Sogni Client for Python

An async Python SDK for image, video, audio, and LLM inference on the Sogni Supernet. It follows the public surface and wire protocol of the TypeScript sogni-client, while using Python naming conventions and async iterators.

The Python port is currently beta. Keep credentials in environment variables or your system keychain; never commit them to source control.

Official quickstart · Examples · Sogni API reference

Install

Install the latest beta directly from the official GitHub repository:

python -m pip install "sogni-client @ git+https://github.com/Sogni-AI/sogni-client-python.git@main"

For an editable source checkout:

git clone https://github.com/Sogni-AI/sogni-client-python.git
cd sogni-client-python
python -m pip install -e .

Python 3.10 or newer is required.

Create an image

import asyncio
import os

from sogni_client import SogniClient


async def main() -> None:
    async with await SogniClient.create(
        api_key=os.environ["SOGNI_API_KEY"],
        app_id="my-image-app",
        app_source="my-app",
    ) as sogni:
        project = await sogni.projects.create(
            type="image",
            model_id="krea2_turbo_fp8_scaled",
            positive_prompt="A tiny observatory above a sea of clouds",
            negative_prompt="text, watermark",
            number_of_media=1,
            width=1024,
            height=1024,
            steps=8,
        )
        print(await project.wait_for_completion())


asyncio.run(main())

Socket clients require a stable app_id. Generate it once per application installation and persist it across process restarts; do not generate a fresh UUID each time the application starts. REST-only clients can omit it by passing disable_socket=True.

The example uses Krea 2 Turbo (krea2_turbo_fp8_scaled) because it is the only model an account's free monthly render credits can be spent on over the API — every other model needs paid credits, so a brand-new key would otherwise fail on its first call. It is an 8-step model, hence steps=8.

Edit an image with Krea 2 Identity Edit

Pass one or two local reference images through context_images. For two-image edits, place the base scene first and the identity or detail reference second.

project = await sogni.projects.create(
    type="image",
    model_id="krea2_identity_edit_v1_2",
    positive_prompt=(
        "Change only the jacket to vivid sapphire blue. Preserve the exact "
        "facial identity, expression, framing, background, and lighting."
    ),
    number_of_media=1,
    width=1024,
    height=1024,
    steps=10,
    guidance=1,
    token_type="spark",
    context_images=["reference.png"],
)
print(await project.wait_for_completion(timeout=900))

The runnable example accepts one or two image paths and can also create a batch:

python examples/krea_identity_edit.py reference.png \
  --prompt "Change only the jacket to vivid sapphire blue; preserve identity."

python examples/krea_identity_edit.py scene.png identity.png \
  --prompt "Use the first image as the base scene and the second for identity." \
  --count 4

Generate speech with Qwen3-TTS

Qwen3-TTS exposes three audio model IDs: studio voices, voice cloning, and voice design. The prompt is the script to read aloud.

project = await sogni.projects.create(
    type="audio",
    model_id="qwen3_tts_1.7b_custom_voice_bf16",
    positive_prompt="Every render on the Supernet runs on somebody else's GPU.",
    number_of_media=1,
    speaker="serena",
    instruct="warm and unhurried, close to the mic",
    output_format="mp3",
)
print(await project.wait_for_completion())

Voice Clone uses qwen3_tts_1.7b_voice_clone_bf16 and requires a 3–30 second reference_audio clip. Supply reference_text with the exact words spoken in that clip whenever possible; the transcript is the strongest control on how closely the clone preserves the source voice and accent. Voice Design uses qwen3_tts_1.7b_voice_design_bf16 and requires instruct to describe the speaker to invent.

Upscale a video with FlashVSR

FLASHVSR_VIDEO_UPSCALE_MODEL_ID (flashvsr_v1.1_tiny_long_bf16) upscales one finished video to 1080p or 1440p on its short edge. It is promptless and separate from video generation: it keeps every source frame, the exact frame rate (including fractional rates such as 24000/1001), the full aspect ratio, and the original audio, and it never trims, crops, restyles, or interpolates.

