Finally! New best local AI image editor is here

What the video covers (Qwen Image 2.1)

• New open-source image generator/editor from Alibaba: Qwen Image 2.1, positioned as a flexible, powerful local tool.

Capabilities

• Text-to-image with strong realism and anatomy/prompt understanding.
• Generates complex visuals like charts/infographics with many elements.
• Interface/design generation examples are shown.
• Transparent image generation: can output alpha channel / transparency.
• Image-to-image editing via natural language (i.e., edits controlled by prompts), unlike some earlier tools that were mainly text-to-image.
• Reference-image editing
  • Accepts up to 10 reference images.
  • Can maintain face/character consistency when composing from multiple images.
  • Can extract and recombine specific parts described in plain language (e.g., jacket from one photo, shoes from another, hat from another).
  • Can compose items into a new scene (e.g., furniture into a room).
• Direct annotation editing
  • Sketch/mark objects to remove or change.
  • Draw on top of an image to add new content while preserving character integrity.
  • Example: generating a garment outfit while keeping garment design intact.
• 3D/panorama workflow example
  • Uses a selfie/photo to generate a panorama, then allows 3D exploration.

---

Local installation + tutorial (ComfyUI workflow guide)

• Recommended UI for local offline use: ComfyUI (free, customizable).

Setup steps

1. Update ComfyUI via update comfyui.bat.
2. In ComfyUI, go to Templates → find Qwen Image 2.1 workflows:
   • Text-based generation
   • Image editing
   • Background removal
3. If workflows show missing models (red frame), download required models from the referenced “Files and versions” pages and place them into the correct ComfyUI folders:

Model components and options

• Diffusion model (UNet)
  • Options mentioned:
    • BF16 full (large; 14.2 GB cited)
    • INT8 / quantized (smaller; ~7.2 GB cited)
  • Notes:
    • Fits on 8GB VRAM or less; mentions a later approach for 4GB VRAM.
• Text encoders
  • An additional “query refinement” encoder exists but can be very large (~9.5 GB each). The video suggests skipping for low VRAM and improving prompts instead.
  • Smaller/compressed options are offered (e.g., W4A8 ~6.3 GB, recommended for low VRAM).
• VAE
  • Relatively small (~676 MB cited).

---

Workflow details and important settings

1) Text-to-Image workflow settings

• Aspect ratio
• Megapixels / resolution
• Prompts
  • Positive prompt + negative prompt
• CFG (guidance strength)
  • Higher CFG = more literal prompt following
  • Lower CFG = more creativity/variability
  • Special note: if CFG = 1, negative prompt won’t work; needs slightly higher CFG.
• Steps
  • More steps = higher quality, slower generation
• Seed
  • Same seed + same prompt + same settings → same output
  • Change seed for variations
• Performance mentioned
  • Example generation took ~29 seconds on RTX 5000 Ada (16GB VRAM) in a laptop setup.

2) Image Editing workflow

• Uses reference images (upload up to 10).
• Demo shows:
  • Upload an input image (example: car photo)
  • Use positive/negative prompts + editing settings similar to text-to-image
  • Custom size switch
    • If disabled, keeps input resolution
    • If enabled, uses specified aspect ratio/resolution instead
• Performance mentioned
  • Example editing run took ~81 seconds.
  • Video compares it as slower than Flux Cline (previous best editing model, per the speaker).

3) Background Removal workflow

• Another ComfyUI template that outputs PNG with transparency.
• Uses a prompt like:
  • “remove the background, leave only the woman… output PNG”
• Custom size switch again controls output dimensions.
• Performance mentioned
  • Example run took ~116 seconds.
• After output, background can be visually replaced (e.g., red background) to confirm transparency/alpha works.

---

LoRA (fine-tuning) section

• LoRA concept: custom models to add effects/characters/actions.
• For Qwen Image 2.1: few/no LoRAs available yet since it was released recently.
  • One example LoRA found: “Comfy 2.1 anime consistency”, aimed at improving character stability in anime-style edits.

How to load LoRA in ComfyUI

• Insert a Load LoRA node into the editing workflow
• Select the downloaded LoRA from the dropdown
• Adjust LoRA strength (example changed from 1.0 to ~0.8)

• Notes on stacking multiple LoRAs:

  • Duplicate nodes and connect, or use LoRA stacking nodes.

• Performance mentioned

  • Example LoRA-based generation took ~128 seconds.

---

Running on limited VRAM (GGUF fallback for ~4GB GPUs)

• Even the quantized model may not start on 4GB VRAM.
• Solution: use GGUF-quantized models (more compressed).
• The video recommends downloading a GGUF variant (example: a smallest ~3.19 GB file).

How to use GGUF in ComfyUI

• Install extension module: ComfyUI GGUF by City96 (if needed)
• Use UNET Loader GGUF node instead of the standard diffusion/UNet loader
• Select the GGUF file via device name dropdown

Quality observation

• Even aggressive quantization like Q3 “looks good,” with quality claimed to be comparable to the full model (as stated in the demo).

---

License notes (important)

• Non-commercial restriction by default per official license.
• Community questions addressed:
  • Generated results/images are not considered “licensed materials.”
  • Creator retains rights to images/content, implying potential commercial usage based on the response described.

---

Main speakers / sources

• Primary speaker: The YouTube presenter/reviewer guiding the installation and demos (not named in the subtitles).
• Product/source referenced: Alibaba Qwen Image 2.1 (open-source model).
• Tool referenced: ComfyUI (workflows + node-based setup).
• Sponsor referenced: Luma (AI creative agents and video model “Ray 3.2” mentioned).