Qwen Image 2.1 Turbo: bad final image but good preview images

[Flix01]

Hi everybody!
I'm not an expert in using stable-diffusion.cpp, so it's probably something on my side.

I'm using sd-server and the web-ui with:
-> qwen_image_2.1_turbo_int8_convrot.safetensors
-> taeqi2_1.safetensors (for preview images)
-> Qwen3-VL-8B-Instruct-UD-IQ3_XXS.gguf (unsloth)
-> every vae for it I could try (bf16 or f32: no changes)
-> other launch options: --steps 8 --cfg-scale 1.0 --sampling-method euler --sigmas "1.0,0.978453,0.95418,0.926626,0.89508,0.845148,0.704534,0.414568,0.0" -v --diffusion-fa --mmap -H 512 -W 512 --cache-mode dbcache --cache-option "threshold=0.125,warmup=5" --offload-to-cpu --max-vram -0.01 --vae-tiling --color

Problem is that when I generate images, the (taesd) preview looks better than the final image.

Example: "A lovely cat"

Preview [4 step/8 steps] taeqi2_1

Final [8 step/8steps] vae

Can this be fixed?

Thank you in advance.

P.S. stable-diffusion.cpp is amazing!

[Flix01]

Well, now I'm explicitly setting: --scheduler discrete in the command line (not sure it's the default or not) and it seems to improve things a bit.

But I've discovered that the model I was using (qwen_image_2.1_turbo_int8_convrot.safetensors by Comfy-Org), despite being 7.3 GB big, it's less precise than qwen-image-2.1-turbo-Q5_K_S.gguf by Unsloth (4.5 GB; but this version, despite being smaller, it's 2 or 3 times slower than the other on my system).

Take a look at the eyes of these cats and see the difference:

qwen_image_2.1_turbo_int8_convrot.safetensors (Comfy-Org) 7.3 GB

qwen-image-2.1-turbo-Q5_K_S.gguf (Unsloth) 4.5 GB

So I think the issue in the first post can be mitigated by using more precise models.

P.S. I've discovered that there's also an Unsloth version of Qwen-Image-2.1-Turbo-INT8-ConvRot.safetensors, but I'm not sure how much it can different from the other...