Qwen-Image

'The original 20B Qwen-Image text-to-image model: a 60-block double-stream MMDiT conditioned on Qwen2.5-VL-7B, working in the 8× / 16-channel latent space of a Wan-style VAE. Strong at text rendering, especially Chinese.'

weights: Qwen/Qwen-Image ↗
org: Qwen (Alibaba)
modality: image
tasks: text-to-image
license: Apache 2.0
released: 2025-08-04
native output: 1328×1328 (1:1), 1664×928 (16:9) and other ~1.76 MP aspect ratios
total params: 28.85B
model.arch: qwen_image

Components

role	model	params	size	dtype	trained
Transformer	Qwen-Image MMDiT (20B) Qwen/Qwen-Image/transformer diffusers.QwenImageTransformer2DModel With the default qtype (qfloat8) and name_or_path Qwen/Qwen-Image, ai-toolkit loads the ComfyUI file Comfy-Org/Qwen-Image_ComfyUI split_files/diffusion_models/qwen_image_fp8mixed.safetensors (20,533,672,381 bytes) instead, using this repo only for the config. Unquantized or other qtypes prefer qwen_image_bf16.safetensors from the same repo.	20.43B 20,430,401,088	40.86 GB	bf16	yes
Text encoder	Qwen2.5-VL-7B-Instruct Qwen/Qwen-Image/text_encoder transformers.Qwen2_5_VLForConditionalGeneration Count includes the vision tower. ai-toolkit drops the vision tower right after loading (before quantization), since text-to-image never uses it.	8.29B 8,292,166,656	16.58 GB	bf16	no
Tokenizer	Qwen2 tokenizer Qwen/Qwen-Image/tokenizer transformers.Qwen2Tokenizer	—	—	—	—
VAE	Qwen-Image VAE Qwen/Qwen-Image/vae diffusers.AutoencoderKLQwenImage A Wan 2.1-style causal video VAE. Images go through it as a single frame. Shared by every Qwen-Image 1.x model.	126.9M 126,892,531	254 MB	bf16	no
Training adapter	Accuracy recovery adapter, 3-bitadapter ostris/accuracy_recovery_adapters/qwen_image_torchao_uint3.safetensors Used with qtype "uint3\|ostris/accuracy_recovery_adapters/qwen_image_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. The example config says 3-bit is required for 24 GB.	148.0M 147,961,856	296 MB	fp16	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: The VAE downsamples 3 times (dim_mult [1, 2, 4, 4]) for 8× spatial. The transformer packs 2×2 latent patches into 64-channel tokens, so one token covers 16×16 pixels. Latents are normalized with the per-channel latents_mean / latents_std from the VAE config.

input	latent (c×t×h×w)	tokens
1024×1024	16×1×128×128	4,096	ai-toolkit sample default
1328×1328	16×1×166×166	6,889	native 1:1
1664×928	16×1×116×208	6,032	native 16:9

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). System-prompt template tokens (34) are dropped; up to 1024 tokens
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
UI label: Qwen-Image (image)
model.name_or_path: Qwen/Qwen-Image
source: 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: 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

Resolution: Buckets and sample sizes snap to multiples of 32 (8× VAE × 2×2 patch).
Transformer source: For Qwen/Qwen-Image the transformer weights come from Comfy-Org/Qwen-Image_ComfyUI: fp8mixed when qtype is qfloat8, bf16 when unquantized or for other qtypes. A local copy in the ComfyUI models folder wins over a download. Set model_kwargs.use_comfy_weights: false to load the Diffusers repo instead.
Text encoder: Qwen2.5-VL from text_encoder/ in name_or_path (or extras_name_or_path), vision tower removed. 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 loads with the Qwen/Qwen-Image transformer config, and the text encoder, tokenizer and VAE come from Qwen/Qwen-Image.
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. Cannot be combined with assistant_lora_path.
Loss: Flow-matching velocity target: noise − latents.
LoRA target: QwenImageTransformer2DModel
Saving: LoRA keys use the diffusion_model. prefix, which ComfyUI loads. Full fine-tunes save a single ComfyUI-format file (the Diffusers key layout is the ComfyUI layout for this model).
Sampling: Flowmatch Euler with true CFG (guidance_scale is true_cfg_scale). low_vram also turns on VAE tiling for the decode. Control images are not supported.
Metadata base version: qwen_image

Example config

not verified

job: extensionconfig:  name: "my_qwen_image_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/images"          caption_ext: "txt"          caption_dropout_rate: 0.05          cache_latents_to_disk: true          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"        arch: "qwen_image"        quantize: true        qtype: "qfloat8"        # 24 GB: qtype: "uint3|ostris/accuracy_recovery_adapters/qwen_image_torchao_uint3.safetensors"        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        prompts:          - "a man holding a sign that says, 'this is a sign'"