multi-model-image-gen 0.2.2

Unified Python library for image generation across OpenAI, fal.ai, Google (AI Studio + Vertex AI), and Higgsfield — aisuite-style addressing, retry + fallback, async-first, plugin-extensible.

multi-model-image-gen

Unified Python library for image generation across OpenAI, fal.ai, Google (AI Studio + Vertex AI), and Higgsfield. One API, aisuite-style addressing, retry + fallback, async-first, plugin-extensible.

from image_gen import generate

result = generate(
    prompt="a neon-lit Tokyo alley at night",
    model="fal:flux/dev",
    fallbacks=["openai", "higgsfield"],
    output="out.png",
)

Supported models

Catalog of models known to the library (current as of 2026-04-17). Shorthands resolve to the provider-native model ID automatically. Any full model ID also works — unknown models just pass through without capability validation.

Single source of truth: src/image_gen/providers/catalog.yaml. Add or update a model = edit that file, no Python changes.

OpenAI — GPT Image family

Shorthand	Model ID	Seed	Neg. prompt	Notes
gpt-image-1.5	gpt-image-1.5	–	–	Current SOTA — natively multimodal (text + image I/O).
gpt-image-mini	gpt-image-1-mini	–	–	Cheapest tier when quality is secondary.
gpt-image	gpt-image-1	–	–	Legacy. Kept for back-compat.

Pricing is token-based (text + image tokens) — see OpenAI's pricing page for current rates. The image API also supports dall-e-2 / dall-e-3 via full model IDs.

fal.ai — curated subset of 1000+ models

Any fal model works by full ID (generate(model="fal:fal-ai/sdxl/lightning", …)). The shorthands below cover the most common ones:

Shorthand	Model ID	Seed	Neg. prompt	Cost / image
flux2-pro	fal-ai/flux-2-pro	✓	–	~$0.03 / MP
flux2-dev-turbo	fal-ai/flux-2-dev-turbo	✓	–	~$0.008
flux-pro	fal-ai/flux-pro	✓	–	$0.050
flux	fal-ai/flux/dev	✓	–	$0.025
flux-schnell	fal-ai/flux/schnell	✓	–	$0.003
seedream / seedream-4.5	fal-ai/bytedance/seedream/v4.5/text-to-image	✓	✓	$0.040
ideogram	fal-ai/ideogram/v3	✓	✓	$0.030–0.090
recraft	fal-ai/recraft-v3	✓	–	$0.040 raster / $0.080 vector
nb2-fal	fal-ai/nano-banana-2	–	–	$0.080
nb-pro-fal	fal-ai/nano-banana-pro	–	–	$0.150
imagen4-fal	fal-ai/imagen4/preview	–	–	variable

fal acts as a unified gateway — you can use Google's Nano Banana and Imagen through fal's billing + API instead of Google's own credentials, which is occasionally useful for multi-tenant apps.

Google — Nano Banana + Imagen (AI Studio or Vertex AI)

Two providers, each with independent auth:

Provider	Env var	Auth method	Use case
google	GOOGLE_VERTEX_API_KEY	Vertex AI (API key)	Railway / production
google	GOOGLE_CLOUD_PROJECT	Vertex AI (ADC)	Local dev
google-studio	GOOGLE_GEMINI_API_KEY	AI Studio	Gemini-only, no GCP project
google	GOOGLE_API_KEY + GOOGLE_GENAI_USE_VERTEXAI=True	Vertex AI (API key)	Legacy / compat
google	GOOGLE_API_KEY alone	AI Studio	Legacy / compat

GOOGLE_VERTEX_API_KEY takes priority for the google provider — set it to use a Vertex-bound API key instead of ADC. Imagen models require Vertex. google-studio shorthands (nb2, nano-banana, nb-pro) also work via google-studio:model-id if you want them on AI Studio billing.

Nano Banana (Gemini Image) — three tiers, shared generate_content API surface:

Shorthand	Model ID	Aliases	Description
nb-pro	gemini-3-pro-image-preview	Nano Banana Pro	Best reasoning ("Thinking"), high-fidelity text, most expensive.
nb2	gemini-3.1-flash-image-preview	Nano Banana 2	Fast, high-volume. Trades some compositional depth for speed.
nano-banana	gemini-2.5-flash-image	original Nano Banana	Legacy speed/efficiency tier.

Imagen 4 — three tiers:

Shorthand	Model ID	Notes
imagen4-ultra	imagen-4.0-ultra-generate-001	Highest fidelity; supports 1K + 2K + 4K.
imagen4	imagen-4.0-generate-001	Default Imagen 4.
imagen4-fast	imagen-4.0-fast-generate-001	Cheapest / fastest.

