUI loads. Full fine-tunes save a single ComfyUI-format file.
Metadata base version: qwen_image (the edit class does not override it)

Example config

not verified

job: extensionconfig:  name: "my_qwen_image_edit_2509_lora_v1"  process:    - type: "sd_trainer"      training_folder: "output"      device: cuda:0      network:        type: "lora"        linear: 16        linear_alpha: 16      save:        dtype: float16        save_every: 250        max_step_saves_to_keep: 4      datasets:        - folder_path: "/path/to/target/images"          control_path_1: "/path/to/reference/images_1"          control_path_2: "/path/to/reference/images_2"          caption_ext: "txt"          caption_dropout_rate: 0.05          resolution: [512, 768, 1024]      train:        batch_size: 1        steps: 3000        gradient_checkpointing: true        noise_scheduler: "flowmatch"        timestep_type: "weighted"        optimizer: "adamw8bit"        lr: 1e-4        dtype: bf16        cache_text_embeddings: true      model:        name_or_path: "Qwen/Qwen-Image-Edit-2509"        arch: "qwen_image_edit_plus"        quantize: true        qtype: "qfloat8"        # 32 GB: qtype: "uint3|ostris/accuracy_recovery_adapters/qwen_image_edit_2509_torchao_uint3.safetensors"        quantize_te: true        qtype_te: "qfloat8"        low_vram: true        model_kwargs:          match_target_res: false      sample:        sampler: "flowmatch"        sample_every: 250        width: 1024        height: 1024        guidance_scale: 3        sample_steps: 25        samples:          - prompt: "the person in Picture 1 wearing the jacket from Picture 2"            ctrl_img_1: "/path/to/image1.png"            ctrl_img_2: "/path/to/image2.png"