Sources must be at most 768px on the short edge and about 1344×768 pixels overall (768×1344 in portrait), 1-60 fps at a constant frame rate, SDR, square pixels with rotation applied, and 100 MB or less. The client sets no frame-count or duration limit: the server enforces the maximum clip length and refuses a source that is too long with a clear error. The output is at most twice the source size, so 1080p needs a source short edge of at least 540px and 1440p at least 720px. You do not send the source's frame count, frame rate, or size: the server probes the upload and uses its verified values. frames, fps, width, and height are optional, and any you do send must match the source.

from sogni_client import FLASHVSR_VIDEO_UPSCALE_MODEL_ID

project = await sogni.projects.create(
    type="video",
    network="fast",
    model_id=FLASHVSR_VIDEO_UPSCALE_MODEL_ID,
    positive_prompt="",
    number_of_media=1,
    reference_video="clip.mp4",
    upscale_resolution=1440,  # or 1080: the output's short edge
)
print(await project.wait_for_completion())  # MP4 with the original audio

Three optional choices tune the render. detail_preference is "stable" (default, More Stable) or "sharper"; processing_speed is "stable" (default, More Stable) or "faster"; seed defaults to 0 for a repeatable result, and -1 asks for a random seed. Sharper, Faster and any seed other than 0 or -1 need a worker release that supports them; until one is connected, the server refuses those requests.

To show a price first, call estimate_video_cost() with the output width and height (the source scaled so its short edge equals the target, both edges rounded to even pixels), the source's frames and fps, steps=1, and source_width/source_height; the job itself is charged from the verified source.

MiniMax H3 two-stage output (720p, 1080p and 2K)

720p, 1080p and 2K MiniMax H3 two-stage output are the FastH3 Two-Stage model ids, not a request option: minimax-h3-fastvideo-int8_t2v_turbo_2stage, minimax-h3-fastvideo-int8_i2v_turbo_2stage, minimax-h3-fastvideo-int8_flf2v_turbo_2stage and the audio-guide minimax-h3-fastvideo-int8_ia2v_turbo_2stage, minimax-h3-fastvideo-int8_flfa2v_turbo_2stage and minimax-h3-fastvideo-int8_a2v_turbo_2stage. Each takes exactly the request of its FastH3 Turbo id (canvas, frames, 4 steps, Euler/simple, inputs, LoRAs). FastH3 renders the canvas, then the worker enlarges it 2× and refines it, so the clip is delivered at exactly twice the canvas width and height with the same frame count, 24 fps timing and audio. Keep width/height on the normal H3 grid and pick the canvas for the delivery you want:

Portrait keeps the aspect: 384×672 delivers 768×1344, 544×960 delivers 1088×1920. Price it with estimate_video_cost() using the _2stage model id and that canvas. projects.create() and estimate_video_cost() raise ApiError before sending anything if the retired output_scale/outputScale is passed (the server refuses it too), naming the two-stage ids to use. One _2stage id per workflow serves every canvas class, so job history and cost reports show the same id at 720p, 1080p and 2K (720p is priced like one-stage FastH3); the short-lived _2stage_720p spellings were retired on 2026-09-14 and the socket answers them with "Model not found". Only a _2stage id renders two-stage: a base FastH3 id always runs one-stage at the canvas it sends. Hosted chat tools select these ids with minimax-h3-fasth3-turbo-2stage (text or first frame), minimax-h3-fasth3-t2v-turbo-2stage, minimax-h3-fasth3-i2v-turbo-2stage and minimax-h3-fasth3-flf2v-turbo-2stage; on those selectors targetResolution names the delivered class (720 renders the 384 px canvas, 1080 the 544 px canvas, 1440 or omitted the 768p canvas for 2K).

project = await sogni.projects.create(
    type="video",
    network="fast",
    model_id="minimax-h3-fastvideo-int8_t2v_turbo_2stage",
    number_of_media=1,
    steps=4,
    positive_prompt="integrated_multimodal_description: [Shot 1] ...",
    duration=8,
    width=1344,
    height=768,  # delivered at 2688x1536; send 960x544 for 1920x1088, 672x384 for 1344x768
)

