Version 2: I’ve added a Turbo LoRA; it really speeds things up significantly. I also included an optional section for upscaling and enhancing image quality.
I posted how to use the Lora Turbo in the comments section.
Versión 2: He añadido una lora turbo, realmente hace las cosas muy rápido. También añadí un apartado que se puede habilitar o no para hacer upscale y añadir calidad a la imagen.
En la sección comentarios puse cómo usar la Lora turbo.
Version 1.5:
I added more explanatory notes and the option to choose between an empty latent and a latent that already includes the prompts and the pose.
Version 1.5:
Añadí más notas explicando cosas y la posibilidad de elegir entre latent vacío o un latent con los prompts y la pose ya metida.
Hello:
This is a workflow designed to run Qwen2.1 with low VRAM. Unfortunately, it cannot be used with SDXL because the core of the workflow is a LoRA specifically made for Qwen2.1.
By using the Qwen 2.1 GGUF model, it can run on 8GB of VRAM. You can also switch it to CPU mode to avoid relying on the GPU and consuming VRAM. Additionally, I’ve included RAM and VRAM cleanup nodes to ensure memory is freed up for the next step of the workflow. I've added a node to enable the use of wildcards. I hope you find it useful.
Graph includes both loaders. Prefer native INT8 + Dynamic VRAM. Use the GGUF branch if that OOMs or Comfy hangs on your card.
The VNCCS_QI2_PoseStudio LoRA is trained so that Qwen 2.1 interprets "image 1 = mannequin, image 2 = character" and transfers the pose without ruining the face, hair, or clothing. The author himself noted that without the LoRA and the Qwen prompt, it works... only halfway. Difficult poses fall apart completely.
La LoRA VNCCS_QI2_PoseStudio: Está entrenada para que Qwen 2.1 lea “imagen 1 = maniquí, imagen 2 = personaje” y migre la pose sin destrozar cara, pelo y ropa. El propio autor lo dejó escrito: sin LoRA y sin el prompt Qwen, funciona… a medias. Poses difíciles se van al garete.
https://civarchive.com/models/2957080/vnccs-posestudio-qwen-image-21
The Qwen-Image 2.1 model is a very recent and powerful technology. The specific issue you are encountering with the qwen-image-2.1-Q4_K_M.gguf file arises because the main GGUF node repository (city96/ComfyUI-GGUF) has not yet formally updated its main branch to recognize the Qwen 2.1 internal architecture, resulting in an "Unknown model architecture!" error. To run Qwen-Image 2.1 in an ultra-lightweight manner on your PC, there is a solution: use the updated GGUF fork (recommended for GGUF). Since the official city96 node is lagging behind this release, the community is actively using the leejet fork, which already includes full support for the Qwen-Image 2.1 architecture:
Go to your custom nodes folder: ComfyUI\custom_nodes\. Delete or move the ComfyUI-GGUF folder elsewhere (if you already have it; otherwise, skip this step). Open a terminal in that custom_nodes directory and install the compatible version by running:
git clone https://github.com/leejet/ComfyUI-GGUF
If you don't want to deal with switching Git repositories, the official ComfyUI ecosystem offers native support for Qwen-Image 2.1 via reduced-precision formats that do not require the GGUF node. The base model, compressed to FP8 or INT8, has a similar footprint and consumes only about 7 GB of VRAM. To set this up efficiently: for the Diffusion Model, download the lightweight native version (e.g., in e4m3fn or int8 format) directly from the official Comfy-Org repositories and place it in models/diffusion_models/. It loads using the standard native UNETLoader node. The key trick to saving VRAM is moving the Text Encoder to system RAM: if you download the qwen3vl_8b_int8_convrot.safetensors version (~9 GB) and configure the workflow to run on system RAM (CPU) instead of VRAM, you free up almost all of the graphics card's resources. Since the text is processed only once at the start of generation, the impact on speed is negligible, allowing even low-end PCs to run it smoothly.
In the "MODELS AND SAMPLERS" node, you have the DEVICE and DEVICE_1 options to set it to CPU mode!!!
https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/text_encoders
https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/vae
https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/diffusion_models
https://huggingface.co/leejet/Qwen-Image-2.1-GGUF/tree/main
Hola:
Este es un workflow para poder usar qwen2.1 con poca VRAM. Desafortunadamente no se puede usar con SDXL porque el cerebro del workflow es esta Lora que es para Qwen2.1
Usando el modelo GGUF de Qwen 2.1 se puede usar con 8GB de VRAM. También se puede poner en modo CPU para que no tire de GPU y consuma GB de VRAM. Además le he puesto nodos de limpieza de RAM y VRAM para que deje toda la memoria libre en el siguiente paso del workflow. He puesto un nodo para que se puedan usar wildcards. Espero que se pueda usar.
El grafo incluye ambos cargadores. Se recomienda usar INT8 nativo + VRAM dinámica. Utiliza la rama GGUF si se produce un error de memoria (OOM) o si Comfy se bloquea en tu tarjeta.
