
Testing Setup:
- Base Model:
black-forest-labs/FLUX.1-dev - LoRA Model:
prithivMLmods/Retro-Pixel-Flux-LoRA(scale=0.8) - ControlNet:
Shakker-Labs/FLUX.1-dev-ControlNet-Union-Pro - Prompt:
8 by 8 Retro Pixel Art Character Sheet of NPC characters for pixel art game called Machi.
Tile Control Net Image:
Without control net:
With controlnet_conditioning_scale of 0.25:
With controlnet_conditioning_scale of 0.30:
From what I can tell, the control net is able to help with the layout.
With controlnet_conditioning_scale of 0.50:
Somewhere between 0.3 and 0.5, the control net becomes too strong. At that level, the image just becomes a mere copy with some dittering.
With controlnet_conditioning_scale of 0.35:
0.35 is still strong.
Now, let's keep the controlnet_conditioning_scale at 0.30, and variate the template image resolution. Above ones were done with 1024×1024 templates.
128×128 Template Image:
64×64 Template Image:
64×64 still works on providing the overall layout. There doesn’t seem to be much difference from 128×128.
32×32 Template Image:
At this resolution, the layout is still preserved.
16×16 Template Image:
At 16×16, the grid collapses and does not give much meaning.
Now, let’s fix the template resolution to 1024×1024 and variate zoom with nearest neighbor sampling.
lora_scalestill fixed at 0.8.controlnet_scalestill fixed at 0.3.
2x Zoom Template:
4 by 4 Retro Pixel Art Character Sheet of NPC characters for pixel art game called Machi.
Generated:
4x Zoom Template:
2 by 2 Retro Pixel Art Character Sheet of NPC characters for pixel art game called Machi.
Generated:
When using nearest neighbor downsampling, we get the following:
It does become blurry. We probably can find a better downsampling algorithm…
But, so far from the past generations, 4x zoom seems to be most stable.
To be continued…
Comments
Join the discussionNo comments yet. Be the first to comment.



