genblaze-gmicloud 0.3.1

GMICloud media provider adapters for genblaze (video, image, audio)

genblaze-gmicloud

GMICloud multi-provider video / image / audio adapters for genblaze — access Seedance, Kling, Veo, Sora, Wan, Seedream, FLUX, Gemini image, ElevenLabs, MiniMax and more through one API with SHA-256 provenance manifests.

genblaze-gmicloud wraps GMICloud's request-queue API, giving you one-call access to a large catalog of video, image, and audio models — including Kling, Veo, Sora, Wan, Seedream, FLUX-Kontext-Pro, Gemini-2.5-Flash-Image, ElevenLabs TTS, MiniMax TTS, and MiniMax Music — via three genblaze provider classes. Compose into multi-step AI pipelines, persist outputs to Backblaze B2 or any S3-compatible store, and emit a tamper-evident provenance manifest on every run.

Why genblaze-gmicloud

• One API, dozens of models — text-to-video (Seedance, Kling, Veo, Sora, Wan), text-to-image (Seedream, FLUX, Gemini, Reve), audio (ElevenLabs, MiniMax TTS/Music).
• LLM access too — standalone chat() wrapper for Llama, DeepSeek, Qwen over GMICloud's OpenAI-compatible inference endpoint (see below).
• Provenance by default — SHA-256-verified manifest with provider, model, prompt, params, cost.
• Cost tracking — register a pricing strategy from docs/reference/pricing-recipes.md and step.cost_usd is populated automatically.
• Production-ready — retries, timeouts, progress streaming, step caching.
• Durable storage — plug genblaze-s3 in for Backblaze B2 / AWS S3 / R2 / MinIO persistence.

Providers + models

Provider class	Modality	Example models
GMICloudVideoProvider	video	Kling-Text2Video-V2.1-Master, Kling-Image2Video-V2.1-Master, Veo3, wan2.6-t2v, seedance-1-0-pro-250528, sora-2-pro
GMICloudImageProvider	image	seedream-5.0-lite, gemini-2.5-flash-image, reve-edit-fast-20251030, flux-kontext-pro
GMICloudAudioProvider	audio	elevenlabs-tts-v3, minimax-tts-speech-2.6-turbo, minimax-music-2.5

Registered via entry points as gmicloud, gmicloud-image, and gmicloud-audio. Any model on GMICloud's queue is supported — pass the exact model slug.

Slug casing — GMICloud's request queue is case-sensitive and uses per-slug casing (per their published catalog): lowercase for Sora / Pixverse / Seedance / Seedream / Reve / Wan / Bria / Gemini-Image, the audio families (elevenlabs-tts-*, minimax-tts-*, minimax-music-*, minimax-audio-voice-clone-*, inworld-tts-*), and the newer Kling V2.5 / V3 series; PascalCase for Kling V2.1 (Kling-Text2Video-V2.1-Master, Kling-Image2Video-V2.1-Master) and Veo3. As of genblaze-gmicloud 0.3.1, the connector rewrites caller-supplied casing to GMICloud's published wire form for the families above via the new ModelFamily.canonical_slug mechanism — so model="ElevenLabs-TTS-v3" and model="veo3" (legacy mixed/lower-case) keep working, with a one-time INFO per (family, input) nudging callers toward the canonical form. Slugs that don't match a known family still pass through verbatim; refer to console.gmicloud.ai as the source of truth when a model rotates.

Install

pip install genblaze-gmicloud

Quickstart — video (Kling)

pip install genblaze-core genblaze-gmicloud
export GMI_API_KEY="..."

from genblaze_core import Modality, Pipeline
from genblaze_gmicloud import GMICloudVideoProvider

run, manifest = (
    Pipeline("gmicloud-video-demo")
    .step(GMICloudVideoProvider(), model="Kling-Text2Video-V2.1-Master",
          prompt="A drone shot flying over a misty mountain valley at sunrise, cinematic",
          modality=Modality.VIDEO, duration=10, aspect_ratio="16:9")
    .run(timeout=600)
)
print(run.steps[0].assets[0].url)
# step.cost_usd is None unless you've registered pricing — see
# docs/reference/pricing-recipes.md for the GMICloud rate sheet.