La LoRA VNCCS_QI2_PoseStudio: Está entrenada para que Qwen 2.1 lea “imagen 1 = maniquí, imagen 2 = personaje” y migre la pose sin destrozar cara, pelo y ropa. El propio autor lo dejó escrito: sin LoRA y sin el prompt Qwen, funciona… a medias. Poses difíciles se van al garete.
https://civarchive.com/models/2957080/vnccs-posestudio-qwen-image-21
El modelo Qwen-Image 2.1 es una tecnología muy reciente y potente. El problema exacto que estás experimentando con el archivo qwen-image-2.1-Q4_K_M.gguf se debe a que el repositorio principal de nodos GGUF (city96/ComfyUI-GGUF) aún no se ha actualizado formalmente en su rama principal para reconocer la arquitectura interna de Qwen 2.1, arrojando el error Unknown model architecture!. Para que puedas correr Qwen-Image 2.1 de forma ultra-ligera en tu PC, tienes una solución: Usar el fork actualizado de GGUF (Recomendado para GGUF)Dado que el nodo oficial de city96 se ha quedado un paso atrás con este lanzamiento, la comunidad está utilizando activamente el fork de leejet, el cual ya incluye soporte completo para la arquitectura de Qwen-Image 2.1:
Ve a tu carpeta de nodos personalizados: ComfyUI\custom_nodes\. Borra o mueve a otro lugar la carpeta ComfyUI-GGUF (si ya la tienes, sino no). Abre una terminal en esa ruta de custom_nodes e instala la versión compatible ejecutando:
git clone https://github.com/leejet/ComfyUI-GGUF
Si no quieres lidiar con cambiar de repositorios Git, el ecosistema oficial de ComfyUI da soporte nativo a Qwen-Image 2.1 a través de formatos de precisión reducida que no requieren el nodo GGUF. El modelo base comprimido en FP8 o INT8 ocupa casi lo mismo y consume solo unos 7 GB de VRAM. Para configurarlo de esta manera de forma muy ligera: El Modelo de Difusión: Descarga la versión nativa ligera (por ejemplo, en formato e4m3fn o int8) directamente desde los repositorios oficiales de Comfy-Org y colócala en models/diffusion_models/. Se carga con el nodo nativo UNETLoader estándar. El truco maestro para ahorrar VRAM (Mover el Text Encoder a la RAM)
Si descargas la versión qwen3vl_8b_int8_convrot.safetensors (~9 GB) y configuras el flujo para que se ejecute en la memoria RAM del sistema (CPU) en lugar de la VRAM, liberarás casi toda la tarjeta gráfica. Como el texto solo se procesa una vez al inicio de la generación, el impacto en la velocidad es casi nulo y permite que PCs de gama baja lo corran sin problemas.
¡¡¡En el nodo "MODELS AND SAMPLERS" TIENES LA OPCIÓN DEVICE Y DEVICE_1 PARA PONERLO EN MODO CPU!!!
https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/text_encoders
https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/vae
https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/diffusion_models
Description
FAQ
Comments (6)
Why are you still using GGUF, GGUF is slow and redundant, dynamic in VRAM makes GGUF pretty much worthless.
I'm on 8gb VRAM here and never looked back to GGUF, maybe if someone is on a 4gb GPU with 16gb of RAM there might be a case but GGUF just isn't recommended anymore.
I just say as your post makes it look like low VRAM users should use GGUF.. ?
Ah, well, that’s what I found when I looked up how to use Qwen2.1 with low VRAM. The thing is, I have an RTX 5090, and if I use dynamic VRAM, ComfyUI crashes on me. Could you tell me how to change that setting, or where I can find information about it?
Oh, I forgot to mention that I also included both options in the workflow—the standard one and the GGUF version—in case they wanted to switch.
@saehara151 Bad info still going around I guess. The comfy devs actively dissuade GGUF use.
I'd recommend you join the banodoco discord if you use discord. That is probably the best place to keep up with the goings on and some developers are there (like Kijai) who can help you.
As for Dynamic VRAM issues sorry I don't know enough about your system but if you search banodoco or ask there or the comfyui discord you might get some helpful suggestions. I do know and have seen people mentioning turning off Dynamic VRAM so that is a thing which I guess you've done? Not sure why you'd get that but if you have a beefy system which it seems you do I think most things you can get away without it? It is particularly helpful for people with less resources though.
@MrTitsworth I created this workflow because a user with limited VRAM asked me if using Qwen 2.1 was mandatory; that’s why I made this version. The original one works fine for me—here it is: https://civitai.red/models/2958698/ai-character-pose-director-qi21
I’ve found a Turbo LoRA for Qwen2.1.
You can choose between the 4-step, 5-step, or 6-step versions.
You can download it here:
https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo/tree/main
You need to copy this file into the documents/comfyui/custom_nodes folder:
He encontrado una Lora turbo para Qwen2.1.
Ahí podéis elegir entre la de 4 pasos, 5 pasos ó 6 pasos.
Podeis descargarla aquí.:
https://huggingface.co/Viggle/Qwen-Image-2.1-viggle-turbo/tree/main
Teneis que copiar este archivo en la carpeta documents/comfyui/custom_nodes
I’ll update the workflow so you can use it.
Haré una actualización del wrokflow para que podais usarla.