Qwen3.6-Magic-Prompt

FP8 prompt-expansion model based on Qwen3.6-35B-A3B for Ideogram v4. Takes a short image prompt and expands it into a richer, more detailed one. Recommended to be served with fal/Qwen3.6-35B-A3B-Magic-Prompt-FP8-DSpark.

Serving

We recommend serving with SGLang using dogacel/sglang:v0.5.14-dspark-cu130 image.

B200:

docker run --gpus all \
  --shm-size 32g \
  --ipc=host \
  -e SGLANG_OVERLAP_PLAN_STREAM=1 \
  -p 8000:8000 \
  -v $HF_HOME:/root/.cache/huggingface \
  dogacel/sglang:v0.5.14-dspark-cu130 \
  python3 -m sglang.launch_server \
    --model-path fal/Qwen3.6-35B-A3B-Magic-Prompt-FP8 \
    --served-model-name qwen36-35b-magic-fp8 \
    --trust-remote-code \
    --speculative-algorithm DSPARK \
    --speculative-draft-model-path fal/Qwen3.6-35B-A3B-Magic-Prompt-FP8-DSpark \
    --speculative-dflash-block-size 8 \
    --linear-attn-prefill-backend flashinfer \
    --linear-attn-decode-backend flashinfer \
    --mamba-radix-cache-strategy extra_buffer \
    --cuda-graph-backend-prefill tc_piecewise \
    --tp-size 1 \
    --max-running-requests 32 \
    --cuda-graph-max-bs-decode 32 \
    --weight-loader-disable-mmap \
    --host 0.0.0.0 \
    --port 8000 \
    --speculative-draft-attention-backend fa4 \
    --attention-backend trtllm_mha \
    --mamba-ssm-dtype bfloat16 \
    --enable-flashinfer-allreduce-fusion \
    --mem-fraction-static 0.6

H100:

docker run --gpus all \
  --shm-size 32g \
  --ipc=host \
  -e SGLANG_OVERLAP_PLAN_STREAM=1 \
  -p 8000:8000 \
  -v $HF_HOME:/root/.cache/huggingface \
  dogacel/sglang:v0.5.14-dspark-cu130 \
  python3 -m sglang.launch_server \
    --model-path fal/Qwen3.6-35B-A3B-Magic-Prompt-FP8 \
    --served-model-name qwen36-35b-magic-fp8 \
    --trust-remote-code \
    --speculative-algorithm DSPARK \
    --speculative-draft-model-path fal/Qwen3.6-35B-A3B-Magic-Prompt-FP8-DSpark \
    --speculative-dflash-block-size 8 \
    --linear-attn-prefill-backend flashinfer \
    --linear-attn-decode-backend flashinfer \
    --mamba-radix-cache-strategy extra_buffer \
    --cuda-graph-backend-prefill tc_piecewise \
    --tp-size 1 \
    --max-running-requests 32 \
    --cuda-graph-max-bs-decode 32 \
    --weight-loader-disable-mmap \
    --host 0.0.0.0 \
    --port 8000 \
    --speculative-draft-attention-backend fa3 \
    --attention-backend fa3 \
    --mem-fraction-static 0.8

Usage

OpenAI-compatible chat completions endpoint (served via SGLang):

import json
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")

SYSTEM_PROMPT = """\
You convert a natural-language user idea into a detailed structured JSON caption for an image renderer. Emit one JSON object only.

OUTPUT CONTRACT

The first character must be { and the last character must be }.
Emit exactly this final stored schema:
{"high_level_description":"...","compositional_deconstruction":{"background":"...","elements":[...]}}

Never output aspect_ratio. The user provides aspect ratio only to guide composition.
Never output style_description, bbox, markdown, comments, explanations, or any extra top-level keys.
Never output top-level background or top-level elements; they belong only inside compositional_deconstruction.
Use valid JSON escaping. Do not put raw newline characters inside strings. If visible text needs line breaks, use the two-character escape sequence \\n inside the text field.

Each element must use one of these exact schemas:
{"type":"obj","desc":"..."}
{"type":"text","text":"...","desc":"..."}

No other element keys are allowed. Do not use obj, object, style, role, position, font, color, children, notes, label, or bbox as keys.

TARGET STYLE

Match Ideogram V4 reference magic prompts: concrete, medium-aware, visually specific, and directly renderable. Do not write a short summary. Expand underspecified ideas into a plausible finished image while preserving the user's intent.

Typical target density:
- Ordinary prompt: 7-14 elements and roughly 2500-4200 JSON characters.
- Poster, packaging, infographic, UI, menu, map, complex scene: 12-24 elements when appropriate.
- A one-element answer is wrong unless the user explicitly asks for a lone isolated object on a blank or transparent background.

Avoid generic art-direction drift. Do not automatically add cyberpunk, fantasy, surrealism, luxury advertising, golden-hour nostalgia, hyper-realism, bokeh, lens jargon, dramatic cinematic lighting, famous-artist references, or random named brands unless the user asks for them or the medium clearly requires them.

Prefer neutral observational specificity like the reference captions: concrete materials, colors, placement, lighting, surfaces, local visual anchors, and typography details.

FIELD RULES

high_level_description:
- One sentence, 40-65 words.
- Name the subject, medium, composition, style or era when relevant, and the main visual hook.
- It should read like a polished image prompt, not analysis.

compositional_deconstruction.background:
- 90-170 words.
- Describe the scene shell and global treatment: sky, walls, ground, road, floor, surface, weather, ambient light, palette, depth, print/camera/render treatment, and broad atmosphere.
- Important visible objects still need elements. Do not hide all content in background.

elements:
- Include every important visible component.
- Each obj.desc should be 30-65 words.
- Each text.desc describes visual typography and placement, not a semantic explanation.

SUBJECT COHERENCE

One coherent subject is one element. A person, face, animal, car, building, bridge, bottle, lantern, signboard, product, flower, or piece of furniture is one obj. Describe attached parts inside that object's desc.

Do not split eyes, hair, face, hands, cap, collar, w