Link to your sampler custom node ?

Why post it without links? It won't work without the custom node, which doesn't show up in search results.

Hey, I tested your workflow I installed your custom sampler and it runs perfectly, but internally I'm still getting pretty catastrophic results. I'm starting to think your great results come from your custom checkpoint rather than the workflow itself, since that file isn't publicly available

Thanks. Any hints about your prompt generation workflow? It feels that you have some interesting features.

I tried following your workflow as closely as possible, but my results still look pretty bad. The overall quality is poor, and the details still have a very obvious AI-generated look. I’m trying to achieve the same kind of genuinely realistic, photographic results that you get, but I’m starting to think the main issue might be the model/checkpoint I’m using

hey, any update on the public checkpoints? it really seems that your checkpoint is the goldmine of it all!

What's your method/prompt for converting images (i assume from pinterest) to krea 2 prompts?

A bad share—just like those images posted without the workflow or LoRA. Here, you've got a workflow with custom nodes but no Git. I have zero interest in following this kind of content.

Works great! Do you use more loras than those that u listed here? Seems like. Could u share? Also did u make that checkpoint yourself or is it a lora merge?

Thanks for sharing it. The scheduler makes a big difference here, you can get closer results with the following manual sigmas:

1.000000, 0.967357, 0.922307, 0.862136, 0.785192, 0.692099, 0.586719, 0.476054, 0.368630, 0.272025, 0.190941, 0.126735, 0.078197, 0.042776, 0.017586, 0.000000

However, you left the 2 most important things out of the workflow: your base model and how you generate those long prompts. Would appreciate it if you could share which LoRAs you used to merge your model, and which system prompt (and LLM) to get the prompts.

Cheers!

I appreciate that you shared the workflow, but calling this a reproducible workflow feels a bit misleading.

You literally state that the checkpoint/diffusion model is the most important part, yet the workflow uses your private MrTeez_MainV2.safetensors, which nobody else has access to. You also don’t explain what RAW Krea 2 model it is based on, whether it is a merge, or what was merged into it. So the single most important component of the workflow is missing.

The same goes for the prompts. You say your prompts are often 2000–3000 words long, but you don’t share how you generate them, which LLM you use, or the system prompt. That is another major part of the result that users are left to guess.

There is also a strange scheduler inconsistency. The workflow you shared uses flow_linear, but the manual sigma schedule posted in the comments matches flow_sigmoid with 16 steps, midpoint 0.4, sharpness 6 and sigma shift 1.25 exactly. So people are basically having to reverse-engineer the missing details just to get closer to your examples.

At that point, what has really been shared is mostly the node layout and some LoRA settings, while the parts that seem to matter most for reproducing the actual images are still private or undocumented.

If the example images really were made with this exact public workflow, it would be helpful to clarify the base model, the prompt-generation method, and whether the examples used flow_linear or flow_sigmoid. Otherwise this is useful as inspiration, but not really as a reproducible workflow