Updated: Feb 3, 2026
style
v1.1 (Experimentation Continues)
Both models use a Qwen model in the first stage and then a WAN model in the second.
These are heavy models. Choose an appropriate gguf model if you have memory contrains.
Disable samplers that you don't need. Will post more extensive instructions but remember this is a workflow in progress.
It's no where near stable
v1.0
Workflow:
This update is long overdue since there have been many release to the qwen family of models. I haven't touched qwen in 6 months and felt this needed a modernisation before I commit completely to the zImage models.
2nd Stage Notes:
Step Count and CFG ranges:
* 5, 1.0cfg - low detail images , e.g comics, illustrations etc
* 30,2.0cfg - photorealism, documentary style etc
This workflow is bad at humans in general. It's bad at portraits - head & shoulders, close up portraits and cowboy. I had a much easier time with the previous version of qwen - Please drop any pointers in the comments because i feel prompting is a large part of the equation.
However, It does fine with people in dynamic groups , like crowds, streets, battle scenes.
Notes:
I found the first qwen model to be lighter, faster and sharper images. This workflow is about twice as long. You can adjust your Wan2.2 gguf to Q2 if you are having memory problems.
This workflow made me feel GPU poor on a 3090/24GB. If you have a 50x0 (possible even 40x0) generations the speed up you have in inference may offset the swapping of models. For 50x0 you will be fine with 16gb at the highest settings.