MiniMax H3 audio guide (image, first/last frame, or audio only)

The FastH3 audio guide drives the video with an uploaded reference_audio from frame 0 and keeps that audio in the output, trimmed to the video length:

Each also has a _2stage id that takes the same request and delivers twice the canvas. A mode refuses any upload it does not take. The optional audio_start (seconds, 0 or greater) offsets the audio window; generate_audio=False, audio_duration and LoRAs are refused before anything is sent, and every other H3 id refuses audio_start. get_minimax_h3_frames_for_audio_duration(seconds) returns the smallest valid frame count covering the audio (124-362), and is_minimax_h3_audio_guide_model() recognizes all six ids. The hosted sound_to_video selectors are minimax-h3-fasth3-ia2v-turbo, minimax-h3-fasth3-flfa2v-turbo, minimax-h3-fasth3-a2v-turbo and their -2stage forms.

from sogni_client import get_minimax_h3_frames_for_audio_duration

project = await sogni.projects.create(
    type="video",
    network="fast",
    model_id="minimax-h3-fastvideo-int8_flfa2v_turbo",
    number_of_media=1,
    steps=4,
    positive_prompt="The dancer crosses the studio in time with the music.",
    reference_image="first.png",
    reference_image_end="last.png",
    reference_audio="song.m4a",  # the output keeps this audio
    audio_start=12,
    frames=get_minimax_h3_frames_for_audio_duration(audio_seconds - 12),
    width=1344,
    height=768,
)

GPT Image 2.5

gpt-image-2.5-flare and gpt-image-2.5-sunburst join gpt-image-2. All three accept up to 16 context_images references (never trimmed) and custom sizes up to 3840px. Quality must be a concrete value: low, medium or high, plus xhigh and max on 2.5; "auto" is rejected because every request is quoted, charged and rendered at the quality it names. 2.5 also supports gpt_image_background="transparent" (PNG or WebP output only), and gpt_image_output_compression (0-100) applies to JPEG or WebP output.

To edit part of the first reference, pass a PNG alpha mask as gpt_image_mask (bytes or a path) or gpt_image_mask_url (a URL, or a data:image/png;base64,... URI under 50 MB, which is uploaded like gpt_image_mask). Transparent mask regions are edited. In chat tools, gpt-image-2.5 and flare select Flare; sunburst selects Sunburst.

Seedance 2.5 export options

seedance-2-5 can deliver a MOV container (output_format="mov"; video defaults to mp4) and export a separate image of the final frame (return_last_frame=True). The frame is available as job.last_frame_url, and await job.get_last_frame_url() mints a fresh signed URL for it, ready to use as the first frame of a follow-up clip. Both options are Seedance 2.5 only: projects.create() raises ApiError before sending anything for another model, for an output format other than mp4/mov, or for a non-boolean return_last_frame. Chat tool results list lastFrameUrls when a frame was exported.

Reusable subscriber uploads

On servers that support saved uploads, eligible subscribers reuse the same image, video or audio file across projects. Pass files to projects.create() as usual: the client checks the account's saved copies by SHA-256 before transferring bytes, so a repeated reference is not uploaded again. Uploads stay private to the signed-in account.

saved = await sogni.projects.assets.upload(open("product.png", "rb").read(), "image/png", "Product")
listing = await sogni.projects.assets.list()  # {"assets": [...], "limits": {...}}
await sogni.projects.assets.remove(saved["id"])  # already-bound project inputs stay

Older servers and ineligible accounts keep using ordinary project uploads, but only when saved storage cannot be prepared; a transfer, checksum or binding failure after preparation stops project submission. Saved IDs do not replace file parameters in projects.create(); assets.bind(id, {"projectId": ..., "type": ...}) is available for callers that manage project input slots directly. Project history may include byolUsed, personalLoras (public-source snapshots) and reusedAssetCount; missing fields on older projects mean unknown, not zero.

