Phosor AI
Create videos with MiniMax H3 Standard or Turbo (text-to-video, image-to-video, and reference-to-video) through Phosor AI, then use the same skill for images, speech, LLM chat (Qwen3.8 27B, OpenAI-compatible), LoRA workflows, and e-commerce product or model photography. Use when the user wants to create H3 video from text, images, or references, compare Standard and Turbo, use the Phosor API, generate images or speech, chat with or call the Qwen3.8 27B LLM, use a custom LoRA, create e-commerce photography, or manage generation jobs.
Phosor AI
MiniMax H3 Standard and Turbo are the primary video workflows in this skill. Use H3 for text-to-video, image-to-video, or reference-to-video, then use the same Phosor API for images, speech, LLM chat, LoRA workflows, and e-commerce product or model photography when needed.
For detailed API endpoints, parameters, pricing, and limits, see references/api.md.
Before preparing or troubleshooting an H3 prompt, read references/h3-prompt-guide.md. It covers actual dialogue versus camera/visual text, silence, reference/voice binding and enhancement preview.
Submit the original prompt. Every H3 video submission performs mandatory prompt processing, including structured input; there is no opt-out. If processing fails, stop generation and follow the refund path for any deducted credits. Previews cannot bypass submission checks or lock the later execution text. The optional avoid_model_limitations boolean defaults to false and controls adaptation only. The optional unrestricted_enhancement boolean (default false, CLI --unrestricted-enhancement) runs processing on an open model that does not refuse mature or sensitive creative themes; use it only when the user asks for it or the default processing refuses a legitimate creative prompt. Safety checks are unchanged.
Setup
Set your API key:
export PHOSOR_API_KEY="your-api-key-here"
Get an API key at phosor.ai → Settings → API Keys.
The CLI script is at scripts/phosor_client.py. All commands output JSON to stdout.
Error handling
API diagnostics are in English regardless of the website language or
Accept-Language. Explain errors to the user in their preferred language, using
the HTTP status and documented error codes rather than matching English wording.
Keep any request_id / X-Request-ID for troubleshooting.
HTTP 503 with error: maintenance or code: SYSTEM_UPGRADING means the system is
being upgraded and this request was not accepted. Do not mark it as a failed
generation job or automatically resubmit it. Tell the user to wait until maintenance
ends; the public GET /api/v1/system/status endpoint reports maintenance: true
or maintenance: false. Retry-After is retry advice, not a completion estimate.
A generic 503 is temporary unavailability and does not establish a planned upgrade.
For 429 responses, inspect reject_reason / X-Reject-Reason and Retry-After
when provided. Distinguish API key quota, server capacity and queue wait limits.
See references/api.md for the response contract.
Base URL
The client talks to https://phosor.ai by default, over HTTPS. Get your key from phosor.ai → Settings → API Keys.
python3 scripts/phosor_client.py --api-key <your-key> check-key
The bundled client connects only to https://phosor.ai over HTTPS. Redirects are rejected.
Note: studio-analyze / studio-suite --image-url must be the full https S3 URL returned by upload-image, not the bare S3 key path — a bare key errors with unsupported URL scheme.
H3 model map
Use an exact model ID. Each H3 mode has a Standard and Turbo variant; Turbo is the faster option for the same input shape. H3 uses duration, resolution tier, and aspect ratio rather than Wan's frame controls.
| Input | Standard | Turbo |
|---|---|---|
| Text to video | minimax/h3/text-to-video | minimax/h3/text-to-video/turbo |
| Image to video | minimax/h3/image-to-video | minimax/h3/image-to-video/turbo |
| Reference to video | minimax/h3/reference-to-video | minimax/h3/reference-to-video/turbo |
Choose Standard or Turbo from the exact model ID before submitting. H3 output is 24fps and billed per output second; see the pricing section for current rates.
LLM chat — Qwen3.8 27B
qwen3.8-27b is a text model (category LLM) behind the OpenAI-compatible
POST /api/v1/chat/completions. Unlike the generation models it is synchronous:
the answer is in the response (or the SSE stream) — there is no request_id, status or result step.
