Qwen Txt2img

Qwen Image 2512 Text-to-Image Workflows

Overview

Qwen Image 2512 is the latest (December 2025) text-to-image model from the Qwen family. It uses a vision-language model (Qwen2.5-VL) as the text encoder and generates high-quality images from natural language prompts. Two workflow approaches:

1. QwenImageIntegratedKSampler: All-in-one node (recommended for simplicity)
2. Separate component loading: UNETLoader + CLIPLoader + VAELoader + standard KSampler (more flexible)

Models

Standard Components

Fine-tuned Variants (Installed)

Lightning LoRAs

4-Step Lightning (General Qwen / txt2img)

{
  "class_type": "LoraLoaderModelOnly",
  "inputs": {
    "model": ["<unet_node>", 0],
    "lora_name": "Qwen-Image-Lightning-4steps-V1.0.safetensors",
    "strength_model": 1.0
  }
}

Settings: steps=4, cfg=1.0, sampler=euler, scheduler=simple, denoise=1.0

8-Step Lightning (Higher Quality)

{
  "class_type": "LoraLoaderModelOnly",
  "inputs": {
    "model": ["<unet_node>", 0],
    "lora_name": "Qwen-Image-Lightning-8steps-V1.0.safetensors",
    "strength_model": 1.0
  }
}

Settings: steps=8, cfg=1.0 (or 2.5 for character detail), sampler=euler, scheduler=simple

Sampler Settings

ModelSamplingAuraFlow

For standard (non-lightning) presets, apply flow matching shift:

{
  "class_type": "ModelSamplingAuraFlow",
  "inputs": { "model": ["<unet_or_lora>", 0], "shift": 3.1 }
}

Shift=3.1 is the standard value for Qwen Image. Not needed with lightning LoRA (baked into the distillation).

Resolutions

Qwen operates at ~1.6 megapixels natively:

Approach 1: QwenImageIntegratedKSampler (All-in-One)

The QwenImageIntegratedKSampler custom node handles model patching, conditioning, sampling, and output in a single node. Simplest workflow: 4 nodes for model loading + 1 integrated sampler + 1 save.

Node Inputs

Required:
  - model: MODEL (from UNETLoader)
  - clip: CLIP (from CLIPLoader, type=qwen_image)
  - vae: VAE
  - positive_prompt: STRING
  - negative_prompt: STRING
  - generation_mode: "文生图 text-to-image" or "图生图 image-to-image"
  - batch_size: INT (default 1)
  - width: INT (default 0, step 8)
  - height: INT (default 0, step 8)
  - seed: INT
  - steps: INT (default 4)
  - cfg: FLOAT (default 1)
  - sampler_name: euler, dpmpp_2m, etc.
  - scheduler: simple, sgm_uniform, beta, etc.
  - denoise: FLOAT (default 1)

Optional:
  - image1-5: IMAGE (reference images for i2i or multi-ref)
  - latent: LATENT
  - controlnet_data: CONTROL_NET_DATA
  - auraflow_shift: FLOAT (default 3)
  - cfg_norm_strength: FLOAT (default 1)

Outputs:
  [0] IMAGE — generated image
  [1] LATENT — output latent (optional)
  [2] IMAGE — scaled input image (for i2i)

Complete Workflow: Integrated Sampler (Lightning 4-Step)

{
  "1": { "class_type": "UNETLoader", "inputs": { "unet_name": "qwenImageEditRemix_v10.safetensors", "weight_dtype": "default" }},
  "2": { "class_type": "LoraLoaderModelOnly", "inputs": { "model": ["1", 0], "lora_name": "Qwen-Image-Lightning-4steps-V1.0.safetensors", "strength_model": 1.0 }},
  "3": { "class_type": "CLIPLoader", "inputs": { "clip_name": "qwen_2.5_vl_7b_fp8_scaled.safetensors", "type": "qwen_image" }},
  "4": { "class_type": "VAELoader", "inputs": { "vae_name": "qwen_image_vae.safetensors" }},
  "5": { "class_type": "QwenImageIntegratedKSampler", "inputs": {
    "model": ["2", 0],
    "clip": ["3", 0],
    "vae": ["4", 0],
    "positive_prompt": "<detailed natural language prompt>",
    "negative_prompt": "",
    "generation_mode": "文生图 text-to-image",
    "batch_size": 1,
    "width": 1024,
    "height": 1344,
    "seed": 42,
    "steps": 4,
    "cfg": 1,
    "sampler_name": "euler",
    "scheduler": "simple",
    "denoise": 1,
    "auraflow_shift": 3,
    "cfg_norm_strength": 1
  }},
  "6": { "class_type": "SaveImage", "inputs": { "images": ["5", 0], "filename_prefix": "qwen_t2i" }}
}

