This is the UPDATE version of the Rockwelly S1 model.

Feel free to test and use it.

The Update Version follows the same principles as Version 1 (previous)!

Reason for updating Project S1:

1: Many errors were found in areas such as eyes and hands.

2: Due to the model's sensitivity to prompts (if the prompt wasn't detailed enough), the images don't come out uniform.

3: Compatibility with multiple loras.

Version S1 v1 at1 is an update version of the project. The issues mentioned above have been fixed.

Recommendations:

1: Know how to describe your prompt (putting things like "looking down" and then putting "looking up" in the same prompt) will cause aberrations in your generation. Therefore, REVIEW YOUR PROMPTS.

2: Note that (in this model's case), you can "play" with the Denoise. You can keep it at (low ha: 0.75) up to its maximum (high: 1.0). Doing so (with the seed fixed) will give you different effects and sometimes different results.

SETTINGS:

The Ksampler settings haven't changed much. However, you can have more freedom with the settings and steps.

Minimum recommendations:

Steps: 22

CFG: 1.4

Sampler_Name: Euler_A

Scheduler: Karras

For different effects, you can use Sampler_name: Res_Multistep with Scheduler: Sgm_uniform; they also give good results (depending on what you plan to do).

Once again, I humbly ask that if you liked this model, please post your creations and feedback after downloading it. This will help me understand how this model can be improved.

And if you liked my work, follow me on social media. All the images I post on social media are generated from my models, whether test or finished versions.

Below is a summary of the first version with information about it.

Since this is my first model, there may be some flaws (nothing that is a problem and cannot be fixed with a refiner or a second attempt with KSampler or USDU (Ultimate Upscale)). Currently, my only priority has been trying to infuse the TensorArt TurboX LoRA 8-step [SD3.5 large] (https://civitai.com/models/1362210/tensorart-turbox-lora-8-steps-sd35-large) with one of my Checkpoint models and Qwen Image Lightning 4-8 steps

(https://civitai.com/models/1862507/qwen-image-lightning-workflow-4-8-steps).

The reason is clear:

"My graphics card is a GTX 1060/6GB," so you can see how much I struggle to generate good images.

That's why I've been working (slowly and carefully) on a checkpoint that allows me to balance speed and quality. ** IT'S NOT PERFECT YET (THE WAY I WANT IT), BUT IT'S ALREADY SURPRISED ME ENOUGH TO SHARE IT WITH YOU.

Please help with the testing and post your results. The current settings are:

Steps: 22

cfg: 1.0 / 1.5 (my recommendation is not to exceed this value if using EULER A)

Scheduler: Karras

I sometimes tested with Res_Multistep + Sgm_ufiform

Steps: 22 / 25

cfg: 1.0 / 1.5

The results were sometimes satisfactory.

This is because I tested with the simplest workflow possible.

Time to beg:

"If you liked this checkpoint and would like to support my work, please consider following me on social media and following me, as I will always be looking for ways to improve this project and develop new ones (even with the limited resources of a GTX 1060/6GB)."

Enviar feedback

Painéis laterais

Histórico

Salvas