Home
Creators
Search
Recent
Random
Posts
DMs
Tags
Random
Importer
Import
FAQ
Account
Register
Login
Favorites
Keys
Logout
Telegram
Qwen Image 2.1 - 20260921 - For Patreon Supporters (Patreon)
Downloads
Content
Hey everyone.
If you watched the Qwen Image 2.1 video https://www.patreon.com/aifuturetech/posts/qwen-image-2-1-170156721 , you saw the model handle text to image, image edit, transparent backgrounds, and up to ten reference images, all in one 7 billion parameter model. The free posts give you the two official ComfyUI workflows. This post is my custom one-pass workflow with the prompt enhancer node, plus the download links and settings I actually used.
First, the downloads. Everything you need is below with direct links.
All model files come from the Comfy-Org repack:
https://huggingface.co/Comfy-Org/Qwen-Image-2.1
Diffusion model (pick one):
- qwen_image_2.1_bf16.safetensors — 14.2 GB, full precision
- qwen_image_2.1_int8_convrot.safetensors — 7.26 GB, lower VRAM
Folder: ComfyUI/models/diffusion_models/
Text encoder (pick one):
- qwen3vl_8b_bf16.safetensors — 17.5 GB, full precision
- qwen3vl_8b_int8_convrot.safetensors — 9.35 GB, lower VRAM
- qwen3vl_8b_w4a8.safetensors — 6.31 GB, lowest memory
Folder: ComfyUI/models/text_encoders/
VAE:
- qwen_image_2.1_vae_bf16.safetensors — 676 MB
https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/vae/qwen_image_2.1_vae_bf16.safetensors
Folder: ComfyUI/models/vae/
Prompt Enhancer custom node (mine):
- https://github.com/benjiyaya/ComfyUI-Qwen-Image-2.1-Prompt-Enhancer
Install via ComfyUI Manager, or git clone into ComfyUI/custom_nodes/
Prompt Enhancer models (both required for the prompt rewrite nodes):
- qwen3.5_9b_qwen_image_2.1_pe_t2i.int8_convrot.safetensors — 9.47 GB, text-to-image prompt rewriter
- qwen3.5_9b_qwen_image_2.1_pe_i2i.int8_convrot.safetensors — 9.47 GB, image-edit prompt rewriter
Folder: ComfyUI/models/text_encoders/
T2I With Prompt Enhancer
Image Edit With Prompt Enahcner
The workflow ships pointing at the BF16 files. If you want the int8 build instead, repoint the loaders to the int8 files. The rest of the workflow is identical.
----------------------------------------
THE WORKFLOW
----------------------------------------
This is one workflow that does both text to image and image edit, one pass each. The main thing that is different from the official template is the prompt enhancer. You type a short idea, and the node expands it into the formatted prompt the model actually wants.
Why that matters: Qwen Image 2.1 does not reward long essay prompts. It wants a structured prompt, and the Qwen team ships a system prompt for that. I built the enhancer node on top of that same logic, so four words in gives you a full formatted paragraph out, and for text to image it picks the aspect ratio for you too.
So the workflow has two sides. Top is text to image. Bottom is image edit. Same model, same loaders, one pass sampling on both.
----------------------------------------
QUICK START
----------------------------------------
1. Download the three model files above (diffusion model, text encoder, VAE) and drop them in the folders listed.
2. Install the Prompt Enhancer node.
3. Load QwenImg_2-1_t2i-imgedit-PromptEnhancer.json.
4. Check the UNETLoader, CLIPLoader and VAELoader point at the files you downloaded.
5. Type a short idea into the prompt enhancer node.
6. Set the aspect ratio in the ResolutionSelector.
7. Queue. That's the whole loop.
----------------------------------------
WHAT TO INPUT (PROMPT FORMAT)
----------------------------------------
Text to image: type a short idea. "A mechanical automatic watch in exploded view showing the gears and assembly." The enhancer expands it and sets the aspect ratio.
Image edit: reference your inputs with tags. The model reads <image1>, <image2> and so on, up to 10. Example:
"Put the jacket and dress from <image2> onto the character in <image1>, keep the face, hair, pose and background the same."
Feed the reference images into the LoadImage nodes in order. <image1> is the first one you connect.
Transparent background: if you want an RGBA cutout, say it in the prompt. The recommended format is:
"This is an RGBA image with transparency. <your description>. The image has alpha channel and the background is transparent."
----------------------------------------
KEY SETTINGS (AND WHY)
----------------------------------------
STEPS: 30, CFG 1, euler, simple. CFG stays at 1, this model was trained for it, and higher CFG burns the output. 30 steps is plenty. The model is fast, so I run 50 sometimes and it is still quick.
RESOLUTION: the ResolutionSelector drives it. 1 megapixel at 16:9 is the default. The model supports native 2K up to 2752 x 1536 for 16:9, so push it higher if you have the VRAM.
PROMPT REWRITE (temperature 1, top_p 0.95): these are the enhancer node settings and they ship correct. Leave them. The only one you might touch is max_new_tokens, but the defaults are already generous.
SAGE ATTENTION: auto, leave it. If SageAttention is not installed, the node falls back on its own.
----------------------------------------
TIPS
----------------------------------------
1. Start with the official free t2i workflow first to confirm the model loads. Then move to this one. Do not debug the enhancer node and the model at the same time.
2. For image edit, give the model enough information. I tested a camera angle from outside the window and it failed because the reference image had no face. Feed a character sheet or a front view and it predicts the opposite side correctly.
3. Numbers and text in generated images are not reliable. A blueprint comes out looking right until you read the measurements. Treat it as a layout concept, then fix the numbers yourself.
4. For try-on, the clothing pattern matters more than the pose. The flower patterns and jacket textures carried over almost perfectly, but a pair of pants came out with two straps untucked. Check the fabric details before you ship the image.
5. Interior design mockups work, but sizes are not exact. The model has no spatial measurement, so a coffee table can come out the wrong scale. Good for a concept, not for a professional floor plan.
----------------------------------------
COMMON ISSUES
----------------------------------------
"Out of memory" — This is the big one. The BF16 build is about 32 GB total. Use the int8 diffusion model (7.26 GB) and the int8 or w4a8 text encoder. 12 GB VRAM with offloading, 16 GB comfortable.
"Custom node not found" — The prompt enhancer node is not in default ComfyUI. Install ComfyUI-Qwen-Image-2.1-Prompt-Enhancer via Manager or git clone.
"Transparent image comes out with a background" — Check your prompt uses the RGBA format line. The model decides transparency from the prompt, so if you do not ask for it, you get an RGB image.
"Camera angle is wrong" — The reference image did not have enough information. Add a front view or a character sheet, or a second angle of the same room.
----------------------------------------
Questions? Drop a comment with your error. I check these regularly.
Thanks for supporting the channel. You guys make these deep dives possible.
— Benji
Files
