Wan2GP vs ComfyUI: Modelos ocultos, dependencias y VRAM

Escrito el por griquelme
Facebook Twitter WhatsApp LinkedIn

Introducción

Wan2GP presents itself as an all-in-one image generation tool. When I first opened the application, I was immediately struck by the sheer volume of models it seems to support. Initially, I thought it was just a registry of models, but comparing it side-by-side with ComfyUI quickly revealed some major friction points in how these tools handle their ecosystems.

Imagen mostrando la diferencia entre las interfaces de dos programas de generación de imágenes, destacando cómo uno lista muchos modelos y el otro se centra en el flujo de trabajo manual.
Comparación visual entre la interfaz de Wan2GP y el entorno de nodos de ComfyUI.

El misterio de los nombres de archivo

The first thing that caught my eye were the model filenames. I noticed qwen_image_edit_20B_quanto_bf16_int8.safetensors, which looked drastically different from ComfyUI’s standard 40GB version of the same model. My initial assumption was that this was a completely different backbone, but the reality is much simpler: it’s just a quantized INT8 version of the exact same architecture.

La opacidad de la interfaz

I quickly noticed a major friction point in Wan2GP’s user interface. The software lists every single model it supports, but there is no clear visual indicator of which ones are actually present on disk versus which ones need to be downloaded. It creates an ambiguous state where the UI suggests availability without confirming installation.

Descubriendo dependencias ocultas

Then came the real shocker: after I finished downloading what I thought was just a single image editing model, Wan2GP silently started downloading Qwen2.5-VL-7B-Instruct_bf16.safetensors. This Vision-Language backbone is an implicit dependency for the pipeline that wasn’t explicitly mentioned or requested by me.

Vista detallada de una ventana de explorador de archivos con dos ficheros pesados y una flecha roja indicando que uno arrastra al otro.
Los archivos de los modelos descargados en el sistema de archivos, mostrando la dependencia oculta del modelo Vision-Language.

No todo es una trampa: el upscaler

Interestingly, I later stumbled upon a small file named upscale2x_imageonly_real_v1.safetensors. Unlike the massive backbone models that act as gateways for hidden dependencies, this upscaler is completely autonomous. It’s just an image processor and doesn’t trigger any silent downloads of extra encoders.

Diagrama visual que compara el tamaño en memoria de dos modelos de IA diferentes, demostrando cómo uno excede los límites del hardware.
Gráfico de barras mostrando la memoria ocupada por el modelo principal y comparado con un pequeño upscaler.

El límite de la memoria física

The session concluded with a harsh reality check: the model was too large to fit entirely in my 24 GB GPU’s VRAM. The system was forced into partial pinning, offloading data to RAM to survive the load.

Conclusión

In conclusion, while Wan2GP offers a convenient plug-and-play experience with its massive registry of models, it lacks transparency regarding dependencies and hardware limits. For someone who values control and clarity over disk space, ComfyUI remains the superior choice for serious work.