Personal LoRAs

Personal imports belong to the authenticated account. Discover compatible models and limits through the library instead of hard-coding them:

library = await sogni.projects.personal_loras.list()
catalog = await sogni.projects.available_loras(include_personal=True)
ready = await sogni.projects.personal_loras.catalog(model_id="krea2_turbo_fp8_scaled")

Use personal_loras.import_lora(url=..., name=..., model_id=..., rights_confirmed=True) only after confirming permission to use the file. Imports are asynchronous: poll personal_loras.get(imported["id"]) until the status is ready, rejected, or revoked. Remove an entry with personal_loras.remove(imported["id"]).

Use ready catalog IDs with their listed model compatibility and strength ranges. The private catalog is fetched afresh and never enters the public catalog cache. Importing, ready-catalog discovery, and generation require the server's active subscription entitlement; library inspection and removal remain available after it lapses. JavaScript-style personalLoras and includePersonal aliases are also supported; Python uses import_lora because import is a language keyword.

Chat

Socket-backed completion:

result = await sogni.chat.completions.create(
    model="qwen3.6-35b-a3b-gguf-iq4xs",
    messages=[{"role": "user", "content": "Give me three visual concepts."}],
)
print(result["content"])

Hosted OpenAI-compatible completion:

result = await sogni.chat.hosted.create(
    model="qwen3.6-35b-a3b-gguf-iq4xs",
    messages=[{"role": "user", "content": "Describe a surreal album cover."}],
)

For streaming socket chat, pass stream=True and iterate over the returned ChatStream with async for.

Durable workflows

workflow = await sogni.workflows.start(
    input={"prompt": "Create a four-panel character turnaround"},
    idempotency_key="turnaround-001",
)

async for event in sogni.workflows.stream_events(workflow["id"]):
    print(event["event"], event["data"])

The client also exposes:

• sogni.account for authentication, balances, rewards, transactions, and subscriptions
• sogni.projects for generation, uploads, model discovery, and estimates
• sogni.chat for socket, hosted, tool, and durable-run APIs
• sogni.workflows and sogni.workflows.templates
• sogni.replay and sogni.stats

Python snake_case arguments are preferred. Common JavaScript-style aliases remain accepted to simplify migration.

Resuming projects after a reconnect

Generation keeps running on the Supernet while your socket is down. A dropped connection is a transport gap, not a failure: tracked projects stay alive, the client reconnects with capped exponential backoff for as long as the session is authenticated, and on every authenticated handshake it reconciles with the server. Whatever the client missed is replayed through the normal project / job events, so listeners attached before the gap keep receiving updates and wait_for_completion() still resolves.

Projects the server knows about but this client does not (a restart, a second client sharing the account, cleared local state) are rebuilt as tracked Project instances with project.recovered is True. Their params are reconstructed from the original request; asset inputs are not recoverable.

# Every reconciliation reports what changed. `snapshot` is the raw server view,
# for apps that keep their own project store.
sogni.projects.on("projectsSynced", lambda r: print(r["reason"], r["active"], r["lost"]))

# In-flight projects this client was not tracking; they are tracked now, so
# `project` / `job` events follow as usual.
sogni.projects.on("activeProjectsRecovered", lambda projects: ...)

# Projects that finished while this client was away, result URLs already resolved.
sogni.projects.on("completedProjectsRecovered", lambda projects: ...)

# Ask for a fresh reconciliation yourself, e.g. after waking from sleep.
await sogni.projects.sync()

A project the server no longer lists is looked up on the REST API (which only stores finished projects) a few times before it is declared lost; it then fails with an error where is_project_lost_error(error) is True. Apps that persist project ids themselves can run the same lookup with sogni.projects.resolve_missing(ids).

