I used a merge of 2 LoRAs to make the colors more contrasting and add a bit of detail.
Resources for launch:
Text Encoders:
VAE:
I recommend:
resolution: 1- 2 MP
sampler: euler
scheduler: simple
steps 40 - 50
cfg 3
Positive prompt: best quality, highres,
Negative prompt: worst quality, bad anatomy, watermark, jpeg artifact
Best for anime characters.
Description
Format: int8_tensorwise + ConvRot (comfy-kitchen, groupsize 256)
Quantized: only 42×4 block Linears (proj.qkv / proj.o / proj.ffn_up_gate / proj.ffn_down)
Kept bf16: norms, modulations, embedders, final_linear_video, rope
Verified on ComfyUI 0.43.0, RTX 50 series:
- size 5.41 GB (bf16: 9.47 GB)
- mean relative weight error 0.876 %, max 0.980 %
- VRAM staged 5162 MB vs 9693 MB (bf16)
- ≈2.5x faster than bf16 at 1024×1024
Loader: Load Diffusion Model → weight_dtype = default
Do NOT enable --fp8_e4m3fn-unet for this file.
VAE / text encoder are NOT included: use flux2-vae + qwen3.5_9b (int8 variant recommended for low VRAM).
Base model: Custom merge based on MOSS-Image-SFT (OpenMOSS-Team), Apache-2.0 — quantized derivative.
