Pretty simple workflow for Krea 2 turbo, similar as my other workflows for Z-image, Flux 2 Klein and Qwen 1.
I had to do a downgrade of my node Save Image because I spent several days battling with the Comfyui registry scanner and couldn't find another solution. If a prompt is sent to KSampler via any transformation like my anyconcan node, but it won’t extract or save metadata. The solution: manually enter positive/negative in Save Image.
Description
v3.0:
Added TXT2IMG + depth mask mode.
Added Onlymasked inpaint mode.
Changed delimiter between trigger words and prompt.
Changed note for Style Rebalance.
Added note for TXT2IMG + depth mask mode.
Some other internal and visual changes.
Requaring DemonAlone-JS_Addon-ComfyUI v1.9.0 or newer.
v3.1:
Few minor fixes.
v3.2:
Some visual adjustments.
FAQ
Comments (2)
is it possible? in TXT2IMG + Depth mask, it just generating same ref image style with hairs also , like in ref image girl has long hairs and my girl lora has short hairs like high bun, but it generting same ref image can i control that? or not
@hdmtrails
Depth mask will be more important than LORA. However, you can still control its strength by increasing the power of your LORA. In practice, 'depth mask' is something between text-to-text and image-to-image around 0.65 denoise. An alternative approach involves using a sketch in IMG2IMG mode with a higher denoising level (around 0.7). Another option is to use an image and apply only-masked inpainting; for example, at civitai.red/images/138631434, there are roughly four or five iterations of inpainting (with partial overlap in only one instance where Alice appears with the LORA effect). Krea2 is generally considered the worst modern model for editing or post-processing. Z-Image and Flux2Klein often prove more useful under these circumstances. However, neither was designed primarily for editing.
In short: if you want a complex image, you’ll likely encounter significant suffering; having a good prompt and reference doesn’t guarantee a great final result.