All Google image models output includes a SynthID watermark.

Higgsfield — async-task platform (submit → poll → download)

Polling is built in via poll_until. Higgsfield also gives day-0 access to third-party video models.

Soul (photorealistic image generation):

Shorthand	Model ID
higgsfield-soul	higgsfield-ai/soul/standard
higgsfield-soul-pro	higgsfield-ai/soul/pro

DoP (image-to-video, three quality/speed tiers):

Shorthand	Model ID
higgsfield-dop	higgsfield-ai/dop/lite
higgsfield-dop-preview	higgsfield-ai/dop/preview
higgsfield-dop-turbo	higgsfield-ai/dop/turbo

Kling (video via Higgsfield platform):

Shorthand	Model ID	Notes
higgsfield-kling	kling-video/v3.0/master/image-to-video	Kling 3.0 — unified video + audio + images, multi-shot.
higgsfield-kling-v2	kling-video/v2.1/pro/image-to-video	Kling 2.1 Master.

Other Higgsfield platform models (Sora 2, WAN 2.5, MiniMax Hailuo 02, Seedance Pro, etc.) work by passing their full ID: generate(model="higgsfield:sora-2/preview", …).

All models support

Aspect ratios 9:16, 16:9, 1:1. Resolutions 720p, 1080p. Capabilities beyond these are rejected pre-flight with UnsupportedCapabilityError — the SDK call is never made. Add a model, or update capabilities, by editing src/image_gen/providers/catalog.yaml — no Python changes needed.

Install

pip install git+https://github.com/oj-rivas/multi-model-image-gen.git
# or pinned:
pip install git+https://github.com/oj-rivas/multi-model-image-gen.git@v0.2.0
# or editable from local checkout:
pip install -e /path/to/multi-model-image-gen
# with Prometheus metrics:
pip install "multi-model-image-gen[metrics] @ git+…"

Credentials

Set only the keys for providers you'll actually call:

# OpenAI
OPENAI_API_KEY=sk-…

# fal.ai
FAL_KEY=…

# Google — Vertex AI (production / Railway)
GOOGLE_VERTEX_API_KEY=…                   # "google" provider → Vertex with API key
GOOGLE_GEMINI_API_KEY=…                   # "google-studio" provider → AI Studio
GOOGLE_CLOUD_PROJECT=your-gcp-project     # required for Vertex (Imagen 4)
GOOGLE_CLOUD_LOCATION=us-central1         # default
# Google — local dev (ADC)
# GOOGLE_CLOUD_PROJECT=your-gcp-project   # Vertex via gcloud auth application-default login
# Google — legacy single key
# GOOGLE_API_KEY=…                        # AI Studio, or Vertex with GOOGLE_GENAI_USE_VERTEXAI=True

# Higgsfield
HIGGSFIELD_API_KEY=…
HIGGSFIELD_API_SECRET=…
HIGGSFIELD_BASE_URL=https://platform.higgsfield.ai  # default

The library does not load .env for you. Use python-dotenv or your framework's loader.

Addressing

Three ways to name a model:

Form	Example	Provider chosen by
provider:model	"fal:flux/dev"	explicit prefix
Bare shorthand	"flux"	catalog
Bare full model ID	"higgsfield-ai/soul/standard"	catalog → higgsfield default

Unknown provider prefix raises ValueError. Unknown bare models route to higgsfield (which uses full IDs rather than shorthands).

Library usage

Sync

from image_gen import generate

result = generate(
    prompt="a neon fox",
    model="flux",                 # catalog → fal-ai/flux/dev
    output="fox.png",             # written when status == "completed"
    aspect_ratio="9:16",
    resolution="720p",
)
assert result.status == "completed"
print(result.request_id, result.cost_usd)

Async — for FastAPI / high-concurrency

import asyncio
from image_gen import generate_async

async def batch():
    return await asyncio.gather(*[
        generate_async(prompt=f"scene {i}", model="flux")
        for i in range(10)
    ])

Fallback chains

result = generate(
    prompt="…",
    model="fal:flux/dev",
    fallbacks=["openai", "higgsfield"],   # tried in order on retry-exhausted failure
    correlation_id="trace-42",
)

Direct provider client (bypass routing)

from image_gen import FalImageClient

with FalImageClient() as client:
    result = client.generate_image(
        prompt="cinematic forest",
        model="fal-ai/flux/dev",
        aspect_ratio="16:9",
        num_inference_steps=40,      # provider-specific kwarg
    )
    client.download(result.image_url, "forest.png")

Result shape

@dataclass
class GenerationResult:
    request_id: str
    status: str                # "completed" | "failed" | "blocked"
    image_bytes: bytes | None  # inline payload (OpenAI, Google)
    image_url: str | None      # remote URL (fal, Higgsfield)
    video_url: str | None      # Higgsfield video models
    cost_usd: float | None     # populated from catalog when available
    raw: dict | None           # full provider response + correlation_id if set

Callers that pass output= to generate() don't need to touch this — bytes/URL are handled automatically.