Socket server restarts

A Sogni platform release restarts the socket server: every connection closes with code 1001 for a few seconds. The SDK is built so apps need no special handling for it:

• create() and chat requests made during the gap wait (up to 30 seconds) for the reconnected, authenticated socket instead of failing.
• A project request that reached the server while it was shutting down is refused by id; the SDK sends the same request again after reconnecting, once. Projects created moments before a reconnect are re-checked when they become old enough to judge, rather than minutes later.
• LLM jobs are not carried across a restart. The server refunds them, and a stream that was open fails with a ChatJobError whose retryable is True (error_type "server_restarting" or "transport_lost") rather than waiting forever. After a plain network blip the server keeps the job for 30 seconds and the stream simply continues. Re-issue retryable failures as new requests:

from sogni_client import is_retryable_chat_error


async def complete_with_retry(**params):
    try:
        return await sogni.chat.completions.create(**params)
    except Exception as error:
        if not is_retryable_chat_error(error):
            raise
        return await sogni.chat.completions.create(**params)  # waits for the reconnect

The same snapshot answers "is anything rendering elsewhere on this account?" — sogni.projects.list_projects_elsewhere() returns those in-flight projects read-only (appSource, status, model, per-job step counts). The socket rate-limits it to 20 calls per 10s per account, so poll on the order of tens of seconds.

Recovery is per app instance: the server hands projects back to the appId that created them, so persist your appId and reuse it across restarts.

Announcements

Admin-authored in-app announcements — maintenance notices, launches — arrive on the appAlert socket event. It is opt-in, so an integration that does not ask for it is unaffected:

sogni = await SogniClient.create(
    api_key=os.environ["SOGNI_API_KEY"],
    app_id="my-announcements-app",
    app_source="my-app",
    socket_event_subscriptions={"appAlert": True},
)

sogni.api_client.on("appAlert", lambda announcement: print(announcement["title"]))

# What is live right now, for a client that just started up.
for announcement in await sogni.announcements.active("my-app"):
    print(announcement["title"], announcement["bodyMarkdown"])

# Dismissal is stored per ACCOUNT, so it sticks across the user's devices.
await sogni.announcements.dismiss(announcement["id"])

appAlert is not at-most-once: a live pinned announcement is re-sent on every reconnect, so a user who was offline when it published still receives it. Deduplicate on id.

Segmentation and 3D models

These workflows transform a source image instead of generating from a prompt, so each needs a starting_image. Ask the SDK rather than hardcoding model ids: requires_starting_image(), is_segmentation_model(), is_model_artifact_model(), is_pixal3d_model() and is_pixal3d_multiview_model(), alongside the SAM3_IMAGE_SEGMENT_MODEL_ID, PIXAL3D_IMAGE_TO_3D_MODEL_ID and PIXAL3D_MULTIVIEW_IMAGE_TO_3D_MODEL_ID constants.

SAM 3 returns one lossless mask PNG the same size as the source. The request carries a bounded sam3_prompt: points (label positive/negative), boxes (a negative box excludes one instance of a text-prompted concept and requires text), text, threshold, multimask (point prompts only), apply_mask (return the selection cut out as RGBA instead of the bare mask), and max_instances (1 to 16). Coordinates are normalized from 0 to 1.

from sogni_client import SAM3_IMAGE_SEGMENT_MODEL_ID

project = await sogni.projects.create(
    type="image",
    model_id=SAM3_IMAGE_SEGMENT_MODEL_ID,
    positive_prompt="",
    number_of_media=1,
    starting_image="room.png",
    sam3_prompt={"text": "the teapot", "apply_mask": True, "max_instances": 1},
)

Pixal3D returns a binary glTF, so job.type is "model" and the artifact downloads as model/gltf-binary. Four options — texture_size, mesh_target_faces, normal_map_size, and ambient_occlusion_size — are reduce-only and default to their maximum. shape_resolution defaults to 1024 and can be raised to the priced 1536 maximum-detail step. mesh_target_faces is the one worth setting: the 700,000-triangle default is far heavier than a real-time engine wants.

pixal3d_int8_i23d reconstructs from starting_image alone. PIXAL3D_MULTIVIEW_IMAGE_TO_3D_MODEL_ID (pixal3d_multiview_int8_i23d) takes starting_image as the required FRONT view plus any subset of three optional orbit views, each uploaded in a fixed slot. The views must show the same object at the same height, 90 degrees apart around it at eye level, like a character turnaround sheet. Name them from the subject's own point of view, not the viewer's:

