search

aredden / flux-fp8-api

Flux diffusion model implementation using quantized fp8 matmul & remaining layers use faster half precision accumulate, which is ~2x faster on consumer devices.
Apache License 2.0
209 stars 22 forks source link

Certain lora not applied correctly. #36

Open fyepi opened 1 week ago

fyepi commented 1 week ago

The lora is :

https://civitai.com/models/819754/iced-out-diamonds-by-chronoknight-flux

The image generated with flux-fp8-api: image

with comfy and flux1-dev: image

The prompt is: coca cola can

aredden commented 1 week ago

Ah- This is probably the result of the fused qkv lora not being applied correctly. This is actually good, since I can use it to test whether my new implementation is correct. So thank you 😆.