Quickstart — image (Seedream)

from genblaze_gmicloud import GMICloudImageProvider

run, manifest = (
    Pipeline("gmicloud-image-demo")
    .step(GMICloudImageProvider(), model="seedream-5.0-lite",
          prompt="A photorealistic macro shot of morning dew on a spider web, soft bokeh",
          modality=Modality.IMAGE, aspect_ratio="16:9")
    .run(timeout=120)
)

Quickstart — audio (ElevenLabs via GMICloud)

from genblaze_gmicloud import GMICloudAudioProvider

run, manifest = (
    Pipeline("gmicloud-audio-demo")
    .step(GMICloudAudioProvider(), model="elevenlabs-tts-v3",
          prompt="Welcome to Genblaze — the fastest way to build generative AI pipelines.",
          modality=Modality.AUDIO)
    .run(timeout=120)
)

Persist to Backblaze B2

from genblaze_core import KeyStrategy, ObjectStorageSink
from genblaze_s3 import S3StorageBackend

storage = ObjectStorageSink(
    S3StorageBackend.for_backblaze("my-bucket"),
    key_strategy=KeyStrategy.HIERARCHICAL,
)
# pass sink=storage to .run(…)

Backblaze B2 is the recommended default sink — cost-efficient, S3-compatible, Object Lock for immutable manifests.

LLM access — standalone chat()

For callers driving a media pipeline from an LLM — caption expansion, prompt rewriting, scene description — genblaze-gmicloud ships a chat() callable over GMICloud's OpenAI-compatible inference endpoint. It sits outside the Pipeline / Step machinery (text generation doesn't benefit from the polling / manifest / asset machinery built for media).

from genblaze_gmicloud import chat

resp = chat("deepseek-ai/DeepSeek-V3", prompt="A cinematic sunset over Tokyo")
print(resp.text, resp.tokens_out)

Any GMICloud-hosted chat model is accepted — model ids pass through to the inference endpoint verbatim, so you can use models the connector hasn't been updated for. cost_usd is always None for this connector; compute cost from tokens_in / tokens_out yourself if needed.

Full signature and ChatResponse shape: docs/features/llm-calls.md.

Credentials

Only API-key auth is supported. Set GMI_API_KEY (obtain from https://console.gmicloud.ai/) or pass api_key= to any provider ctor or to chat().

Configuring the endpoint (staging, proxies, VPC)

All three provider classes and chat() accept a base_url= ctor kwarg (or GMI_BASE_URL env var) to override the default endpoint, and an http_client= kwarg for injecting a pre-built httpx.Client — useful for shared connection pools across multi-modality pipelines or for mocking in tests.

import httpx
from genblaze_gmicloud import GMICloudVideoProvider, GMICloudImageProvider

shared = httpx.Client(
    base_url="https://my-vpc-proxy.example/gmi",
    headers={"Authorization": f"Bearer {key}"},
    timeout=120,
)
video = GMICloudVideoProvider(http_client=shared)
image = GMICloudImageProvider(http_client=shared)
# Caller owns `shared` — providers never close externally-supplied clients.

Naming reference

GMICloud surfaces five related names; they look interchangeable but come from different namespaces:

Surface	Value
PyPI package	genblaze-gmicloud
Python import	import genblaze_gmicloud
Provider class prefix	GMICloud* (e.g. GMICloudVideoProvider)
Entry-point slug	gmicloud, gmicloud-image, gmicloud-audio
Env vars	GMI_API_KEY, GMI_BASE_URL

The GMI_ env prefix is short on purpose; the class / import / PyPI names use the full gmicloud for precision and to leave room for future genblaze-gmi* packages if needed.

Reading outputs safely

step.assets[0] is only valid when the step succeeded. Always check step.status first — especially in fan-out runs where one step may fail and others succeed:

for step in run.steps:
    if step.status == "succeeded" and step.assets:
        print(step.assets[0].url)
    elif step.status == "failed":
        print(f"failed ({step.error_code}): {step.error}")

Documentation

• Main repo: https://github.com/backblaze-labs/genblaze
• Examples: gmicloud_video_pipeline.py · gmicloud_image_pipeline.py · gmicloud_audio_pipeline.py

Related packages

• genblaze-core — the pipeline SDK
• genblaze-s3 — durable storage on Backblaze B2 and other S3-compatible backends

License

MIT