Images

Image generation runs on the GPU engine, so it needs a Mac. The models are written out in this repository: CLIP, T5, the UNet, the MMDiT, the VAEs and both schedulers. Only convolution and matrix multiplication come from a framework.

kvad pull stabilityai/stable-diffusion-xl-base-1.0
kvad images make "a red fox in fresh snow" \
    --model stabilityai/stable-diffusion-xl-base-1.0

The picture is written to image-ID.png, or to --out FILE. It is also kept on the server with the settings that made it, where the web UI's Images page and kvad images list it.

#Models

SDXL and Qwen-Image run the denoiser twice per step while guidance is on, and those figures include both passes. FLUX.1-schnell needs no guidance. FLUX.1-dev takes a guidance scale, 3.5 unless said, as a number the model reads: one pass a step, and no negative prompt. Its repository is gated and its licence non-commercial, so it needs a Hugging Face token for an account that has accepted the licence.

#Fine-tunes

SDXL and Stable Diffusion 1.5 fine-tunes load in any of the three layouts people publish: a diffusers folder, a single checkpoint file in Stability's own layout, or a file on your disk.

kvad pull Lykon/dreamshaper-8
kvad images make "a lighthouse at dusk" --model Lykon/dreamshaper-8
kvad images make "a lighthouse at dusk" --model ~/Downloads/some-sdxl-finetune.safetensors

A repository with several checkpoints is named as repo:file.safetensors.

stabilityai/sdxl-turbo draws in one step with no guidance, at 512², and those are what it is given when a request does not say: about 0.3 s of denoising on an M5 Pro, and 3 s to decode. Four steps are a little sharper.

kvad images make "a red fox in fresh snow" --model stabilityai/sdxl-turbo

#GGUF

Community GGUF files of Qwen-Image's and FLUX.1-schnell's transformers load by the repository and the quantisation together. The rest of the model comes from the base_model on the model card.

kvad pull city96/Qwen-Image-gguf:Q4_K_S
kvad images make "a red fox in fresh snow" --model city96/Qwen-Image-gguf:Q4_K_S

A Q4_K_S holds 9.1 GB less than q8 for Qwen-Image, and 5.5 GB less for FLUX, for a step that is 5 to 18% longer.

#Options

kvad images make PROMPT [--out FILE] [--model MODEL] [--size WxH] [--steps N]
                        [--guidance F] [--negative TEXT] [--seed N]
                        [--lora NAME[:SCALE]]...

Anything left out is the model's own default. --lora applies a LoRA for this picture only.

kvad images            # the gallery, as a list
kvad images rm 12

#From a picture

kvad images edit makes a picture from one you give it. The picture is noised part of the way and drawn over from there, so the prompt says what the result should be, not what to change: "a green apple on a wooden table", not "make the apple green". SDXL and Stable Diffusion 1.5 do this, and their fine-tunes.

kvad images edit "a green apple on a wooden table" --image apples.jpg --strength 0.5
kvad images edit "a blue ceramic bowl" --image apples.jpg --mask right-half.png

--strength is how far the picture is taken towards noise, above 0 and at most 1, and 0.75 unless said. Near 0 it comes back as it was; at 1 nothing of it is left. An edit runs that share of its steps.

With --mask only part of the picture is drawn anew, at strength 1 unless said, and the rest comes back exactly as it went in. The mask is a picture the same size: transparent where the model should draw, as OpenAI's masks are, or white on black if it has no transparency. The web UI's Images page lets you paint one on the picture.

Without --size, the result has the picture's own shape, at about as many pixels as the model's own size. The server reads the picture with ffmpeg, so it needs one installed.

These models were not trained to inpaint, so a mask's edge can show where what is drawn does not meet what was kept. Models trained to follow an editing instruction (FLUX.1 Kontext, Qwen-Image-Edit) are not here yet.

#From the API

POST /v1/images/generations is OpenAI's images endpoint, with the settings OpenAI has no field for beside its own.

curl http://127.0.0.1:5823/v1/images/generations \
  -H 'content-type: application/json' \
  -d '{
    "model": "stabilityai/stable-diffusion-xl-base-1.0",
    "prompt": "a red fox in fresh snow",
    "size": "1024x1024",
    "steps": 30,
    "guidance_scale": 5,
    "seed": 5,
    "response_format": "url"
  }'

With "stream": true the server sends an event per denoising step, a rough preview beside each when partial_images asks for one, and a completed event per image. See OpenAI-compatible API.

POST /v1/images/edits is OpenAI's edits endpoint: the same fields, with an image, an optional mask and strength, as a multipart form or as JSON with each file a data: URL.

curl http://127.0.0.1:5823/v1/images/edits \
  -F image=@apples.jpg \
  -F prompt="a green apple on a wooden table" \
  -F strength=0.5 \
  -F response_format=url

#Memory

The VAE decode is the most memory an image asks for: it is the only stage at full resolution. If a picture fails with a message about GPU memory, another model is probably loaded. kvad ps shows what is, and kvad unload frees it.

#The pictures on the front page

The four pictures on the front page were made with the commands below, on an M5 Pro, with SDXL's defaults: 1024², 30 steps, guidance 5. Each took 82 to 86 seconds to denoise and 9 to decode. The same seed draws the same picture.

M=stabilityai/stable-diffusion-xl-base-1.0
kvad images make "a lighthouse on a rocky Norwegian coast at dusk, calm sea, warm light in the lamp room, long exposure photograph" --model $M --seed 11
kvad images make "a red fox in fresh snow at the edge of a birch forest, early morning light, wildlife photograph" --model $M --seed 5
kvad images make "a wooden stave church in a green valley under low clouds, watercolour" --model $M --seed 3
kvad images make "an open book on a wooden desk by a window, a single candle, oil painting, Dutch golden age" --model $M --seed 21