Sign in

b560c0ad0618

latestOpenAPI 3.0.22026-08-11115250435.0 KB
prompt

Generate Magic Prompt

Transforms basic prompts into a magic prompt. Internal use only (feature flagged). TODO: update description if enabled externally.

post/magic-prompt

Request body

promptstring

The prompt to enhance with a magic prompt. Mutually exclusive with messages; one of the two must be set.

magic_prompt_version'V_0' | 'V_0_1' | 'V_0_2' | 'V_0_3' | 'V_0_4' | 'V_0_5' | 'V_0_6' | 'V_0_7' | 'V_0_8' | 'V4_QWEN_3_5_27B' | 'V4_QWEN_3_5_27B_PREFUSED' | 'V4_QWEN_3_5_27B_FULL_FINETUNE' | 'V4_QWEN_3_5_27B_FULL_FINETUNE_LORA' | 'V4_QWEN_3_5_27B_EXPERIMENTAL' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_Q36_2P5K' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_NONNAS_2P5K' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_NONNAS_5K' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_ALTCAP_BAL_9K' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_ALTCAP_BAL_10K' | 'V4_QWEN_3_5_27B_SGLANG_EXPERIMENTAL_Q36_2P5K' | 'V4_QWEN_3_5_27B_SGLANG_EXPERIMENTAL_Q36_2P5K_TP2' | 'V4_QWEN_3_5_27B_PRUNA_FP8' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_01' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_02' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_03' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_04' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_05' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_06' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_07' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_08' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_09' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_10' required

The magic prompt version to use when magic prompt option is set to AUTO or ON.

classify_prompt_categoryboolean required

Whether to classify the prompt into a use case category.

style_type'AUTO' | 'GENERAL' | 'REALISTIC' | 'DESIGN' | 'CUSTOM' | 'FICTION' | 'STYLIZED'

The style type to generate with.

seedinteger

Random seed for reproducible generation

character_reference_imagesstring[]

Generations with character reference are subject to the character reference pricing. A set of images to use as character references (maximum total size 10MB across all character references), currently only supports 1 character reference image. The images should be in JPEG, PNG or WebP format.

system_promptstring

A system prompt to use for the magic prompt.

aspect_ratio'1x3' | '3x1' | '1x2' | '2x1' | '9x16' | '16x9' | '10x16' | '16x10' | '2x3' | '3x2' | '3x4' | '4x3' | '4x5' | '5x4' | '1x1'

The aspect ratio to use for image generation, which determines the image's resolution. Cannot be used in conjunction with resolution. Defaults to 1x1.

temperaturenumber float

Sampling temperature override. Defaults to the trainer-conditioned value when omitted (0.7 for V4_QWEN_3_5_27B). Only honored with V4 magic-prompt versions; supplying it with another version returns 400.

top_pnumber float

Nucleus-sampling top_p override. Only honored with V4 versions.

top_kinteger

Top-k sampling override. Only honored with V4 versions.

min_pnumber float

min_p sampling override. Only honored with V4 versions.

max_tokensinteger

Cap on the number of tokens the model may generate. Defaults to 4096 (the trainer's reference cap) when omitted. Only honored with V4 versions.

presence_penaltynumber float

Presence-penalty override. Only honored with V4 versions.

frequency_penaltynumber float

Frequency-penalty override. Only honored with V4 versions.

repetition_penaltynumber float

Repetition-penalty override. Only honored with V4 versions.

checkpointstring

Optional override of the LoRA adapter the V4 magic-prompt wrapper downloads and runs. Pass a gs:// URI pointing at a checkpoint directory containing adapter_config.json. When omitted, the wrapper-preloaded production checkpoint is used. Honored with the wrapper-routed LoRA versions V4_QWEN_3_5_27B and V4_QWEN_3_5_27B_FULL_FINETUNE_LORA; supplying it with V4_QWEN_3_5_27B_PREFUSED or V4_QWEN_3_5_27B_FULL_FINETUNE (which serve fixed weights and cannot swap adapters per request) or any other version returns 400.

Example request

{
  "prompt": "a cat",
  "magic_prompt_version": "V_0_4",
  "classify_prompt_category": true,
  "style_type": "AUTO",
  "seed": 12345
}

Response

Magic prompt generated successfully.

magic_promptstring required

The generated magic prompt

magic_prompt_version'V_0' | 'V_0_1' | 'V_0_2' | 'V_0_3' | 'V_0_4' | 'V_0_5' | 'V_0_6' | 'V_0_7' | 'V_0_8' | 'V4_QWEN_3_5_27B' | 'V4_QWEN_3_5_27B_PREFUSED' | 'V4_QWEN_3_5_27B_FULL_FINETUNE' | 'V4_QWEN_3_5_27B_FULL_FINETUNE_LORA' | 'V4_QWEN_3_5_27B_EXPERIMENTAL' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_Q36_2P5K' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_NONNAS_2P5K' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_NONNAS_5K' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_ALTCAP_BAL_9K' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_ALTCAP_BAL_10K' | 'V4_QWEN_3_5_27B_SGLANG_EXPERIMENTAL_Q36_2P5K' | 'V4_QWEN_3_5_27B_SGLANG_EXPERIMENTAL_Q36_2P5K_TP2' | 'V4_QWEN_3_5_27B_PRUNA_FP8' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_01' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_02' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_03' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_04' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_05' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_06' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_07' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_08' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_09' | 'V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_10'

The magic prompt version to use when magic prompt option is set to AUTO or ON.

served_model_namestring nullable

Identifier of the underlying model that produced this prompt. For experimental versions, whose backing model can change over time, this reports the model that actually served the request. Null when not applicable.

Example response

{
  "magic_prompt": "A photorealistic cat with striking green eyes, sitting gracefully on a vintage wooden chair. The cat's fur has beautiful orange and white markings, and soft natural lighting illuminates the scene from a nearby window.",
  "magic_prompt_version": "V4_QWEN_3_5_27B_EXPERIMENTAL_SLOT_01",
  "served_model_name": "qwen35-27b-magic-prompt-experimental-2026-06-17"
}