refactor: img-cond->img_uncond

[stduhpf]

Summary

Re-implement some changes that were dropped from #929

• Optimizes inference speed with --img-cfg-scale 1
• Fixes inconsistency with the way kontext refs and concat refs are handled

Before this change the conditionning were:

• cond: positive prompt, concat, kontext refs
• uncond: negative prompt, no concat, kontext refs
• img_cond: negative prompt, concat, no kontext refs

With this PR, they become:

• cond: positive prompt, concat, kontext refs
• uncond: negative prompt, concat, kontext refs
• img_uncond: negative prompt, no concat, no kontext refs

In a way it's just swapping the roles of uncond and img_(un)cond for inpaint / unet edit models.

Related Issue / Discussion

N/A

Additional Information

Checklist

• I have read and confirmed this PR follows the contribution guidelines.

[stduhpf]

Alternatively I could just revert the name changes and just include the optimisations, but this naming makes more sense to me.

[leejet]

I think there are some issues with the PR implementation, which may degrade editing quality.

.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-base-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 4.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu

PR:

Master:

[stduhpf]

I think there are some issues with the PR implementation, which may degrade editing quality.

.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux-2-klein-base-4b.safetensors --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'klein.cpp'" --cfg-scale 4.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu

PR:

Master:

What if you set --img-cfg-scale 1?

[leejet]

I don’t think this PR should change the default behavior.

[stduhpf]

@leejet I think the default img cfg scale should always be 1.

I know it has been set to = txt-cfg for a while, but this is not right. As I shown in this comment: #929 (comment), or in your exemple, higher img cfg scale can cause issues at a higher degree than higher txt cfg. (apg can help handle it but that's besides the point)

The results with edit models are changed from current Master, because there is a discrepency between the way kontext-like ref images and concat latents are handled with img cfg (forgot to mention that in the PR summary)

Master:

• uncond: negative prompt, no concat, kontext refs
• img_cond: negative prompt, concat, no kontext refs

PR:

• uncond: negative prompt, concat, kontext refs
• img_uncond: negative prompt, no concat, no kontext refs

[leejet]

I’ve set the default image CFG to 1.f, and everything looks OK now. Thanks for your contribution.