Lahr Carpet Cleaning — Image Generation Pipeline

Model Stack

Prompt
gen-images-flux.py
28 images (16 hero + 12 svc)
→
API
ComfyUI
localhost:8188
→
UNet (model)
FLUX.1 Schnell
Q8_0 GGUF · 12GB · 12B params
→
Sampler
KSampler
4 steps · euler · cfg=1.0
→
Decode
FLUX AE
ae.safetensors · 108MB
→
Output
JPEG → WebP
1024×576 · q92 → q80

Text Encoders

CLIP-L
clip_l.safetensors
235MB · short prompts
+
T5-XXL fp8
t5xxl_fp8_e4m3fn
4.6GB · long prompt understanding
→
Node
DualCLIPLoader
type: flux

Hardware

GPU
AMD Radeon (2GB VRAM)
Execution
CPU only (VRAM too small)
Speed
~4 min / image
Total ETA
~1h50m for 28 images

Model Files on Disk

UNet
flux1-schnell-Q8_0.gguf · 12GB
T5-XXL
t5xxl_fp8_e4m3fn.safetensors · 4.6GB
CLIP-L
clip_l.safetensors · 235MB
VAE
ae.safetensors · 108MB (official BFL)

Generation Progress

4 / 28
14% — reload page to update

Prompt Strategy

Low-angle perspective (35mm / 24mm lens specified in prompt)
Carpet/floor texture sharp in foreground — subject recedes into bokeh
Shallow depth of field + vanishing point for depth cues
No people, no machines, no equipment
Finger Lakes / upstate NY context for residential scenes

Previous model: RealVisXL V5.0 fp16 (SDXL 3.5B) — rejected: flat angles, poor depth
Current model: FLUX.1 Schnell (12B transformer) — better spatial understanding