About these variants
Four precision levels are available for Qwen-Image-2.1, trading VRAM and speed against generation quality:
- BF16 — full-precision reference (~14 GB). Highest quality, largest footprint.
- INT8 (W8A8) — 8-bit weights + 8-bit activations (~7 GB). Near-lossless quality, ~2× smaller than BF16, runs on INT8 tensor cores (RTX 30-series and up).
- INT6 (W6A8) — 6-bit weights + 8-bit activations (~5.5 GB). Middle-ground: smaller than INT8 with only a small quality trade-off.
- INT4 (W4A8) — 4-bit weights + 8-bit activations (~4 GB). Smallest footprint; uses ConvRot with a per-tensor codebook that decodes to INT8 for compute, so it runs on the same INT8 hardware as W8A8. Larger quality trade-off than INT6.
Qwen-Image-2.1 7B — INT4 (W4A8) & INT6 (W6A8) ConvRot for ComfyUI
INT4 and INT6 quantized weights of Qwen-Image-2.1 for fast, low-VRAM inference in ComfyUI. This is a modified (quantized) version of the Qwen-Image-2.1 model. It is not an official Qwen release and is not endorsed by the Qwen team.
About Qwen-Image-2.1
A unified text-to-image generation and image editing model in the Qwen family. With just 7B parameters in its visual generation component (32 Single-Stream DiT layers), Qwen-Image-2.1 balances generation quality, inference efficiency, and versatility.
Highlights
- Efficient Image Generation — combines strong visual performance with fast inference and a compact design, making high-quality image creation accessible across a wide range of creative workflows.
- Flexible Creative Control — supports diverse inputs, outputs, and localized edits, giving creators the flexibility to explore ideas and refine details within a unified workflow.
Key improvements in 2.1
- Compact and Efficient — lightweight architecture with mixed-granularity attention and prefix KV cache reuse delivers strong image quality at low computational cost.
- Native Transparency, Unified Creation and Editing — generate regular or transparent (RGBA) images from text, edit transparent layers, and extract subjects from photographs — all in one model.
- Versatile Editing — support up to 10 reference images, specify local edits via circles, painted annotations, or separate masks, and preserve identity for people and products.
- Realistic Textures and Refined Aesthetics — improved typography, portrait lighting, and fine details for more visually compelling results.
License
Qwen-Image-2.1 is licensed under the [Qwen Research License] These files are a quantized derivative and are distributed under the same license — see the original repository for full terms, permitted uses, and any commercial-use restrictions.
Description
FAQ
Comments (34)
damn Qwen has this lifeless ai vibe aesthetic in most images
Yeah, it seems like it, LoRAs are definitely needed for T2I. The edit capabilities are a bit more interesting.
@tsolful yeah loras can fix this without problem however artifacts are very difficult to fix. This image has that
As long as it's open sourced, trainable, and the base knowledge and quality is ok,
all you need is to trust the community. 💪
@ikekph5 I trust the community 100%, I'll probably train a new fantasy realism refiner when AItoolkit supports. Non-commercial licensing tho :/ https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE
@tsolful the license is ... unfortunate ... Why just follows other qwen model like qwen3.8 with a revenue cap. Qwen image 2.1 is an old model to them anyways.
edit: GPT tells me they probably don't care about the old model. But the model doesn't seem to have safety filter and is too good at nsfw so they probably want the "research or evaluation purposes only" shield for lawsuits.
@ikekph5 This was posted to the QwenDev twitter account, so maybe it is not as bleak as it seems.
"We’ve received so much love for Qwen-Image-2.1 over the past 24 hours, thank you!! Also gotten a lot of questions about the license, especially around model outputs. So here’s the answer: Outputs are not part of the licensed Materials. Users retain the rights to images and other content they generate using the model."
@ikekph5 As mentioned by @tsolful devs don't care about output, as long as the old school civitai FOSS community does their thing on training finetunes and loras we'll all be eating good.
It's really poor in comparison of Krea 2. I hope the editing features are better. Thanks for the INT8 anyway.
the license kills this model
It is far and away the best editing model I've messed with, robust and surprisingly fast.
It's an omni edit model so it's more like a replacement to the Flux 2 Klein.
@VeerGeer Yes, the editing is very nice. I keep Krea 2 for image generation and replace Flux Klein 9B by Qwen 2.1 for editing !
And now INT8 is gone...