# One-shot answer (JSON response on stdout, includes usage.credits_charged)
python3 scripts/phosor_client.py chat "Summarize the plot of Hamlet in 3 sentences." --max-tokens 300
# Stream text to stderr as it arrives; assembled JSON still goes to stdout
python3 scripts/phosor_client.py chat "Write a haiku about autumn." --stream
# System prompt, JSON mode, thinking, images (URL or local file, max 2, 4 MB each)
python3 scripts/phosor_client.py chat "Extract name and email" --system "Reply as JSON" --json-mode
python3 scripts/phosor_client.py chat "Is 391 prime?" --thinking --max-tokens 4096
python3 scripts/phosor_client.py chat "What is in this photo?" --image ./photo.jpg
# Full conversation or tool calling: pass OpenAI-format JSON (inline or @file)
python3 scripts/phosor_client.py chat --messages @messages.json --tools @tools.json --tool-choice auto
Rules an agent must follow:
- Send only the documented fields (
model,messages,max_tokens,temperature,top_p,stop,stream,stream_options,response_format,tools,tool_choice,enable_thinking,seed,presence_penalty,frequency_penalty,n=1,user). Any other field returns400 invalid_request. - Limits: 48K-token context,
max_tokens1–8192 (default 1024), at most 2 images per request. - Cost: USD 0.2125 per 1M input tokens and USD 1.275 per 1M output tokens (2.125 / 12.75 credits).
Credits for the estimated input plus
max_tokensare reserved up front, then settled to actual use — keepmax_tokensclose to what you need. Thinking tokens are billed as output. Readusage.credits_chargedfrom the response. 402 insufficient_credits: lowermax_tokensor ask the user to add credits.429 rate_limited(60 requests/min per key) or429 capacity_busy: waitRetry-Afterseconds, then retry.503 service_unavailable: not charged; retry later. Do not retry a400unchanged.
Direct HTTP (any OpenAI SDK works with base_url="https://phosor.ai/api/v1"):
curl https://phosor.ai/api/v1/chat/completions \
-H "Authorization: Bearer $PHOSOR_API_KEY" -H "Content-Type: application/json" \
-d '{"model":"qwen3.8-27b","messages":[{"role":"user","content":"Hello"}],"max_tokens":256,"stream":true}'
Full field, streaming and error reference: references/api.md.
Generation examples
Read generation examples for copyable video, image, speech and LoRA commands. Read only the section matching the selected model.
CLI Commands
| Command | Description | Key Arguments |
|---|---|---|
check-key | Validate API key | — |
submit | Submit inference job (T2V/I2V/S2V/Animate/T2I/I2I) | prompt, --width, --height, --num-frames, --fps, --steps, --guidance, --image-url, --audio-url, --video-url, --lora-id, --lora-scale, --loras, --seed, --negative-prompt, --model, --num-images, --strength, --output-format |
status | Get job status | request_id |
result | Get job result (video or image URL) | request_id |
poll | Poll all pending jobs | — |
list | List locally tracked pending jobs | — |
history | Get job history | --limit |
upload-image | Upload image for I2V or I2I | file |
import-image | Import image from URL | url, --filename |
upload-lora | Upload LoRA (two .safetensors for video) | high_noise_file, low_noise_file, --name |
import-lora | Import LoRA from URLs (one or two files) | high_noise_url, [low_noise_url], --name |
loras | List LoRA models | --limit, --offset |
lora-status | Get LoRA upload/import status | lora_id |
save-lora | Activate a LoRA (extends expiry to 7 days) | lora_id, --name |
delete-lora | Delete a LoRA model | lora_id |
chat | Chat with the Qwen3.8 27B LLM (OpenAI-compatible, synchronous; --stream for SSE) | prompt, --system, --messages, --image, --model, --max-tokens, --temperature, --top-p, --stop, --seed, --json-mode, --tools, --tool-choice, --thinking, --stream |
