Flux.1 Dev Lite 8B FP8 = 7.5 GB Pruned Model of Flux.1 Dev 12B

p0zp0z
15
485
base model
UNET v1.0
model-image-0
model-image-1
model-image-2
model-image-3
model-image-4
model-image-5
model-image-6
model-image-7
model-image-8
model-image-9
model-image-10
model-image-11
model-image-12
model-image-13
model-image-14
model-image-15
model-image-16
model-image-17
model-image-18
model-image-19

Details

Download Files (1)

Model description

This is a smaller version 8B (better for 6 to 12 GB VRAM) parameter self-merge of the 12B Flux.1-dev model into the new Flux Lite causing a 90% pruned Dev version. FP8 quantized from fp16.

I highly recommend using a brightener VAE like https://civitai.com/models/1546042/flash-photo-flux-vae

This is not an all-in-one (AIO) model, you still need t5xxl_fp8_e4m3fn_scaled.safetensors, clip_l.safetensors and flux_vae.safetensors.

I suggest using this model with ComfyUI portable (browse workflows) github.com/comfyanonymous/ComfyUI/releases/ (free open source, windows/macOS/linux app, unlimited/private/offline).

Sample prompts here: fb.com/groups/1712618889252391/user/100000357446008/

All my uploaded images can be drag/dropped to comfyui as embedded workflows :)

Images made by this model