Skip to content

image_generation.md

Latest commit

 

History

History
137 lines (104 loc) · 3.5 KB

image_generation.md

File metadata and controls

137 lines (104 loc) · 3.5 KB

Image Generation

Cake supports image generation with Stable Diffusion and FLUX models.

Supported Models

Model Architecture VRAM Resolution Steps Quality
Stable Diffusion 1.5 SD ~4 GB 512x512 20-50 Good
Stable Diffusion 2.1 SD ~5 GB 768x768 20-50 Good
Stable Diffusion XL SD ~7 GB 1024x1024 20-50 High
SDXL Turbo SD ~7 GB 512x512 1-4 Good (fast)
FLUX.2-klein-4B FLUX ~8 GB 512x512 4 Good (fast)
FLUX.1-dev (FP8) FLUX ~12 GB up to 1024x1024 20 Excellent

FLUX

FLUX.1-dev (FP8)

High-quality 12B parameter flow-matching transformer. Runs in FP8 precision on a single GPU with 16GB+ VRAM. Models are downloaded automatically from HuggingFace.

cake run evilsocket/flux1-dev \
  "a photorealistic landscape at golden hour, dramatic clouds" \
  --model-type image-model --image-model-arch flux1 \
  --flux-height 768 --flux-width 1024 \
  --image-output landscape.png

FLUX.2-klein-4B

Faster 4B variant, 4 denoising steps, best at 512x512:

cake run black-forest-labs/FLUX.2-klein-4B \
  "a fluffy orange cat sitting on a wooden table" \
  --model-type image-model --image-model-arch flux

FLUX Arguments

Argument Default Description
--flux-height 1024 Image height in pixels
--flux-width 1024 Image width in pixels
--n-steps 20 Denoising steps (FLUX.2-klein uses 4)
--guidance-scale 3.5 CFG guidance scale
--image-output output.png Output file path (PNG)

Stable Diffusion

Single Node

cake run stabilityai/stable-diffusion-xl-base-1.0 \
  "An old man sitting on the chair at seaside" \
  --model-type image-model \
  --sd-version xl --sd-num-samples 1 --image-seed 2439383

Distributed Generation

Define the model parts in topology.yml:

gpu_worker:
  host: 192.168.1.2:10128
  description: NVIDIA RTX 4090 24GB
  layers:
  - unet

macbook:
  host: 192.168.1.3:10128
  description: Macbook M2
  layers:
  - clip
  - vae

Start worker and master:

# Worker
cake run /path/to/hf/cache \
  --name gpu_worker --model-type image-model \
  --topology topology.yml --address 0.0.0.0:10128

# Master with API
cake serve /path/to/hf/cache \
  --model-type image-model --topology topology.yml

API Endpoints

OpenAI-compatible (/v1/images/generations) — returns raw PNG by default:

curl http://master-ip:8080/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{"prompt": "An old man sitting on the chair at seaside"}' \
  -o seaside.png

For base64 JSON (OpenAI client compatibility), add "response_format": "b64_json".

Legacy (/api/v1/image):

curl http://master-ip:8080/api/v1/image \
  -H "Content-Type: application/json" \
  -d '{
    "image_args": {
      "prompt": "An old man sitting on the chair at seaside",
      "sd-num-samples": 1,
      "image-seed": 2439383
    }
}'

See the full REST API Reference for details.

SD Versions

Flag Value Model
v1-5 (default) Stable Diffusion 1.5
v2-1 Stable Diffusion 2.1
xl Stable Diffusion XL
turbo SDXL Turbo

SD Arguments

Argument Default Description
--sd-version v1-5 SD version to use
--sd-num-samples 1 Number of images to generate
--image-seed random Seed for reproducibility