Shea

...

Image Generation Suggestions

I generate these images via Mcmonkeyprojects' fantastic branch of the SwarmUI (a powerful but easy to use front end for ComfyUI and other backends). I find presently other UIs producing lacklustre results. YMMV.

Core Parameters

• Model: Flux1-dev with clip_l, t5xxl_fp16.

• Steps: 30

• CFG Scale: 1

Sampling

I've had success with a good amount of samplers and schedulers, but at present am using the following,

• Sampler: DEIS

• Scheduler: Beta

• Flux Guidance Scale: 3.5

Prompting

With SwarmUI use the <Segment:> feature to inpaint particular portions of the output. This can be nearly anything the segment model used was trained on, face, eyes, feet, top, etc.

Segmentation (spot in-painting)

A general statement I tag to most of these prompts is:

• <segment:object,0.7>prompt

Where object is the particular object in the image you'd like to inpaint and the number is the strength. SwarmUI explains additional variables. Use like a superior aDetailer. Yolo models can also be added to SwarmUI and used as well.

Refine/Upscale (Hi-Res Fix)

These settings can be used to get higher resolution outputs without the banding it produces. Starting with a higher resolution and using this as a Refiner mechanism is often the way to maintain detail fidelity. I generally stay between 1 and 1.25 upscales for this reason.

• Refiner Control Percentage: 0.3 - 0.7

• Refiner Method: Post-Apply

• Refiner upscale: 1 - 1.5

• Refiner Upscale Method: Latent: Bislerp

• Refiner Do Tiling: Checked (Important!)

Posted for civitai.com and tensor.art only

Each of these LoRAs take me dozens of hours and GPU usage out of pocket. Every review is appreciated and encourages me to do more of these.

-MoC