Image Models (Instant)

Parel BYOM Instant image lane lets you import a public Hugging Face image generation model and call it through /v1/images/generations without provisioning a GPU. Pricing is per image (Parel applies its 1.8x margin on the upstream cost).

Supported Models

Model	HF ID	License	Notes
FLUX.1 schnell	black-forest-labs/FLUX.1-schnell	apache-2.0	4-step, fastest Flux variant
FLUX.1 dev	black-forest-labs/FLUX.1-dev	non-commercial	License acceptance required
FLUX.1 pro	black-forest-labs/FLUX.1-pro	commercial	Closed-weight
Stable Diffusion 3.5 large	stabilityai/stable-diffusion-3.5-large	community	Multi-component
SDXL base 1.0	stabilityai/stable-diffusion-xl-base-1.0	openrail++	Lightweight
Qwen-Image	Qwen/Qwen-Image	apache-2.0	Strong on multilingual prompts
PixArt-Sigma	PixArt-alpha/PixArt-Sigma-XL-2-1024-MS	openrail++	Lightweight, fast

Other Hugging Face models are validated dynamically: if inferenceProviderMapping lists a live text-to-image provider (Fal AI primary), import succeeds.

Flow

1. Validate. Confirm the model is reachable and capture pricing.
2. Import. Create a tenant-scoped tm_* model ID. If the model carries a non-commercial license (e.g. FLUX.1-dev), pass accepted_license: true.
3. Generate. Call /v1/images/generations. The endpoint returns 202 Accepted with a task_id because image generation runs through the async queue. Poll /v1/tasks/{id} to read the result URL.

import time
import httpx

API = "https://api.parel.cloud/v1"
HEADERS = {"Authorization": "Bearer pk-dev-YOUR_KEY"}

# 1. Validate
val = httpx.post(
    f"{API}/my-models/validate-hf",
    headers=HEADERS,
    json={"hf_model_id": "black-forest-labs/FLUX.1-schnell"},
).json()
assert val["supported"], val["errors"]

# 2. Import
imported = httpx.post(
    f"{API}/my-models/import-hf",
    headers=HEADERS,
    json={
        "validation_id": val["validation_id"],
        "execution_mode": "instant",
        "accepted_license": val.get("requires_license_acceptance", False),
    },
).json()
model_id = imported["parel_model_id"]   # tm_flux_1_schnell_xxxx

# 3. Generate (async)
queued = httpx.post(
    f"{API}/images/generations",
    headers=HEADERS,
    json={
        "model": model_id,
        "prompt": "a serene mountain lake at sunset, photorealistic",
        "n": 1,
        "size": "1024x1024",
    },
).json()
task_id = queued["task_id"]

# 4. Poll
for _ in range(30):
    time.sleep(2)
    state = httpx.get(f"{API}/tasks/{task_id}", headers=HEADERS).json()
    if state["status"] == "completed":
        print(state["result"]["data"][0]["url"])
        break
    if state["status"] == "failed":
        raise RuntimeError(state["error"])

Endpoints

Endpoint	Description
POST /v1/my-models/validate-hf	Returns pipeline tag, pricing, license flag
POST /v1/my-models/import-hf	Registers tm_* model id
POST /v1/images/generations	Async; returns task_id
GET /v1/tasks/{id}	Poll for status=completed and result.data[].url

Pricing

Pricing is per image. Parel applies a 1.8x margin on the upstream provider rate. Examples (USD):

Family	Upstream	Parel
FLUX.1 schnell	$0.003	$0.0054
FLUX.1 dev / SD 3.5	$0.025	$0.045
FLUX.1 pro	$0.04	$0.072
SDXL / PixArt	$0.005	$0.009
Qwen-Image	$0.018	$0.0324

The actual rate is returned in the validate response under pricing.per_image.

License Acceptance

Some models (e.g. FLUX.1-dev) ship under a non-commercial license. The validator returns requires_license_acceptance: true and a warning. Pass accepted_license: true in the import body to confirm responsibility for commercial usage. Without acceptance the import returns 400 license_acceptance_required.

Cold Start and Latency

Instant image generation is routed via Hugging Face Inference Providers (Fal AI by default). End-to-end latency depends on the provider:

• FLUX.1 schnell: 4-7 seconds warm
• FLUX.1 dev: 9-14 seconds warm
• SD 3.5 large: 6-9 seconds
• Qwen-Image: 5-8 seconds

There is no dedicated GPU and no idle charge. For deterministic latency or LoRA support, see Dedicated GPU (image dedicated path lands in BYOM Image Faz 2).

Out of Scope (Faz 2)

• Video models (Wan2.2, HunyuanVideo)
• Image edit / inpaint endpoints
• LoRA upload
• ControlNet conditioning

These ship in BYOM Image Faz 2 (PAR-44).