submit-tts | Submit a text-to-speech job (Qwen3-TTS) — keys off text, not a prompt | text, --speaker, --language, --seed, --temperature, --top-p, --top-k, --repetition-penalty |
models | List available video/image models (static offline reference) | — |
studio-features | List Image Studio endpoints, fields, billing (static offline reference) | — |
studio-pricing | Get live Image Studio pricing | — |
studio-analyze | AI-analyze a product/garment image or reference URL (freemium) | --target agent|product|model|reference, --image-url, --url, --prompt, --language |
studio-layouts | List the layout template library (query, then select) — static asset on the web front (phosor.ai), not a /api/v1 endpoint; needs no key; on a self-hosted deployment set PHOSOR_WEB_BASE_URL if the web front is not the gateway host | --module (product|clothing), --type (selling_point|aplus|white_bg|scene|closeup|size_chart) |
studio-suite | Generate a product image suite | --image-url, --layout-types, --count-per-type, --custom-suggestions, --template-ids (ids from studio-layouts, auto-expanded to custom_suggestions like the UI's manual pick — use this to get text-callout selling-point / A+ layouts and model templates), --product-info, --aspect-ratio, --gen-language, --model, --same-style-reference |
studio-clothing-suite | Generate a model/garment image suite | --image-urls, --main-image-types, --aplus-types, --product-info, --brand-config, --aspect-ratio, --gen-language, --model, --same-style-reference |
studio-status | Get Image Studio job status (separate id space, same request_id key) | request_id |
studio-cancel | Cancel a running generation — queued images refunded, already-generating ones charged | request_id |
studio-my-works | List past Image Studio generations | --task-type, --limit, --offset |
studio-call | Generic call for any other Image Studio endpoint (remove-bg, replace, inpaint, erase, handheld, translate, outpaint, recolor, enhance, upscale, scene-compose, scene-variation, real-model-swap, mannequin-swap, model-scene-swap, ai-outfit, pose-variation, ai-wearable) | method, path, --json |
Image Studio (Product & Model Photography)
Image Studio is a separate product surface for e-commerce sellers — AI product photography and model/clothing photography — reached through the same gateway and API key as video/LoRA, under the /api/v1/image-studio prefix. It has its own async namespace - the same key name request_id, but a separate id space: an Image Studio request_id is not valid on /api/v1/inference/status/... and vice versa - and its own pricing (flat per-image rate + freemium analyze quota, not per-frame). Full endpoint/parameter reference: references/api.md.
Quick Start: Product Suite
# 1. Upload the product photo
python3 scripts/phosor_client.py upload-image /path/to/product.jpg
# Returns: {"file_id": "img-xxx", "s3_key": "images/img-xxx.jpg", ...}
# 2. (Optional) AI-analyze it first for richer generation context
python3 scripts/phosor_client.py studio-analyze --target product --image-url "images/img-xxx.jpg"
# 3. Generate a product image suite
python3 scripts/phosor_client.py studio-suite --image-url "images/img-xxx.jpg" \
--layout-types "white_background,lifestyle_scene" --count-per-type 2
# 4. Poll for the result
python3 scripts/phosor_client.py studio-status <request_id>
Quick Start: Clothing/Model Suite
python3 scripts/phosor_client.py upload-image /path/to/garment.jpg
python3 scripts/phosor_client.py studio-clothing-suite \
--image-urls "images/img-xxx.jpg" \
--main-image-types '{"model_shot":2,"selling_point":1}' \
--aplus-types '{"standard_aplus":1}'
python3 scripts/phosor_client.py studio-status <request_id>
Quick Start: One-off Edits (remove-bg, inpaint, translate, etc.)
