Image Generation

UniLM.jl supports image generation via the OpenAI Images API using models like gpt-image-2.

Basic Usage

result = generate_image(
    "A watercolor painting of a friendly robot reading a Julia programming book",
    size="1024x1024",
    quality="medium"
)
println("Success: ", result isa ImageSuccess)
if result isa ImageSuccess
    imgs = image_data(result)
    println("Images: ", length(imgs))
    println("Base64 length: ", length(imgs[1]))
    save_image(imgs[1], joinpath(@__DIR__, "..", "assets", "generated_robot.png"))
    println("Image saved to assets/generated_robot.png")
else
    println("Images: 0")
    println("Image generation failed — see result for details")
end

The ImageGeneration Type

For full control, construct an ImageGeneration object:

ig = ImageGeneration(
    prompt="A minimalist logo for a Julia programming package",
    model="gpt-image-2",
    size="1024x1024",
    quality="high",
    background="transparent",
    output_format="png"
)
println("Model: ", ig.model)
println("Size: ", ig.size)
println("Quality: ", ig.quality)
println("\nRequest JSON:")
println(JSON.json(ig))

Model: gpt-image-2
Size: 1024x1024
Quality: high

Request JSON:
{"background":"transparent","model":"gpt-image-2","output_format":"png","prompt":"A minimalist logo for a Julia programming package","quality":"high","size":"1024x1024"}

Configuration Options

Multiple Images

Generate multiple images in a single request:

result = generate_image("A cute robot learning to program", n=3, size="1024x1024")

if result isa ImageSuccess
    imgs = image_data(result)
    for (i, img) in enumerate(imgs)
        save_image(img, "robot_$i.png")
    end
end

Transparent Backgrounds

Perfect for logos and icons:

result = generate_image(
    "A simple geometric icon of a butterfly",
    background="transparent",
    output_format="png",
    quality="high"
)

if result isa ImageSuccess
    save_image(image_data(result)[1], "butterfly_icon.png")
    # => PNG with transparent background
end

Result Structure

# Show the type hierarchy for image results
println("ImageSuccess <: ", supertype(ImageSuccess))
println("ImageFailure <: ", supertype(ImageFailure))
println("ImageCallError <: ", supertype(ImageCallError))

ImageSuccess <: LLMRequestResponse
ImageFailure <: LLMRequestResponse
ImageCallError <: LLMRequestResponse

result = generate_image("A sunset over mountains")

if result isa ImageSuccess
    r = result.response

    r.created                 # Unix timestamp
    r.data                    # Vector{ImageObject}
    r.data[1].b64_json        # base64-encoded image data
    r.data[1].revised_prompt  # revised prompt (may be nothing)
    r.usage                   # token usage Dict

    # Convenience accessors
    image_data(result)        # Vector{String} of base64 data
    save_image(image_data(result)[1], "sunset.png")
end

Saving Images

The save_image helper decodes base64 and writes to disk:

# Demonstrate save_image with a tiny test payload
tmpfile = tempname() * ".txt"
UniLM.save_image("aGVsbG8=", tmpfile)  # "hello" in base64
println("File saved to: ", basename(tmpfile))
println("Contents: ", read(tmpfile, String))
rm(tmpfile)

File saved to: jl_Uvwz8PErG9.txt
Contents: hello

Error Handling

result = generate_image("A sunset over mountains")

if result isa ImageSuccess
    save_image(image_data(result)[1], "sunset.png")
elseif result isa ImageFailure
    @warn "HTTP $(result.status): $(result.response)"
elseif result isa ImageCallError
    @error "Call failed: $(result.error)"
end

Editing Images

edit_image edits an existing image from a text prompt, optionally masked to an inpainting region. The convenience form takes the source image (a file path, or a vector of paths) and the prompt; an optional mask path restricts edits to the mask's transparent area:

# Inpaint: change only the masked region, described by the prompt
result = edit_image("room.png", "Add a large window with a sea view"; mask="room_mask.png")

if result isa ImageSuccess
    save_image(image_data(result)[1], "room_edited.png")
else
    println("Edit failed — see result for details")
end

For full control, build an ImageEdit and call edit_image(e). The model defaults to gpt-image-2; image editing needs the :image_edits capability (OpenAI), and other providers reject it at request time.

Retry Behaviour

generate_image retries transient HTTP statuses (408, 429, 500, 502, 503, 504, 529) within the request budget configured by RequestConfig (max_attempts, default 3), with exponential backoff and jitter; on 429 the Retry-After header is respected. Override per call with config=RequestConfig(...).

See Also

• ImageGeneration — request configuration type
• ImageResponse — response type
• generate_image — request function
• image_data, save_image — accessor/utility functions