How I looked into better pose and angle control in Qwen Image Edit
Contents
Qwen Image Edit running on an M1 Max was good at preserving a character’s shape, but it was hard to steer pose and angle. Once I looked into the Phr00t AIO base I was using, the version history made the situation a bit clearer:
| version | base |
|---|---|
| v16 | Qwen-Image-Edit-2509 |
| v18 | shifting toward 2511, mixed with 2509 |
| v20 and later | 100% Qwen-Image-Edit-2511 |
That means the v18 behavior I felt as “the model is worse at keeping character structure” may have been the transition to 2511 showing through.
The developer also said:
“v19 is likely best for consistency in edits, while v23 is likely best for prompt adherence.”
| version | characteristic |
|---|---|
| v19 | best consistency for edits, strongest at preserving the character |
| v23 | best prompt adherence, weaker character preservation |
If you want to keep the character design stable, v19 is the one to try. v23 leans harder toward prompt obedience. The Phr00t AIO project is also scheduled to end.
nunchaku quantized build
The Nunchaku quantized build gives a very different tradeoff. It is meant to run lighter, but the price is that control becomes looser and image quality may vary.
If your goal is direct, reliable pose control, quantization alone is not the answer. It is useful when memory is tight, but not when you want the model to obey a very specific pose or camera angle.
VNCCS Pose Studio
VNCCS Pose Studio is the most interesting external attempt I found.
relation to Qwen Image Edit
It sits on top of Qwen Image Edit rather than replacing it. The idea is to make pose and composition control easier by feeding the model more structured guidance.
structure
The setup combines the base model with pose-oriented controls so you can keep the subject’s structure while changing the angle or body position.
when to use it
It is a good fit when you want to preserve the character but change the shot, such as a different camera angle, a rotated body, or a more deliberate pose.
required specs
It is still a local image-generation workflow, so memory pressure is real. The heavier the model, the more likely you are to hit VRAM limits.
setup gotchas
The usual issues apply: model version mismatch, missing components, and the need to make sure the pose controls line up with the exact Qwen base you are running.
what the official 2511 model improves
The official 2511 base is better at the overall balance between prompt adherence and edit consistency. It is not a magic “pose fix,” but it is clearly a better foundation than the earlier builds.
how it differs from older i2i
Compared with ordinary image-to-image workflows, the Qwen Image Edit line is trying to preserve identity while allowing more structured edits. That is why the model choice matters so much.
how to run 2511 on an M1 Max
qwen-image-mps
One path is the qwen-image-mps route. It is the most direct Apple Silicon option, but it is still slow and memory-sensitive.
ComfyUI + GGUF
Another path is ComfyUI with GGUF quantization. That is more practical if you want to fit the model into limited memory, but it still is not a lightweight workflow.
what I would try next
The most practical next step is to compare:
- the official 2511 base
- VNCCS Pose Studio
- the Nunchaku quantized build
and see which one gives the best balance of character preservation and pose control for your hardware.
references
- Qwen Image Edit
- VNCCS Pose Studio
- ComfyUI