v3: Cleaner, more eye-catching, more vibrant... otherwise it has the same features as v2. all settings are the same...
v2: Created using Qwen Image 2512. Produces extremely realistic and detailed images. Provides cleaner outputs compared to the first version. Ideal for NSFW content but not very creative for hardcore scenes.
v1: This is my first qwen share. It was created to achieve realistic and cinematic style visuals. It may not be suitable for anime-style visuals. It is EXTREMELY suitable for NSFW style content. Despite its large size, it is extremely fast. I created it for fast use. If you are going to use it with LORAs, keep the LORA weight values very low. Otherwise, the images you create may be distorted.
The most suitable usage I have found:
Steps: 6-8
cfg: 1
Resolution: 1328x1328 or 1440x1440 resolution and 2:3, 3:4, 16:9, 1:1, and all other ratios
Sampler: Euler a + beta/ bong tangent/ DDIM Uniform
Sigma Shift: 2-3
Using these settings, it takes about 25-30 seconds to generate an image at 1440x1440 resolution and 8 steps on my RTX3090 computer.
IMPORTANT: If you use settings other than the ones I provided (different samplers other than Euler-A, clip skip:2, high-value LORAs, high sigma shift values, etc.), you will most likely get corrupted images.
Description
This is "Qwen Image 2512" version.
FAQ
Comments (10)
Any chance you could post the clip and vae separately (if they're custom). Or post the model without clip/vae?
I created my own model by adding certain LORA weights to the Qwen 2512 base model in specific proportions. I then converted the resulting BF16 model to FP8 and published it. During this process, I used Qwen's default clip and VAE files. When creating my model, it also generated clip and vae files, but I don't use them (in fact, I deleted them to avoid taking up unnecessary space). I use Qwen's own vae and clip files. I use SwarmUI to generate visuals. Swarm automatically uses the clip and vae files it downloads for Qwen.
@53rt5355iz So there's no baked in VAE/CLIP in the v2 model?
My friend, I don't know the technical details. Does the process I performed embed the VAE and clip files into the model? If so, I don't have a separate version of these files. And I don't know how to separate them. Although the size of this all-in-one model is large, it's quite fast, and that works for me. But if there's someone who knows how to separate it and is interested in doing so, they can do it and publish it on their own account.
@53rt5355iz No worries friend :). This is an amazing model you've made here.
Something to easily separate them if there is an included VAE and Clip is just use ComfyUI's normal "load checkpoint" node with the 3 separate ouputs, then connect just the "checkpoint" output to a "Save diffusion model", and leave the "clip" and "vae" outputs not connected to anything.
Run that, and it will just pull the checkpoint weights from the bundled file and save it without the clip or vae.
I did that with another Qwen model and cut out a few GB of space.
@Griphen116 Thank you, I'll try that...
@Griphen116 Thanks for the info :). I managed to split them. TBH I didn't realize the checkpoint loader worked ;)
Thank you for your work, but is it possible to create a 4KM gguf version? At the current size, it's too big for my pc so it would be great if you could release a quantized version instead. Thanks!
Unfortunately, I can only convert to FP8.