@nonemo13 Yeah, I was messaged by the civit team after they forcibly archived my whole page, as they said: "as not to confuse people with the official release." Went back and forth and i had to delete the int8 and they unarchived so i can keep up the int4. Going forward, I think I'll be uploading int8/int4 until they officially release their versions, then unpublish the int8. Disappointing but understandable from a certain perspective.
cant wait for the Loras, its actually good
still seeing the old halftone thing going on in image edit, so they didn't fix that in their new vae, lol
very light NSFW test showing that fine‑tuning is certainly needed, but even the bare knowledge is quite good. At the very least, there’s no censorship.
i honestly dont even know why loras are necessary, its a multireference image model, just place whatever nsfw stuff you want into imgag2 and tell it to do that (although good lora never hurts)
@Azzyboi Lora is useful for clean txt2img. Plus, in this case, you can come up with an idea for which it’s hard to find a reference. Or the model can sometimes come up with a funny gem on its own.
The camera angle changes are weak, but it's the best image enhancer (open source) I've seen so far. The consistency of character is very good. The screen-shot improvement is quite good.
It worked well when I used a prompt like this.
---
General Order: Rotate the camera angle 90 degrees to face the front of the character in image1.
Order1: The character is now directly facing the camera in a full front-facing view.
Order2: Both eyes and the front of the face/body should be fully visible, looking straight at the viewer.
Note: Original image1 is left side view of the subject.
---
@Nagatsuki16 Yours is not work. I came up with this idea and it worked 90% of the time.. A bullet-proof-shot. place the camera in the described location of the **motionless-woman** in picture 1;
back-left quarter view eye-level shot wide shot
Really good, thank you.
Image Edit is the best out there.
Image generation is good but the first example that came with comfyui's workflow gave me bad hand and fingers and it took 30second to generate on my 8gb 3060ti. i made better verion of that image with Krea 2 Trubo in around 14 seconds.
PS: it also does face swap. sometimes the head is too small for body and you have to generate couple more images to get it right
"swap the face of woman in <image1> with the face of the woman in <image2> also transfer the original facial features, nose, eyes and hair from <image2>"
it needs a turbo loda
@RedditUser9811 Yep
@RedditUser9811 @infernahermit846 https://civitai.red/models/2958738/qwen-21-turbo-loras?modelVersionId=3351727
@tsolful superb, Thank you. good lora for generation but not really good for editing, it edits but changes the character's position.
like this: step1 character is correct, step 2 character slightly moved to the left (same pose), step 3 character moves slightly to right and step 4 character moves slightly to left.
@infernahermit846 Yeah they mentioned it when they posted, i added it to description. hopefully 1.0 improves upon the editing capabilities
Its a good model. I wanted it for sharpening/detailing images for reference in MiniMaxH3. Is was used in a few images that fit what I'm going for so I thought id check it out.
It did okay, but I had problems with facial consistency when using same seed, same prompt, and same ref. Not anything most people would notice, but it was there. I had to switch to the full model. That is not to say *ANYTHING* against this model by any means. It's output was great. The fault was mine. I was trying to use the wrong size screw driver for the screw I was trying to turn.
It is a good model. just not what i needed.
For users of Forge, is this or another model a better choice for an edit model?
I havent got to actually test it yet, in fact I didnt even realize neoforge supports it, I guess thats cause its the same architecture as the last QWEN model?
That said this one is almost certainly going to be your best bet- its the latest and all the examples I've seen are way better than anything you could do with the old qwen or flux edit models.
This model likes to simply copy+paste reference image faces (without regard for lighting or proportions) instead of using them as starting points and blending them into images naturally. I'm sure you can prompt it into submission but I'm surprised this is the default behavior.
it weirdly does that better at 1mp, at least paired with tha tfamous bfs_face swap lora
Details
Files
qwen_image_2.1_vae_bf16.safetensors
Mirrors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
QwenImage2.1-ae.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwen_image_2.1_vae_bf16.safetensors
qwenImage21INT8INT4_int8_txt.safetensors
Mirrors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
arcpoly_vision_reader.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwen3vl_8b_int8_convrot.safetensors
qwenImage21INT8INT4_int8.safetensors
Mirrors
qwen_image_2.1_int8_convrot.safetensors
qwen_image_2.1_int8_convrot.safetensors
qwen_image_2.1_int8_convrot.safetensors
qwen_image_2.1_int8_convrot.safetensors
qwen_image_2.1_int8_convrot.safetensors
qwen_image_2.1_int8_convrot.safetensors
QwenImage2.1-FP8.safetensors
qwen_image_2.1_int8_convrot.safetensors
qw21.int8cr.safetensors
qwen_image_2.1_int8_convrot.safetensors
qwen_image_2.1_int8_convrot.safetensors
qwen_image_2.1_int8_convrot.safetensors
