If you are familiar with earlier Qwen mage model, you have probably seen it yourself. Faces that look slightly plastic. Skin that feels too smooth. Even when the prompt is perfect, that AI look sneaks in and breaks immersion.
And it is not just people but natural scenes like landscapes or animal body detailing often lack the fine, organic details that make an image feel real. Text rendering is another pain point. Misspelled words or text that looks pasted on instead of naturally integrated into the image. The improved version Qwen Image 2512 released by Qwen Team (a part of Alibaba Group) that fixes these problems.
 |
| Qwen Image 2512 Showcase(Ref-Official page) |
Compared to the August release, this improved version focuses heavily on making images feel less AI generated and more real world believable. The goal is not flashy upgrades but it's practical, visible improvements that you notice immediately when you generate images.
The improvements in Qwen Image 2512 clearly show that the team studied where users struggled the most in the earlier version.
(a) Enhanced Human Realism
 |
| Enhanced Human Realism |
Human realism has been reworked with richer facial details, better
age representation, and improved environmental context. Faces now carry
subtle imperfections and expressions that reduce that artificial feel.
(b) Improved Text Rendering
 |
| Improved text rendering |
Text rendering, which was already a strength of Qwen Image, has been pushed further. The model now delivers better accuracy, cleaner layouts, and stronger multimodal integration, meaning text and image work together instead of fighting each other.
(c) Finer Natural Detail
 |
| Fine Natural details |
Natural elements like landscapes, animal fur, and textures are rendered with finer detail, making scenes look more alive and less flat.
Installation
1. New user have to install ComfyUI. Portable users need to Update ComfyUI from Manager by selecting Update All if you are familiar to it. Desktop version will be updated automatically.
2. Download Qwen Image 2512 BF16/FP8 (qwen_image_2512_bf16.safetensors
or qwen_image_2512_fp8_e4m3fn.safetensors). Save it into ComfyUI/models/diffusion_models directory. Higher VRAM user(more than 24GB) can use BF16 that will generate better results. Lesser VRAM users(12-16GB) need to use FP8 variant.
For low VRAM users, download Qwen Image 2512 GGUF by unsloth. Save it into your ComfyUI/models/unet folder.
All the VRAM usage with model size has been provided below:
2-bit- Q2_K 7.22 GB
3-bit-Q3_K_S 9.04 GB, Q3_K_M 9.74 GB
4-bit- Q4_K_S 12.3 GB, Q4_0 11.9 GB, Q4_1 12.8 GB, Q4_K_M 13.1 GB
5-bit- Q5_K_S 14.3 GB, Q5_0 14.4 GB, Q5_1 15.4 GB, Q5_K_M 15 GB
6-bit-Q6_K 16.8 GB,
8-bit-Q8_0 21.8 GB
16-bit-BF16 40.9 GB, F16 40.9 GB
Make sure you have already installed ComfyUI-GGUF custom node by Author-city96. If not done, just install it from the Manager by selecting Custom Nodes Manager option. If you do not know what's GGUF models, you can get detailed overview from our model quantization tutorial.
3. All the other models (text encoders, VAE, lora) will be same as used for basic Qwen Image model. You do not need to download again. But if you want then download them :
(a) Download Text encoder (qwen_2.5_vl_7b_fp8_scaled.safetensors) and save it into your ComfyUI/models/text_encoders directory.
(b) Download Vae (qwen_image_vae.safetensors) and save it into ComfyUI/models/vae folder.
(c) Download Lightning Lora model (Qwen-Image-Lightning-4steps-V1.0.safetensors) if you want then save it into your ComfyUI/models/loras directory.
4. That's it, just restart and refresh ComfyUI to take effect.
Workflow
1. You need to download the workflow (Qwen_Image_2512.json) from our Hugging face repository.
2. Next follow the steps:
(a) Drag and drop workflow into ComfyUI. You can use lora that has been embedded into it. Just enable the node and upload the Qwen-Image-Lightning-4steps-V1 model.
(b) Load Qwen image 2512 model into Load diffusion model loader node. Use Unet loader node instead for GGUF models.
(c) Load the Vae, text encoders and lora.
(d) Set the KSampler Settings for different model variants:
BF16 variant-
CFG-4
Steps- 50
FP8 variant-
CFG-2.5
Steps- 20
FP8 + lighting lora variant-
CFG-1
Steps- 4
Perfect Supported aspect ratios:
1:1 - 1328x1328, 16:9 - 1664x928, 9:16 - 928x1664, 4:3 - 1472x1104, 3:4 - 1104x1472, 3:2 -1584x1056, 2:3 -1056x1584
(e) Put the prompt into prompt box. To get the better output the model needs detailed prompting. Finally, hit the Run button to start generation.
Prompt- fixed camera extreme macro cinematic close-up of a human mouth partially closed, lips and skin textured and softly lit, the mouth opens and reveals teeth fitted with custom metallic grills spelling out bold sculptural letters "DIFFUSION". , reflective silver-gold material with imperfect handcrafted texture, moisture catching highlights, shallow depth of field isolating the mouth, gritty high-detail realism, intimate and confrontational framing, dramatic contrast between organic skin and cold metal, moody cinematic lighting with warm highlights and deep shadows, tactile film grain, fashion-editorial meets underground music-video aesthetic, hyperreal, high resolution
seed- 595540274
resolution-16:9
steps-50
CFG-4
Instead of chasing gimmicks, it focuses on fixing what actually frustrated users like unnatural humans, weak fine details, and inconsistent text rendering. The result is a model that feels far more usable for creators who care about realism and accuracy.
If you are working with human centric visuals, nature scenes, or text heavy image compositions, this update is a meaningful step forward from the August release. It doesnot just generate images but it generates images you are more likely to actually use.
- Improved Realism){kind=link}