Z-Image Base

This version was trained on Base, but looks terrible when used with Base. Use it with Turbo at a strength of around 2 and it looks great. The sample images were created with Turbo, 9 steps, Res_multistep/simple. Euler_ancestral and FlowMatchEulerDiscreteScheduler are a nice alternative if you have it.

Z-Image Turbo

Use at a strength of 0.8 to minimize mutations. I'm using the KSamplerWithNag node in Comfyui so I can use "Pixelated" as a negative. For some reason Z-Image Turbo thinks "low poly" means "pixelated".

ComfyUI workflow is in the images, as always, if you want to see my other settings.

Wan 2.1/2.2

Trained on 64 still images, so I'm not sure how stable it is for video. All sample images have their ComfyUI workflows embedded, so feel free to grab those for the settings. Otherwise, here were my general settings:

• LoRa strength: 1.0

• Euler/Beta

• 18 steps

• CFG: 2.5

• 1024x1344

There are a few other nodes for things like NAG in the workflow if you want to take a peek. This is my daily driver workflow for Wan images right now and I'm super happy with the quality.

Flux.1-Dev

This model creates images of "handmade" low poly papercraft scenes, with little perforations at many of the folds.

The trigger is Low poly papercraft scene

The default look is a bit muted, but the model responds well to "dramatic lighting" and "vibrant" if you want a brighter, more contrasty/saturated scenes. "Dappled sunlight" also gives a fun effect.

Settings for sample images:

• Flux1-dev.safetensors

• Guidance 3.0

• Sampler: Euler

• Scheduler: Normal

• Steps: 24

• LoRA Strength: 1.0