Qwen-Image-Edit

'The first image-editing version of the 20B Qwen-Image model. One reference image goes to both Qwen2.5-VL (for meaning) and the VAE (for appearance), so it handles both semantic edits like restyling and precise local edits, including editing text in the image.'

weights: Qwen/Qwen-Image-Edit ↗
org: Qwen (Alibaba)
modality: image
tasks: image-editing
license: Apache 2.0
released: 2025-08-18not verified
native output: About 1 MP at the input image aspect ratio
total params: 28.85B
model.arch: qwen_image_edit

Components

role	model	params	size	dtype	trained
Transformer	Qwen-Image-Edit MMDiT (20B) Qwen/Qwen-Image-Edit/transformer diffusers.QwenImageTransformer2DModel Same shape as the Qwen-Image transformer, fine-tuned for editing. ai-toolkit loads it from this repo (no ComfyUI file substitution for the edit models).	20.43B 20,430,401,088	40.86 GB	bf16	yes
Text encoder	Qwen2.5-VL-7B-Instruct Qwen/Qwen-Image-Edit/text_encoder transformers.Qwen2_5_VLForConditionalGeneration Includes the vision tower, which the edit models keep: the reference image is shown to it.	8.29B 8,292,166,656	16.58 GB	bf16	no
Tokenizer	Qwen2 tokenizer Qwen/Qwen-Image-Edit/tokenizer transformers.Qwen2Tokenizer	—	—	—	—
Processor	Qwen2-VL image processor Qwen/Qwen-Image-Edit/processor transformers.Qwen2VLProcessor	—	—	—	—
VAE	Qwen-Image VAE Qwen/Qwen-Image-Edit/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_torchao_uint3.safetensors Used with qtype "uint3\|ostris/accuracy_recovery_adapters/qwen_image_edit_torchao_uint3.safetensors" (the UI option "3 bit with ARA"). The transformer is quantized to 3-bit torchao and this LoRA-shaped adapter runs alongside it to recover accuracy.	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. The reference image is encoded the same way and its tokens are appended after the target tokens, so the sequence is about twice the target token count.

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) of the prompt plus the reference image. Template tokens (64) are dropped
Image conditioning: Reference latents appended to the image tokens as a second frame in RoPE; the same image is also seen by Qwen2.5-VL
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
UI label: Qwen-Image-Edit (instruction)
model.name_or_path: Qwen/Qwen-Image-Edit
source: extensions_built_in/diffusion_models/qwen_image/qwen_image_edit.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.control_path, sample.ctrl_img, model.low_vram, model.layer_offloading

UI defaults

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

Specifics

Training data: Set datasets.control_path to a folder of reference images. Each target image is paired with the file of the same name (any of .jpg, .jpeg, .png, .webp) in that folder. Captions are the edit instruction.
Reference into the text encoder: The reference is resized to about 1 MP (1024×1024 area, its own aspect ratio, snapped to 32 px) and encoded with the prompt by Qwen2.5-VL. With cache_text_embeddings the control path is part of the cache key.
Reference into the transformer: The reference is resized to the target bucket size, VAE-encoded, packed and appended after the noisy target tokens. It stays clean, and only the target tokens are kept from the prediction for the loss.
Sampling: Set sample.ctrl_img per prompt; it is resized to the sample width and height. True CFG (guidance_scale is true_cfg_scale); low_vram turns on VAE tiling.
Resolution: Buckets and sample sizes snap to multiples of 32 (8× VAE × 2×2 patch).
Text encoder: Qwen2.5-VL with its vision tower kept, plus the processor from processor/. Never trained. With low_vram it stays on the CPU and moves to the GPU only to encode.
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 Qwen/Qwen-Image-Edit.not verified
Accuracy recovery adapter: qtype "uint3|<adapter>" quantizes the adapter-covered linears to that qtype and everything else to uint8, then keeps the adapter live on the model.
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_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: "/path/to/reference/images"          caption_ext: "txt"          caption_dropout_rate: 0.05          resolution: [512, 768, 1024]      train:        batch_size: 1        steps: 2000        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"        arch: "qwen_image_edit"        quantize: true        qtype: "qfloat8"        quantize_te: true        qtype_te: "qfloat8"        low_vram: true      sample:        sampler: "flowmatch"        sample_every: 250        width: 1024        height: 1024        guidance_scale: 3        sample_steps: 25        samples:          - prompt: "make the sky a sunset"            ctrl_img: "/path/to/reference.png"