The long tail of single-purpose editing endpoints doesn't get a dedicated subcommand — use studio-call with the exact field names from references/api.md:
python3 scripts/phosor_client.py studio-call POST /product/remove-bg \
--json '{"image_url": "images/img-xxx.jpg", "count": 2}'
python3 scripts/phosor_client.py studio-status <request_id>
Key facts
- Every Image Studio call requires
X-API-Key(PHOSOR_API_KEY), includingGET /pricing— there is no unauthenticated endpoint under this prefix. - All generation/analyze endpoints are async: POST returns
{"request_id": "...", "status": "pending"}; pollstudio-status <request_id>untilstatusis"done","error"or"cancelled". Image Studio ids live in a separate id space from video/LoRA: Usestudio-statusfor these IDs;status/resultare for inference. The automaticpollcommand routes tracked Studio jobs to their own status endpoint. - Cancelling:
POST /jobs/{request_id}/cancelstops a running generation. Images still queued are refunded; images already generating are charged and cannot be stopped; images already delivered bill once through the normal path. The response reports the split asrefunded_queued,charged_runningandalready_done, and the task then polls asstatus: "cancelled"- not an error. - Pricing is per-image, not per-frame: call
studio-pricingfor the live rate. Partial success (e.g. 3 of 5 images) bills only the successes. - Analyze is freemium:
agent/analyze,product/analyze,model/analyzeshare a daily free quota before per-call billing kicks in. model_attrsmatters for model-photography endpoints (real-model-swap, mannequin-swap, ai-outfit, ai-wearable) — pass{gender, age_group, ethnicity, skin_tone, hair_color}explicitly; it is not reliably inferred from the source image alone.- Run
studio-featuresfor the full offline endpoint/field catalog without leaving the terminal.
Key Constraints
Video Resolutions (exact pairs only)
| Preset | Width × Height | Max Frames (turbo) | Max Frames (standard) |
|---|---|---|---|
| 480p landscape | 854 × 480 | 161 | 161 |
| 480p portrait | 480 × 854 | 161 | 161 |
| 720p landscape | 1280 × 720 | 161 | 161 |
| 720p portrait | 720 × 1280 | 161 | 161 |
| 1080p landscape | 1920 × 1080 | 153 | 81 |
| 1080p portrait | 1080 × 1920 | 153 | 81 |
Standard (non-turbo) mode: 1080p is capped at 81 frames due to generation time limits.
S2V / Animate Video Resolutions (exact pairs only)
| Preset | Width x Height | Max Frames |
|---|---|---|
| 480p landscape | 854 x 480 | 161 |
| 480p portrait | 480 x 854 | 161 |
| 512p square | 512 x 512 | 161 |
| 720p landscape | 1280 x 720 | 161 |
| 720p portrait | 720 x 1280 | 161 |
Image Resolutions (exact pairs only)
| Preset | Width × Height |
|---|---|
| Square small | 512 × 512 |
| Square | 1024 × 1024 |
| Landscape | 1024 × 768 |
| Portrait | 768 × 1024 |
| Wide landscape | 1280 × 768 |
| Tall portrait | 768 × 1280 |
MiniMax H3 Frame Sizes (--resolution-tier + --aspect)
| Tier | 16:9 | 4:3 | 1:1 | 3:4 | 9:16 |
|---|---|---|---|---|---|
| 480p | 832 × 480 | 640 × 480 | 480 × 480 | 480 × 640 | 480 × 832 |
| 768p | 1344 × 768 | 1024 × 768 | 768 × 768 | 768 × 1024 | 768 × 1344 |
duration is 4–15 seconds (default 5). Output FPS is fixed at 24 and
frames_per_second is ignored.
Ref2VA Reference Limits (minimax/h3/reference-to-video)
Caps differ per tier, and the image cap is higher when you send only images:
| Limit | 480p | 768p |
|---|---|---|
| Reference images (with videos/audio present) | 5 | 5 |
| Reference images (images only) | 9 | 9 |
| Reference videos | 3 | 3 |
| Reference audios | 3 | 3 |
| Limit | Value |
|---|---|
| Total reference video length | 15.1 s (across all reference videos) |
| Reference video FPS ceiling | 24 |
| Reference audio length | 10 s each |
| Reference image longest edge | 2048 px |
| Reference image aspect ratio | ≤ 4.0 |
FLUX.2-dev Resolutions (exact pairs only)
| Width × Height |
|---|
| 2048 × 1536 · 1536 × 2048 |
| 2048 × 1152 · 1152 × 2048 |
| 2048 × 2048 · 1024 × 1024 |
FLUX.2-dev does not accept the general image resolution list above, and always returns exactly 1 image.
