[Bug] Krea2 Turbo: cfg-scale 0 seems to ignore prompt conditioning

[knb]

Git commit

$git rev-parse HEAD
b5d8120

Operating System & Version

Ubuntu 24.04

GGML backends

Vulkan

Command-line arguments used

sd-cli --diffusion-model "$DIFFUSION" --llm "$LLM" --vae "$VAE" --prompt "A quiet Japanese mountain temple after rain, wet stone steps, soft morning mist, hydrangeas, natural diffused light, detailed documentary photography, restrained colors" --steps 8 --cfg-scale 1.0 --sampling-method euler --width 512 --height 512 --seed 42 --diffusion-fa --offload-to-cpu --output krea2-turbo.png --verbose

Steps to reproduce

1. Load Krea2 Turbo.
2. Generate an image using --cfg-scale 0.
3. Observe the warning:
   "unconditioned mode, images won't follow the prompt".
4. The prompt is ignored.
5. Repeat with --cfg-scale 1; prompt conditioning works as expected.

What you expected to happen

Krea2 Turbo should follow the prompt when using --cfg-scale 0.

According to the official Krea2 documentation, Turbo is intended to be used with:

• 8 sampling steps
• CFG = 0

Therefore, I expected prompt conditioning to work normally with --cfg-scale 0.

What actually happened

When --cfg-scale 0 is used, stable-diffusion.cpp switches to unconditional generation.

The following warning is printed:

unconditioned mode, images won't follow the prompt
(use cfg-scale=1 for distilled models)

The generated image largely ignores the prompt.

Changing only --cfg-scale from 0 to 1 makes prompt conditioning work correctly.

Logs / error messages / stack trace

Version:
stable-diffusion.cpp master-714
commit b5d8120

sample_params:
txt_cfg: 0.00
img_cfg: 0.00
distilled_guidance: 3.50

[WARN] unconditioned mode, images won't follow the prompt
(use cfg-scale=1 for distilled models)

Additional context / environment details

Environment:

• OS: Ubuntu 24.04
• Backend: Vulkan
• GPU: AMD Radeon 780M (RADV PHOENIX)

Models:

• Krea-2-Turbo-Q4_K_M.gguf
• Qwen3VL-4B-Instruct-Q4_K_M.gguf
• wan_2.1_vae.safetensors

Command options:

--steps 8
--sampling-method euler
--cfg-scale 0

The same command with --cfg-scale 1 follows the prompt correctly.

The prompt is successfully tokenized and encoded by the Qwen3VL text encoder, so it appears that conditioning is disabled only because cfg-scale == 0.

I'm not sure whether this is intentional or an implementation issue. If this behavior is expected, it might be worth documenting in docs/krea2.md, since the official Krea2 documentation recommends CFG=0 for Turbo.

[wbruna]

See a discussion about that warning on #1088 . Quoting stduhpf's comment:

The diffusers implementation uses CFG literally like it was proposed in the original paper (https://arxiv.org/abs/2207.12598v1), with CFG=0 meaning pure conditioned and CFG=-1 meaning unconditioned. It's effectively just an offset of -1 compared to every other implementation.

'Every other' including stable-diffusion.cpp: our CFG=1 is the same as Diffuser's CFG=0. If Krea2 docs take Diffusers as reference, it makes sense they would refer to 'pure conditioned' as CFG=0; and to get the same effect on stable-diffusion.cpp, you need CFG=1.

Note this isn't really specific to Krea2 Turbo: Z-Image Turbo docs also mention CFG=0, which was what motivated that PR.

[knb]

Thanks for the clear explanation!

I wasn't aware that stable-diffusion.cpp uses the "traditional" CFG convention, while the Krea2 documentation follows the Diffusers convention.

The mapping now makes perfect sense:

• Diffusers CFG = 0 <=> stable-diffusion.cpp CFG = 1
• Diffusers CFG = -1 <=> stable-diffusion.cpp CFG = 0 (unconditioned)

I confirmed that using --cfg-scale 1 produces the expected results with Krea2 Turbo.

This was a misunderstanding on my side rather than a bug in stable-diffusion.cpp, so I'll close this issue.

Thanks again for the detailed explanation!

It might be helpful to mention this CFG convention in docs/krea2.md, since users coming from the official Krea2 documentation may expect CFG=0 to work the same way.