CLI

image-gen -p "a neon fox" -m flux -o fox.png
image-gen -p "…" -m "openai:gpt-image-1" --aspect 16:9 --resolution 1080p -o wide.png
image-gen -p "…" -m "higgsfield:higgsfield-ai/soul/standard" -o h.png

FastAPI integration

from fastapi import FastAPI, HTTPException
from image_gen import generate_async, FallbackExhausted

app = FastAPI()

@app.post("/images")
async def gen(prompt: str, model: str = "flux", x_correlation_id: str | None = None):
    try:
        r = await generate_async(prompt=prompt, model=model, correlation_id=x_correlation_id)
    except FallbackExhausted as e:
        raise HTTPException(502, str(e))
    if r.status == "blocked":
        raise HTTPException(422, r.raw)
    return {"url": r.image_url, "cost": r.cost_usd, "id": r.request_id}

Resilience

• Retry: tenacity-backed exponential backoff on httpx.TimeoutException, ConnectError, ReadError, and HTTP 429/500/502/503/504. Configurable via IMAGE_GEN_RETRY_ATTEMPTS (default 3) and IMAGE_GEN_RETRY_MAX_WAIT (default 8s). Non-transient errors (ValueError, 4xx) are not retried.
• Fallback: Router iterates [primary, *fallbacks]; each provider's retries run first, then the router moves on. FallbackExhausted is raised if all fail, carrying the list of underlying exceptions.
• Blocked content: NSFW / policy refusals return status="blocked" instead of raising — never retried.

Observability

Every generate() / generate_async() call emits one structured log record image_gen.request with canonical fields:

request_id, provider, model, latency_ms, status, bytes_out,
cost_usd, retry_count, fallback_used, correlation_id

Set IMAGE_GEN_LOG_FORMAT=json for one-line JSON per request (Datadog / CloudWatch / Loki-friendly).

Install the optional [metrics] extra to expose Prometheus counters: image_gen_requests_total, image_gen_retries_total, image_gen_fallbacks_total, image_gen_cost_usd_total, and histogram image_gen_latency_seconds.

Extending — add a provider as a plugin

No fork needed. In your own package:

# runway_image_gen/__init__.py
from image_gen.providers.base import ImageProvider

class RunwayClient:
    def generate_image(self, prompt, model, aspect_ratio, resolution, **kw): ...
    def download(self, url, output_path): ...
    def close(self): ...

# pyproject.toml
[project.entry-points."image_gen.providers"]
runway = "runway_image_gen:RunwayClient"

Then pip install your-package and generate(model="runway:gen-3") just works — entry-point discovery picks it up.

Also works at runtime:

from image_gen import register_provider
register_provider("custom", MyClientFactory)
generate(model="custom:foo")

A complete example lives in examples/runway_plugin/.

Layout

src/image_gen/
├── __init__.py                 # generate(), generate_async(), Router re-export
├── cli.py                      # image-gen CLI
├── config.py                   # lazy env readers
├── router.py                   # Router, FallbackExhausted
├── observability.py            # StructuredLogger, JSONFormatter, correlation_id
├── metrics.py                  # Prometheus (optional, gated on import)
├── py.typed                    # PEP 561 marker
└── providers/
    ├── __init__.py             # registry + plugin discovery (register_provider, get_provider)
    ├── base.py                 # ImageProvider Protocol (sync + async)
    ├── result.py               # GenerationResult
    ├── catalog.yaml            # SINGLE SOURCE OF TRUTH for models
    ├── catalog.py              # ModelEntry, resolve, validate
    ├── _retry.py               # @retry_policy, is_retryable
    ├── _poll.py                # poll_until, poll_until_async
    ├── README.md               # "how to add an async-task provider"
    ├── openai_images.py
    ├── fal_images.py
    ├── google_images.py        # AI Studio + Vertex AI
    └── higgsfield.py
examples/
├── runway_plugin/              # third-party plugin example
└── benchmark_concurrent.py     # async concurrency demo

License

MIT