Frame Alignment (video only)
Frames must follow 1 + 4*k where k >= 1 (e.g. 5, 9, 13, ... 81, 85, ...). Server auto-aligns down.
Video Inference Parameters
| Parameter | Default | Range |
|---|---|---|
frames_per_second | 16 | 4–60 |
num_inference_steps | Model-dependent | Consult the selected model contract |
guidance_scale | Model-dependent | Consult the selected model contract |
Image Inference Parameters
| Parameter | Default | Range | Notes |
|---|---|---|---|
num_images | 1 | 1–4 | Number of images to generate |
num_inference_steps | varies | 1–4 (z-image turbo), 1–40 (qwen-image) | Model-dependent max |
guidance_scale | varies | 1.0–20.0 | — |
strength | — | 0.0–1.0 | Image-to-image only: how much to transform the source |
output_format | png | png, jpeg | Output file format |
Concurrency
Model API inference jobs run concurrently up to a per-account cap; over it the submit
returns 429 and you retry after a job finishes.
| Tier | Plan | Concurrent Model API jobs |
|---|---|---|
| Free | — | 1 |
| 1 | Starter | 2 |
| 2 | Standard | 4 |
| 3 | Pro | 8 |
Image Studio runs on its own pool and does not consume this quota — a suite and a video generation can run at the same time.
Multiple LoRAs
python3 scripts/phosor_client.py submit "A person dancing" \
--loras '[{"lora_id": "lora-abc", "lora_scale": 0.8}, {"lora_id": "lora-def", "lora_scale": 0.5}]'
Two-Step Upload Rule
Files must be uploaded before use — direct URLs are NOT supported in submit --image-url:
- Image →
upload-image/import-image→ returnss3_key→ use as--image-url - LoRA →
upload-lora/import-lora→ returnslora_id→ use as--lora-id
Queue Flow
PENDING → PROCESSING → COMPLETED / FAILED
The poll command checks all locally-tracked pending jobs and removes completed/failed ones.
MiniMax H3 Pricing (per output second)
H3 charges per output second. Confirm currently effective rates with GET /api/v1/pricing/config.
10 credits = USD 1. Prices and billing are in USD.
| Model | 480p USD/sec | 768p USD/sec | 480p credits/sec | 768p credits/sec |
|---|---|---|---|---|
text-to-video, image-to-video | $0.0045 | $0.012 | 0.045 | 0.12 |
text-to-video/turbo, image-to-video/turbo | $0.006 | $0.012 | 0.06 | 0.12 |
reference-to-video | $0.0045 | $0.012 | 0.045 | 0.12 |
reference-to-video/turbo | $0.006 | $0.012 | 0.06 | 0.12 |
Reference-to-video adds reference inputs to the output charge:
| Model | Tier | Output USD/sec | Reference images | Reference audio USD/clip | Reference video USD/sec |
|---|---|---|---|---|---|
reference-to-video | 480p | $0.0045 | first 5 free, then $0.0075 each | $0 | $0.0045 |
reference-to-video | 768p | $0.012 | first 5 free, then $0.0075 each | $0 | $0.012 |
reference-to-video/turbo | 480p | $0.006 | first 5 free, then $0.0025 each | $0 | $0.006 |
reference-to-video/turbo | 768p | $0.012 | first 5 free, then $0.0025 each | $0 | $0.012 |
inputs_usd = round(max(image_count - free_image_count, 0) * image_rate
+ audio_count * audio_rate
+ round(sum(actual_reference_video_seconds), 3) * reference_video_rate, 6)
total_usd = round(output_billable_seconds * output_rate + inputs_usd, 6)
total_credits = round(total_usd * 10, 4)
Reference video uses the actual combined duration of all clips, even when longer than the output; it is not capped at output duration. Reference audio remains free. Rounding follows the Gateway (Python rounding).
Turbo 768p example: 5 output seconds + 9 images + 14.75 reference-video seconds + 3 audio clips = USD 0.247.
Existing task charges remain as recorded; current rates do not reprice past tasks.
