Qwen Image Edit

Qwen Image Edit Workflows

Overview

Qwen Image Edit uses a vision-language model (Qwen2.5-VL) to edit images based on natural language instructions. The model "sees" the source image through CLIP conditioning and generates an edited version.

Models

Required Components

Alternative UNET Models

Conditioning Nodes

TextEncodeQwenImageEditPlusAdvance_lrzjason (Recommended)

From the qweneditutils custom node pack. The Advanced variant is preferred because it:

• Outputs a LATENT directly (no need for separate EmptyLatentImage)
• Has separate VL-resize and non-resize image slots for fine control
• Supports target_size control for output resolution
• Includes a pad/center/disabled crop method with pad_info output

Required Inputs:
  - clip: CLIP
  - prompt: STRING — natural language edit instruction

Optional Inputs:
  - vae: VAE — needed for image encoding and latent output
  - vl_resize_image1-3: IMAGE — images that get VL-resized (downscaled for vision encoder)
  - not_resize_image1-3: IMAGE — images kept at full resolution
  - target_size: [1024, 1344, 1536, 2048, 768, 512] (default 1024)
  - target_vl_size: [392, 384] (default 384)
  - upscale_method: [lanczos, bicubic, area]
  - crop_method: [pad, center, disabled]
  - instruction: STRING — system instruction template (has sensible default)

Outputs (10):
  [0] conditioning_with_full_ref: CONDITIONING — use as positive conditioning
  [1] latent: LATENT — auto-scaled latent, feed directly to KSampler
  [2] target_image1: IMAGE — processed target-size image
  [3] target_image2: IMAGE
  [4] target_image3: IMAGE
  [5] vl_resized_image1: IMAGE — VL-resized version
  [6] vl_resized_image2: IMAGE
  [7] vl_resized_image3: IMAGE
  [8] conditioning_with_first_ref: CONDITIONING — conditioning with only first ref
  [9] pad_info: ANY — padding info for later unpadding

Key advantage: Output [1] (latent) eliminates the need for a separate EmptyLatentImage or VAEEncode node. The Advanced node handles latent creation internally at the correct resolution.

Other Conditioning Variants

• TextEncodeQwenImageEditPlus (Phr00t v2, built-in) is simpler: 4 image inputs, outputs only CONDITIONING. Requires separate EmptyLatentImage. Good for quick edits.
• TextEncodeQwenImageEditPlus_lrzjason: 5 image inputs, resize toggles, but less control than Advance
• TextEncodeQwenImageEditPlusPro_lrzjason: Per-image VL resize selection via vl_resize_indexs string, main_image_index control

Lightning LoRAs (Fast Generation)

4-Step Lightning (2511 Edit)

{
  "class_type": "LoraLoaderModelOnly",
  "inputs": {
    "model": ["<unet_node>", 0],
    "lora_name": "Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors",
    "strength_model": 1.0
  }
}

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

4-Step Lightning (General Qwen)

For non-edit models (txt2img, 2512):

• Qwen-Image-Lightning-4steps-V1.0.safetensors (strength 1.0)

8-Step Lightning

• Qwen-Image-Lightning-8steps-V1.0.safetensors, higher detail than 4-step

Sampler Settings

The sub-1.0 denoise rows REQUIRE a VAEEncode latent. A denoise low enough to shorten the sampling schedule — which 0.5-0.8 certainly is — keeps part of the incoming latent, so that latent has to BE the source image. Wire latent_image from a VAEEncode of the source (or from a node that emits a source-derived latent, like TextEncodeQwenImageEditPlusAdvance_lrzjason output [1]). Pairing these rows with an EmptyLatentImage runs clean and returns a flat, near-uniform field — an empty latent has no source content to preserve. Feeding the reference through TextEncodeQwenImageEditPlus does not rescue it: that image rides on CONDITIONING, which steers denoising but never seeds the sampler's starting state.

Denoise for editing: Lower denoise = closer to source — provided the latent IS the source. 0.5-0.8 range for standard editing on a VAEEncode latent. Lightning uses 1.0 (model handles fidelity internally).

Resolutions

This table is for Qwen-Image TEXT-TO-IMAGE. Do not pick an edit graph's output size from it. An edit graph's geometry is decided by the SOURCE image, not by you — see "Resolution on an edit graph" below. Choosing 1104x1472 here for an edit was #2681.

Qwen-Image operates at ~1.6 megapixels natively:

For video pipelines: Use 832x480 to match WAN 2.2's default resolution.