Approach 2: Separate Component Loading (Standard Pipeline)

More flexible, since it allows inserting additional processing nodes between stages.

Pipeline Flow

UNETLoader → [LoraLoaderModelOnly] → [ModelSamplingAuraFlow (shift=3.1)] → MODEL
CLIPLoader (qwen_image) → CLIP
VAELoader → VAE

CLIPTextEncode (positive) → CONDITIONING
ConditioningZeroOut → negative CONDITIONING

EmptySD3LatentImage (1024x1344) → LATENT

KSampler → VAEDecode → SaveImage

Latent node: use EmptySD3LatentImage, matching the official Comfy-Org image_qwen_image template. Qwen Image’s latent format is Wan21, so its latent is 16-channel; EmptyLatentImage emits 4. A bare EmptyLatentImage → KSampler still renders, because ComfyUI’s fix_empty_latent_channels (comfy/sample.py, called by every sampler node) repeats an all-zero latent up to the model’s channel count. But that rescue is gated on torch.count_nonzero(latent) == 0, so it stops applying the moment a node inserted here writes into the latent — which is exactly what this approach is for. Start 16-channel and the question never arises.

Complete Workflow: Separate Loading (Lightning 4-Step)

{
  "1": { "class_type": "UNETLoader", "inputs": { "unet_name": "qwenImageEditRemix_v10.safetensors", "weight_dtype": "default" }},
  "2": { "class_type": "LoraLoaderModelOnly", "inputs": { "model": ["1", 0], "lora_name": "Qwen-Image-Lightning-4steps-V1.0.safetensors", "strength_model": 1.0 }},
  "3": { "class_type": "CLIPLoader", "inputs": { "clip_name": "qwen_2.5_vl_7b_fp8_scaled.safetensors", "type": "qwen_image" }},
  "4": { "class_type": "VAELoader", "inputs": { "vae_name": "qwen_image_vae.safetensors" }},
  "5": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["3", 0], "text": "<detailed natural language prompt>" }},
  "6": { "class_type": "ConditioningZeroOut", "inputs": { "conditioning": ["5", 0] }},
  "7": { "class_type": "EmptySD3LatentImage", "inputs": { "width": 1024, "height": 1344, "batch_size": 1 }},
  "8": { "class_type": "KSampler", "inputs": {
    "model": ["2", 0],
    "positive": ["5", 0],
    "negative": ["6", 0],
    "latent_image": ["7", 0],
    "seed": 42, "steps": 4, "cfg": 1, "sampler_name": "euler", "scheduler": "simple", "denoise": 1
  }},
  "9": { "class_type": "VAEDecode", "inputs": { "samples": ["8", 0], "vae": ["4", 0] }},
  "10": { "class_type": "SaveImage", "inputs": { "images": ["9", 0], "filename_prefix": "qwen_t2i" }}
}

Complete Workflow: Standard Quality (50-Step)