Wan Video Pricing (per frame)
| Tier | Standard | + LoRA | Turbo |
|---|---|---|---|
| 480p | $0.0009375 | $0.00125 | $0.0003125 |
| 512p | $0.0013125 | $0.001625 | $0.0004375 |
| 720p | $0.001875 | $0.0021875 | $0.000625 |
| 1080p | $0.0025 | $0.003 | $0.0010938 |
LoRA multiplier on turbo: 1.2x.
S2V Pricing (per frame)
| Tier | Cost |
|---|---|
| 480p | $0.0009375 |
| 512p | $0.0013125 |
| 720p | $0.001875 |
Animate Pricing (per frame)
| Tier | Cost |
|---|---|
| 480p | $0.00125 |
| 512p | $0.00175 |
| 720p | $0.0025 |
Image Pricing (per image)
| Model | USD | Credits |
|---|---|---|
| GPT Image 2 T2I | $0.03 | 0.3 |
| FLUX.2-dev T2I | $0.006 | 0.06 |
| FLUX.2-dev Image Edit | $0.012 | 0.12 |
| qwen-image T2I | $0.015 | 0.15 |
| qwen-image T2I + LoRA | $0.018 | 0.18 |
| qwen-image-edit | $0.003 | 0.03 |
| z-image turbo (T2I / I2I) | $0.0025 | 0.025 |
| z-image turbo + LoRA | $0.003 | 0.03 |
Flat per image at every resolution. Multiply by num_images (1–4) where the model
supports it; FLUX.2-dev is fixed at 1 image per request.
GPT Image 2, like FLUX.2-dev, is fixed at one image per request.
Availability and business boundaries
Model API prices do not price Image Studio suites or Drama workflows. Use their own contracts and price sources. Use model IDs returned by the current model catalog.
Price reference checked 2026-11-04; live Model API prices take precedence. A price configuration is not proof of submission availability or successful delivery.
LLM Pricing (per 1M tokens)
| Model | Input | Output | Input credits | Output credits |
|---|---|---|---|---|
qwen3.8-27b | $0.2125 | $1.275 | 2.125 | 12.75 |
Billed on actual usage (cached input at the input rate, thinking tokens as output);
each response reports usage.credits_charged.
Audio Pricing
| Item | Cost |
|---|---|
| Qwen3-TTS | $0.00003 per character |
| Minimum charge | $0.003 per request |
Exchange rate: 10 credits = $1 USD. Credits pre-deducted, auto-refunded on failure.
Live rates: GET /api/v1/pricing/config — always prefer it over any table here.
MiniMax H3 Image pricing
Text-to-image: minimax/h3/text-to-image. Fixed price per generated image:
| Resolution tier | USD/image | credits/image |
|---|---|---|
| 1K | 0.0175 | 0.1750 |
| 1.5K | 0.0180 | 0.1800 |
| 2K | 0.0187 | 0.1870 |
| 3K | 0.0338 | 0.3380 |
| 4K | 0.0550 | 0.5500 |
1 USD = 10 credits. USD and credits use at most four decimal places. Billing uses the supported width/height preset's pixel-area tier; do not infer a tier from the longest edge or invent custom sizes. PNG/JPEG/WebP and aspect ratio do not add a surcharge within a tier. One image per request. Check the target environment's /api/v1/pricing/config for effective prices before submission; availability follows that environment's model catalog. Prompt processing runs automatically and is not a configurable API parameter.
MiniMax H3 Image quick start
python3 scripts/phosor_client.py submit "A ceramic cup in soft daylight" \
--model minimax/h3/text-to-image --width 1024 --height 1024 --output-format webp
One image per request. Use exact supported dimensions. Prompt processing is automatic.
For reference-to-video labels, use --reference-image-labels '["Alice", "Garden"]'
with two reference image URLs in the same order; video and audio labels work likewise.
Wan Turbo uses service-selected frame counts and FPS; do not infer output duration from requested frame controls. Confirm the model contract before quoting duration.
Related skills
- Downloads
- 1.7kAll time
- Security audit
- Pass
- Last updated
- 5d ago
- Current version
- v1.4.1
- License
- MIT-0