Swapping left and right builds a model turned 180 degrees. Some turnaround templates label the photo of the subject's right side "left"; follow the table, not those labels. The single-view model refuses orbit views, both models refuse context_images, and only the single-view model accepts template_variant. Both take the options above.

from sogni_client import PIXAL3D_MULTIVIEW_IMAGE_TO_3D_MODEL_ID

project = await sogni.projects.create(
    type="image",
    model_id=PIXAL3D_MULTIVIEW_IMAGE_TO_3D_MODEL_ID,
    positive_prompt="",
    number_of_media=1,
    starting_image="front.png",
    left_view_image="left.png",  # optional
    back_view_image="back.png",  # optional
    right_view_image="right.png",  # optional
    mesh_target_faces=200_000,
)

When a workflow attests its inputs and outputs, job.provenance carries the worker-signed receipt. Like job.error and project.params, it is the wire record, so its keys stay camelCase: lowercase SHA-256 digests (sha256, sourceImageSha256, samPromptSha256, maskRleSha256) plus, for SAM 3, maskBox, maskCoverage, and the per-selection report (maskDetectedCount, maskReturnedCount, maskSelections) that tells a confident selection from a marginal one. Malformed entries are dropped rather than surfaced half-valid.

Sensitive content

job.is_nsfw means the server withheld the media: the render ran with the Sensitive Content Filter on, a signal fired, and there is nothing to download. When the artist turns the filter off the media is delivered and merely labelled — that case reports job.nsfw_detected with job.nsfw_sources (prompt and/or image), has a result_url like any other result, and leaves job.is_nsfw false. Use job.has_result_media (or job.is_withheld) to decide whether media exists, and the viewer's own filter setting to decide whether to blur it.

Compatibility

This release tracks the TypeScript SDK at 5.51.0. The REST, WebSocket, and SSE contracts are covered by credential-free protocol tests, including authentication refresh, uploads, project state recovery, streaming chat, workflows, templates, replay, and the canonical 27 hosted-tool schemas.

Current model and transport coverage includes LTX 2.5, MiniMax H3 in all four tiers (Standard, 8-step Balanced, 4-step LightX2V Turbo, and the separate FastH3 fastvideo-int8 Turbo engine with its audio-guide ia2v/flfa2v/a2v modes and Two-Stage 720p/1080p/2K ids), Seedance 2.5, Wan 3 and Wan 3.0 Enhanced, RTX VSR, MiniMax Music 3, Qwen3-TTS speech and voice cloning, SAM 3 image segmentation, Pixal3D image-to-3D, FlashVSR v1.1 promptless video upscaling, LoRA catalog discovery, queue start estimates, live-benchmarked render/total time on cost quotes, in-flight project recovery across reconnects, confirmed cancellation, connection/workload attribution, and admin announcements (appAlert plus the announcements read/dismiss pair).

The Python API is async-first; AsyncSogniClient is an alias of SogniClient, not a synchronous wrapper. Browser-only cookie coordination and multi-tab behavior have no Python equivalent. Local image references are uploaded with their detected MIME type, but the TypeScript client's optional browser-side image resizing is not reproduced. All 25 canonical tool schemas are exposed; the local project-backed executor handles the six direct media generation tools, while the remaining tools run through the hosted or durable chat APIs. Live, credentialed smoke tests are intentionally separate from the default test suite.

Token authentication

sogni = await SogniClient.create(app_id="my-token-app", auth_type="token")
await sogni.set_tokens(token=access_token, refresh_token=refresh_token)

Username/password login and signing are available through sogni.account.login. API-key use does not require storing a wallet password.

Development

python -m pip install -e '.[dev]'
pytest
ruff check sogni_client tests
ruff format --check sogni_client tests
python -m build

Live integration tests require explicit credentials and are not run by default.

Documentation

• Python SDK quickstart
• Sogni SDK overview
• REST API reference