# 02 — FLUX.1 Schnell Image Pipeline

## Why FLUX over SDXL

FLUX is a 12B-parameter transformer model. SDXL (RealVisXL) is 3.5B.
FLUX has significantly better:
- Spatial depth and perspective (lens simulation)
- Scene geometry (vanishing points, depth-of-field)
- Prompt following (T5-XXL understands long, detailed prompts)

SDXL was tested on lahrcarpetcleaning.com and rejected: flat angles, no depth,
poor spatial coherence. FLUX replaced it entirely.

## Model stack

| File | Size | Notes |
|---|---|---|
| flux1-schnell-Q8_0.gguf | 12GB | GGUF Q8, needs ComfyUI-GGUF node |
| t5xxl_fp8_e4m3fn.safetensors | 4.6GB | T5-XXL text encoder, fp8 quantized |
| clip_l.safetensors | 235MB | CLIP-L, short prompt encoder |
| ae.safetensors | 108MB | Official FLUX VAE from Black Forest Labs |

## Download (one-time)

FLUX GGUF (public, no auth):
```bash
wget "https://huggingface.co/city96/FLUX.1-schnell-gguf/resolve/main/flux1-schnell-Q8_0.gguf" \
  -O ~/ComfyUI/models/unet/flux1-schnell-Q8_0.gguf

wget "https://huggingface.co/comfyanonymous/flux_text_encoders/resolve/main/t5xxl_fp8_e4m3fn.safetensors" \
  -O ~/ComfyUI/models/clip/t5xxl_fp8_e4m3fn.safetensors

wget "https://huggingface.co/comfyanonymous/flux_text_encoders/resolve/main/clip_l.safetensors" \
  -O ~/ComfyUI/models/clip/clip_l.safetensors
```

FLUX VAE (gated — requires HF login and license acceptance):
```bash
hf auth login   # paste read token
HF_TOKEN=$(cat ~/.cache/huggingface/token)
wget --header="Authorization: Bearer $HF_TOKEN" \
  "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/ae.safetensors" \
  -O ~/ComfyUI/models/vae/ae.safetensors
```

## ComfyUI workflow (what gen-images-flux.py sends)

```
UnetLoaderGGUF    → flux1-schnell-Q8_0.gguf
DualCLIPLoader    → t5xxl_fp8_e4m3fn + clip_l (type=flux)
VAELoader         → ae.safetensors
CLIPTextEncode    → prompt
EmptyLatentImage  → 1024×576, batch=1
KSampler          → steps=4, cfg=1.0, euler, simple
VAEDecode
SaveImage
```

## Settings

| Setting | Value | Why |
|---|---|---|
| Steps | 4 | Schnell is distilled — 4 steps is optimal |
| CFG | 1.0 | Distilled model, higher CFG degrades quality |
| Sampler | euler | Best for FLUX |
| Scheduler | simple | Matches FLUX training |
| Negative prompt | none | Distilled model ignores it |
| Resolution | 1024×576 | 16:9 hero format |

## Running generation

```bash
# ComfyUI must be running first (see 01-comfyui-setup.md)
cd /home/sirdrez/arisingmedia-websites/{domain}
python3 tools/gen-images-flux.py 2>&1 | tee tools/flux-gen.log
```

Monitor:
```bash
tmux attach -t comfyui     # step progress bars
tail -f tools/flux-gen.log  # per-image OK/FAIL
```

Speed: ~4 min/image on CPU (2GB VRAM insufficient for GPU). 28 images = ~1h50m.

## After generation

```bash
python3 tools/convert-to-webp.py          # resize + convert to WebP
rm assets/images/**/*.jpg                  # delete source JPGs
docker compose build --no-cache web        # bake WebP into image
docker compose up -d
```

Verify:
```bash
curl -s -o /dev/null -w "%{http_code}" http://localhost:{port}/assets/images/hero/hero-carpet-cleaning.webp
# must return 200
```