Qwen

/

Qwen-Image-2.1

Like 3.54k

Follow

Qwen 114k

Hi, I tested this model (quantized int8) in my 4060 ti (16GB) GPU
Here includes the prompts and results (both image gen and image editing)

Model Details:

• Qwen Image 2.1 INT8 Convrot
• Qwen3 VL 8B INT8
• Qwen Image 2.1 VAE BF16

No CPU or Disk Offloading

Platform: ComfyUI

IMAGE GENERATION

Generation 1

Prompt: " A premium futuristic AI image-generation emblem designed as a Telegram bot profile picture, centered composition, a sleek abstract letter "R" formed from flowing luminous light and tiny digital particles, subtle neural-network geometry and generative-art patterns surrounding it, cinematic sci-fi aesthetic, intelligent and elegant rather than robotic, deep black background, electric cyan and violet highlights, glossy glass-metal details, soft volumetric glow, extremely clean silhouette, high contrast, symmetrical, minimal visual clutter, professional AI startup branding, sophisticated technology logo, sharp details, polished 3D render, dramatic studio lighting, perfectly centered, square composition, designed to remain recognizable as a small Telegram avatar, no text, no watermark"

Result:

Generation 2

Prompt: "A colossal impossible megastructure constructed around a rotating black hole, an entire futuristic civilization living on enormous concentric rings orbiting the event horizon, thousands of illuminated cities connected by transparent skybridges, gigantic spacecraft navigating between the structures, gravitational lensing bending distant galaxies into enormous luminous arcs across the sky, artificial suns illuminating different sections of the megastructure, waterfalls flowing upward into zero-gravity gardens, microscopic flying vehicles visible as trails of light, enormous holographic mathematics and physics equations projected into the atmosphere, scientists observing the phenomenon from a glass observatory suspended above the accretion disk, physically detailed materials, realistic reflections, cinematic volumetric lighting, extreme sense of scale, scientifically inspired but visually surreal, breathtaking photorealism, countless tiny details that reward zooming in, masterpiece-level composition"

Result

IMAGE EDITING

Generation 1

Prompt:

"change the planet to mars"

Result:

Generation 2

Prompt:

"Use Reference Image 1 as the environment and city reference, and Reference Image 2 as the character reference for Ultron.

Place Ultron in the foreground of the city, making him absolutely enormous — hundreds of meters tall, towering over the surrounding skyscrapers. His feet should be at street level while his upper body rises far above the rooftops. Preserve Ultron's recognizable metallic appearance, facial structure, armor design, glowing red eyes, and mechanical details from the character reference.

Make the scale extremely convincing: nearby skyscrapers should only reach around his lower legs and knees, while taller buildings reach toward his waist. Use strong atmospheric perspective so distant buildings become smaller and hazier.

Multiple military attack helicopters are surrounding Ultron and attacking him from different directions. Some helicopters are flying close to his shoulders, some near his torso, and others are much farther away between the skyscrapers. Their searchlights illuminate his armor while missiles and tracer fire streak through the night sky toward him.

Ultron should remain the dominant subject and occupy most of the vertical frame. The city must remain clearly recognizable behind and around him.

Nighttime cinematic atmosphere, dramatic clouds, smoke, sparks, explosions in the distance, glowing windows, red emergency lights, helicopter searchlights, realistic reflections on metallic armor, volumetric fog, atmospheric haze, physically accurate shadows and lighting, realistic scale and perspective.

Ultra-photorealistic high-budget science-fiction movie scene, extremely detailed mechanical surfaces, realistic helicopters, believable city architecture, cinematic wide-angle camera, low viewpoint from street level, dramatic composition.

Do not make Ultron human-sized. Do not place him behind the buildings. He must visibly tower over the entire city skyline."

Result:

It also took 1 min to generate.

Generation 3

Prompt:

Positive Prompt
*Use the first image as the exact cat reference and the second image as the exact collar reference.

Place the Christmas-themed collar from Reference Image 2 naturally around the cat's neck, precisely following the area indicated by the blue pen marking in Reference Image 1.

Preserve the cat's identity, face, fur pattern, fur color, body proportions, pose, lighting, and background exactly as much as possible. The collar should fit naturally around the neck with realistic thickness and curvature.

Remove the blue pen marking completely after using it as a placement guide. Do not preserve, reproduce, or leave any blue ink on the cat.

Make the collar look physically attached to the cat: realistic contact with the fur, slight compression of the fur underneath the collar, natural shadows between the collar and neck, correct perspective, realistic material texture, and believable lighting.

Preserve the exact Christmas design, colors, decorations, bells, and shape of the collar from Reference Image 2. Make the final result look like a real photograph of the cat wearing that collar.

Photorealistic, seamless image editing, natural fur detail, realistic materials, accurate proportions, high detail.*

Negative Prompt
blue pen, blue markings, ink marks, collar floating, collar detached, collar behind neck, oversized collar, undersized collar, distorted collar, duplicated collar, extra collar, deformed cat, altered face, altered fur pattern, changed cat identity, distorted neck, unnatural fur, unrealistic shadows, cartoon, illustration, blurry, low quality, artifacts, extra objects, text, watermark

Result:

Performance

• Image generation: ~20 seconds
• Image editing: ~60 seconds
• Peak VRAM: ~15 GB
• Typical VRAM usage during generation: ~13–14 GB
• GPU: RTX 4060 Ti 16GB
• CPU/Disk offloading: None

Uncensored generation test

I also tried testing how the model behaves with content that is usually blocked by a lot of hosted image generation services.

Gave it an ai-generated adult female character as a reference and asked it to generate a nude version of the same character.
the local pipeline actually followed the prompt and generated the requested nsfw result without refusing or showing any obvious content filter.

so at least in my local setup, qwen-image 2.1 didn't seem to have the same kind of content restrictions that you usually see on hosted image generation services.

obviously this is just what i observed with my setup. results could be different depending on the model version, quantization, workflow or hosted service being used.

Conclusion / my take

after testing qwen-image 2.1 int8 convrot on my rtx 4060 ti 16gb, i'm honestly pretty impressed with it.

the image generation quality is really good. i tried both simple stuff like the ai logo and much more complicated scenes with huge structures, cities, lighting effects and tons of small details. it handled all of them pretty well.

the image editing is probably what impressed me the most though.

the ultron test was especially interesting. i gave it a nighttime city and an ultron reference and asked it to make ultron absolutely gigantic while keeping the scale, perspective, lighting and overall look believable. the result came out way better than i expected.

the cat + collar test was another nice example. it was able to use two different reference images and put the collar onto the correct part of the cat while removing the blue marking i used as a placement guide.

performance is also pretty good considering i'm running everything on a 4060 ti 16gb with no cpu or disk offloading.

image generation took around 20 seconds for me and the more complicated editing generations took around 1 minute.
vram peaked at around 15gb and usually stayed around 13-14gb during generation.

overall i think qwen-image 2.1 is a really capable local image generation and editing model for 16gb gpus. the combination of image quality, reference understanding and editing ability is honestly pretty crazy for a consumer gpu like the 4060 ti.

for a 16gb gpu, this is seriously impressive.