Resolution on an edit graph

TextEncodeQwenImageEdit and TextEncodeQwenImageEditPlus do not take a size. They scale every reference image to a hard-coded int(1024 * 1024) px — ~1.05 MP, at the source's own aspect ratio — VAE-encode it, and hand it to the model as a reference latent (comfy_extras/nodes_qwen.py).

The model then lays the reference tokens and the tokens it is generating on one shared, centred coordinate grid (comfy/ldm/qwen_image/model.py, process_img), so reference position (i, j) and output position (i, j) mean the same place only when the two grids are the same size. That is the whole reason the official templates run the one image through FluxKontextImageScale and then feed the sampler a VAEEncode of that scaled image — both branches then see the same pixels at the same ~1 MP scale and the same aspect. Every PREFERRED_KONTEXT_RESOLUTIONS entry is ~1.05 MP for the same reason.

It is agreement to within the encoder's round-to-8, not exact equality, and the difference is worth knowing precisely. FluxKontextImageScale snaps to a preferred pair; the encoder then renormalises that to its own 1,048,576 px budget. For 12 of the 18 preferred pairs the two land on the same latent grid. For the other 6 — 688x1504, 800x1328, 832x1248 and their landscape mirrors — the reference lands one latent row or column off: at 800x1328 the sampler's grid is 100x166 and the reference's is 99x165. ComfyUI's own bundled 2511 template does exactly this, so a sub-patch offset is evidently fine in practice. The failure this page is about is one of SCALE, not of rounding — an empty latent at 1104x1472 sits 1.24x away linearly, not one row.

So on an edit graph you do not choose a resolution — you inherit one:

• Right: LoadImage -> FluxKontextImageScale -> (TextEncodeQwenImageEditPlus and VAEEncode) -> KSampler latent_image.
• Wrong: an EmptyLatentImage at a size from the table above. Its dimensions are literals; the reference's are computed from the source when the graph runs. At 1104x1472 (1.63 MP) against a 1.05 MP reference the grids are 1.24x apart linearly, the model cannot copy detail across them, and it re-synthesises the subject instead — materials come back looking plastic/CGI and printed detail comes back as a generic shape (#2681). Nothing errors; the image just is not the edit you asked for.

create_workflow (action:"validate") now flags this pairing as edit_reference_empty_latent.

Prompt Patterns

Edit Instructions (Natural Language)

"Change the black cat into a cute girl with a black bodysuit and jeans"
"Make the sky a dramatic sunset with orange and purple clouds"
"Add a red sports car parked in front of the house"
"Remove the person on the left and fill with the background"

Multi-Angle LoRA (qwen-image-edit-2511-multiple-angles-lora)

Uses <sks> token with structured angle/distance prompts:

<sks> front view eye-level shot close-up
<sks> front-right quarter view low-angle shot medium shot
<sks> back view elevated shot wide shot

Template: <sks> {direction} view {angle} shot {distance}

Directions: front, front-right quarter, right side, back-right quarter, back, back-left quarter, left side, front-left quarter Angles: low-angle, eye-level, elevated, high-angle Distances: close-up, medium shot, wide shot

Negative Conditioning

Always use ConditioningZeroOut for negative conditioning with Qwen edit:

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

Complete Workflow: Lightning Edit (Advanced Node)

Uses TextEncodeQwenImageEditPlusAdvance_lrzjason, which outputs the latent directly, so no EmptyLatentImage is needed.

{
  "1": { "class_type": "UNETLoader", "inputs": { "unet_name": "qwen_image_edit_2511_bf16.safetensors", "weight_dtype": "default" }},
  "2": { "class_type": "LoraLoaderModelOnly", "inputs": { "model": ["1", 0], "lora_name": "Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors", "strength_model": 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": "LoadImage", "inputs": { "image": "<source_image.png>" }},
  "6": { "class_type": "TextEncodeQwenImageEditPlusAdvance_lrzjason", "inputs": {
    "clip": ["3", 0], "prompt": "<edit instruction>", "vae": ["4", 0],
    "vl_resize_image1": ["5", 0],
    "target_size": 1024, "target_vl_size": 384,
    "upscale_method": "lanczos", "crop_method": "pad"
  }},
  "7": { "class_type": "ConditioningZeroOut", "inputs": { "conditioning": ["6", 0] }},
  "8": { "class_type": "KSampler", "inputs": {
    "model": ["2", 0],
    "positive": ["6", 0],
    "negative": ["7", 0],
    "latent_image": ["6", 1],
    "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_edit" }}
}

Key connections:

• "latent_image": ["6", 1]: KSampler gets its latent directly from the Advanced node's output [1]
• "positive": ["6", 0]: conditioning_with_full_ref from output [0]
• "vl_resize_image1": ["5", 0]: source image goes into VL-resize slot (downscaled for vision encoder)