{
  "1": { "class_type": "UNETLoader", "inputs": { "unet_name": "qwenImageEditRemix_v10.safetensors", "weight_dtype": "default" }},
  "2": { "class_type": "ModelSamplingAuraFlow", "inputs": { "model": ["1", 0], "shift": 3.1 }},
  "3": { "class_type": "CLIPLoader", "inputs": { "clip_name": "qwen_2.5_vl_7b_fp8_scaled.safetensors", "type": "qwen_image" }},
  "4": { "class_type": "VAELoader", "inputs": { "vae_name": "qwen_image_vae.safetensors" }},
  "5": { "class_type": "CLIPTextEncode", "inputs": { "clip": ["3", 0], "text": "<detailed natural language prompt>" }},
  "6": { "class_type": "ConditioningZeroOut", "inputs": { "conditioning": ["5", 0] }},
  "7": { "class_type": "EmptySD3LatentImage", "inputs": { "width": 1328, "height": 1328, "batch_size": 1 }},
  "8": { "class_type": "KSampler", "inputs": {
    "model": ["2", 0],
    "positive": ["5", 0],
    "negative": ["6", 0],
    "latent_image": ["7", 0],
    "seed": 42, "steps": 50, "cfg": 4, "sampler_name": "euler", "scheduler": "simple", "denoise": 1
  }},
  "9": { "class_type": "VAEDecode", "inputs": { "samples": ["8", 0], "vae": ["4", 0] }},
  "10": { "class_type": "SaveImage", "inputs": { "images": ["9", 0], "filename_prefix": "qwen_t2i_hq" }}
}

Negative Conditioning

Always use ConditioningZeroOut for Qwen txt2img:

{
  "class_type": "ConditioningZeroOut",
  "inputs": { "conditioning": ["<positive_cond>", 0] }
}

Or use an empty string in CLIPTextEncode, but ZeroOut is more explicit and reliable.

QwenImageDiffsynthControlnet

For ControlNet support with Qwen models. Patches the model with a DiffSynth control signal:

Required Inputs:
  - model: MODEL
  - model_patch: MODEL_PATCH (from DiffSynth ControlNet loader)
  - vae: VAE
  - image: IMAGE (control image)
  - strength: FLOAT (default 1.0)

Optional:
  - mask: MASK

Outputs:
  [0] MODEL (patched)

DiffSynth ControlNets support: canny, depth, inpaint only (NOT pose).

Concept/Style LoRAs (Installed)

Located in loras/Qwen/:

• style/: Figure makers, reality transform, panel painter
• concept/: Various concept LoRAs
• poses/: Pose-specific LoRAs
• character/: Character enhancement
• anime/: Anime style LoRAs
• tool/: Utility LoRAs (anything2real, gaussian splash)
• equirectangular projection/: 360 panorama LoRA

Apply with LoraLoaderModelOnly:

{
  "class_type": "LoraLoaderModelOnly",
  "inputs": {
    "model": ["<unet_or_lightning_lora>", 0],
    "lora_name": "Qwen\\concept\\hinaQwenImageAsianMixLora_v2.safetensors",
    "strength_model": 0.8
  }
}

Prompt Style

Natural language, 1 to 3 sentences. Be descriptive:

Good: "Professional portrait of an Asian woman in her late 20s, wearing a cream linen blazer at a Tokyo rooftop café during golden hour, holding a matcha latte, editorial fashion photography, shot on Sony A7III 85mm f/1.4"
Bad: "1girl, cafe, blazer, matcha"

Tips:

• Put text to render in quotes within the prompt
• "photograph" works better than "photorealistic"
• Negative prompts: use NLP-style descriptions, not keyword spam (or use ZeroOut)

VRAM Considerations

• Always clear_vram before switching to Qwen from another model family
• Lightning 4-step takes ~3-5s per image

Tips

1. QwenImageIntegratedKSampler is the simplest approach for basic txt2img. One node handles everything
2. For LoRA stacking or ControlNet, use the separate component pipeline instead
3. The integrated sampler's auraflow_shift defaults to 3 (close to the recommended 3.1). Adjust only if needed
4. For video pipeline output (feeding into WAN FLF), set resolution to 832x480
5. CopaxTimeless pick: res_multistep + sgm_uniform at CFG 4.0 for ultra-realistic results
6. Multiple concept LoRAs can stack. Reduce individual strength to 0.5-0.7 when combining

Sources

• Official: Comfy-Org workflow template image_qwen_image.json — https://github.com/Comfy-Org/workflow_templates/blob/main/templates/image_qwen_image.json
• Empirical: sampler values, wiring, and prompt notes from working graphs in packs/ and observed renders; not a vendor prompting guide.