How to Run Qwen-Image 2.1 Locally with ComfyUI: Full Setup Guide
Install Qwen-Image 2.1 in ComfyUI: model files, VRAM needs, and workflow setup for this 7B text-to-image and editing model with native transparency.
What is Qwen-Image 2.1?
Qwen-Image 2.1 is a 7 billion parameter image model from Alibaba’s Qwen team that handles text-to-image generation, image editing, and native RGBA transparency in a single model. Published on Hugging Face under Qwen/Qwen-Image-2.1, it ships as a diffusers pipeline with separate transformer, text encoder, and VAE components, the same modular structure that ComfyUI workflows expect. You can generate images from a plain text prompt, edit existing images with a text instruction, or ask for transparent (alpha channel) output directly, without a separate background-removal step.
TL;DR
- Qwen-Image 2.1 is a 7B parameter diffusion model that unifies text-to-image generation, instruction-based editing, and native transparency into one checkpoint.
- Running it locally through ComfyUI requires manually placing three types of files: diffusion model weights, text encoder files, and a VAE, each in its own models subfolder.
- Fully loaded, the model consumes roughly 27 to 30GB of VRAM, putting it in the same bracket as other current-generation image models rather than something a low-VRAM card can handle comfortably.
- On a 4090, a 1024x1024 image generates in about 7.5 seconds, while pushing resolution up to 2048x2048 or higher stretches generation time to 30 seconds or more, and 4096x4096 can take over three minutes and introduce artifacts.
- The model supports multi-image reference composition, feeding several generated objects back in as reference images so it can arrange them into one cohesive scene.
- Independent testing across cultural and architectural prompts (Egypt, Malaysia, Rwanda, Ukraine, Belarus, and more) showed strong stylistic accuracy, though results vary and some scenes need better prompting to nail fine details like signage text or crowd composition.
- Sampler settings matter: Qwen-Image 2.1 reportedly performs best with a CFG value close to 1, meaning minimal steering away from the base prompt.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
How do you install Qwen-Image 2.1 in ComfyUI?
Installation is a manual file-placement process rather than a one-click installer. You need ComfyUI already installed and running before you start. If you haven’t set that up, get ComfyUI installed first since every step below assumes it’s already in place.
From there, the process breaks into three downloads, all sourced from the model’s Hugging Face repository:
- Diffusion model files. Go to the Hugging Face repo’s Files tab, open the
diffusion_modelsfolder, and download the model weight files (the repo splits the transformer into two safetensors shards). Save these into your ComfyUImodels/diffusion_modelsdirectory. - Text encoder files. Back in Files, open the
text_encodersfolder and download the text encoder files. The repo ships the text encoder across four safetensors shards plus config and tokenizer files. Place these inmodels/text_encoders. - VAE file. Open the
vaefolder and download the variational autoencoder weights, the smallest file of the set, intomodels/vae.
Once all three sets of files are in their correct folders, load a Qwen-Image 2.1 workflow in ComfyUI, drop your prompt into the text node, and hit run.
What do the diffusion model, text encoder, and VAE actually do?
Each component handles a distinct part of the pipeline, and understanding the split helps when something goes wrong or you want to swap a component for a lighter version:
- The diffusion model (transformer) is the actual image generator, the network that draws the picture step by step.
- The text encoder converts your written prompt into a numerical representation the diffusion model can act on.
- The VAE translates the model’s internal latent representation into the final pixels you see on screen.
In the sampler node, a few settings control how that generation unfolds. The sampler is the algorithm that removes noise step by step to build the image. The scheduler controls the pacing of that noise removal. CFG (classifier-free guidance) determines how strictly the model sticks to your prompt, and Qwen-Image 2.1 reportedly performs best with a CFG close to 1, meaning minimal steering. Dtype sets the numerical precision the model runs at, where lower precision trades a bit of quality for speed and reduced VRAM use.
How much VRAM does Qwen-Image 2.1 need?
