✨ Save 50% on Annual Plans ✨ Claim 6 months FREE →
All the best AI models in one place
Nano Banana, GPT Image, Flux, Seedream, Seedance and more. Try it now for free!
a year ago
Uploaded
this is just a test for my current new workflow for Qwen image.
using the Nunchaku 4bit model and 8 steps at 1920x1088, and it gens them FAST! but i still have to inpaint the faces, since i haven't trained them for Qwen Image yet, not that i could currently use lora's with the Nunchaku version yet. but unlike previously, i'm doing the face fixes at the low-rez step instead of the 14MPix stage.
but even this brand new workflow is now outdated, as i've finally figured out how to get Qwen Image Edit 2509 to do prompts with reference images of my characters. so i'll be trying to swap to that in a bit.
but it's possible i might like the creativity and freedom plain old Qwen Image gives vs Qwen Image Edit 2509. we'll see.
a note: QI is doing really great with hands, fingers, and legs so far. specifically this image required NO mods to anything but the face.
another side note is that using Krea instead of flux-dev, my old loras no longer have the dreaded flux chin!
Creation Settings
This creation was not made on NightCafe.
- Model
Qwen Image Nunchaku 8stp, Krea Nunchaku InPaint&Up
- Prompts
- two OBP's(random prompts) used to get a basic prompt ideas, then fed through Llama-3.1-8B-Lexi-Uncensored_V2_Q8 using a specialized prompt to take the inputs, guidance of what i would like, 8 random locations, then create a prompt with enhanced details and place the character in the prompt. this is the ...
- Workflow
- ComfyUI used to gen the source, then ComfyUI again to do a ComfyUI using Flux Krea Nunchaku for any inpaints, then ControlNet Tiled upscale with Flux Krea Nunchaku, and finally GIMP to do any pre-fixes, as well as final color correction and enhancing, and signaturing.