Fully loaded, Qwen-Image 2.1 consumes somewhere in the range of 27 to 30GB of VRAM, consistent with other recent-generation image models of similar scale. That puts a 24GB card like an RTX 4090 close to the edge, workable for most standard resolutions but tight if you’re stacking large batch sizes or running at very high resolutions alongside other loaded models.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Generation speed scales with resolution rather than VRAM alone. On a 4090, a 1024x1024 image comes out in about 7.5 seconds. Bumping to 2048x2048 pushes that to around 30 seconds, and a 2688x1536 image landed in a similar 30-second range. At 3072x3072, generation time climbed to roughly 95 seconds. Pushing all the way to 4096x4096 took over three minutes and produced visible texture artifacts and unusual patterning, suggesting that resolution has a practical ceiling before quality degrades. If you’re running on a 3090 instead of a 4090, expect roughly double the generation time at each resolution tier.
Is Qwen-Image 2.1 worth running locally?
For anyone doing serious text-to-image or image-editing work, yes, with caveats. The model’s strengths show up clearly in a few areas:
Native transparency is the standout feature. Rather than generating an opaque image and running it through a separate background removal tool, Qwen-Image 2.1 can output RGBA images with a real alpha channel directly, tested successfully with detailed subjects like a phoenix made of fire and molten gold.
Multi-reference composition also works well. Feeding the model several independently generated images (furniture pieces, food ingredients) as reference inputs and asking it to arrange them into one cohesive scene produced results with correct shadow placement, consistent lighting, and plausible spatial arrangement, even for unusual, likely out-of-training-distribution requests like assembling a traditional South Asian paan from individually generated ingredients.
Cultural and stylistic range tested reasonably well across a wide set of prompts spanning Egyptian calligraphy, Bulgarian cobblestone streets, Malaysian street scenes, Rwandan textile patterns, Ukrainian architecture, and New York diner signage. Results weren’t uniformly excellent (some scenes felt sparse or the framing was off), but the model generally captured recognizable cultural and architectural cues rather than defaulting to generic outputs.
Editing also holds up: instructing the model to change clothing, add jewelry, or remove an object from a person’s hand while preserving pose worked cleanly in testing, with the model making sensible small adjustments beyond the literal instruction.
Weak points show up at extremes: very high resolutions introduce artifacts, text rendering in generated scenes (signage, storybook plates) can include misspellings or nonsensical characters, and fine anatomical or historical detail (a cat’s fur, the Sphinx’s original appearance) doesn’t always land. Feeding the model iterative feedback, or pairing it with an agent that can look up official prompting guidance and refine outputs automatically, noticeably improved results in testing compared to single-shot prompting.
Frequently Asked Questions
What GPU do I need to run Qwen-Image 2.1?
You need a GPU with at least 24GB of VRAM to load the model comfortably, since it consumes roughly 27 to 30GB when fully loaded. A 4090 handles standard resolutions in single-digit to tens-of-seconds generation times; a 3090 works but runs roughly twice as slow.
Can Qwen-Image 2.1 generate transparent images directly?
Yes. Native RGBA transparency is one of the model’s built-in capabilities, meaning it can output images with a real alpha channel without a separate background removal step.
Does Qwen-Image 2.1 support image editing, not just generation?
Yes. You can feed it an existing image along with a text instruction (for example, changing an outfit or removing an object from a subject’s hand) and it will apply the edit while preserving pose and composition.
What resolution should I use for the best quality?
Testing suggests 1024x1024 up to around 2688x1536 or 3072x3072 produces reliable results. Pushing to 4096x4096 introduced visible artifacts and texture patterning in testing, so it’s not recommended as a default setting.
Where do the model files go in ComfyUI?
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
Diffusion model weights go in models/diffusion_models, text encoder files go in models/text_encoders, and the VAE file goes in models/vae. All three are downloaded from the model’s Hugging Face repository’s Files tab.
