starnodes/ qwen3.8-vl-27b-promptrefiner-abliterated

99 8 hours ago

vision 
ollama run starnodes/qwen3.8-vl-27b-promptrefiner-abliterated

Models

View all →

1 model

qwen3.8-vl-27b-promptrefiner-abliterated:latest

20GB · 256K context window · Text, Image · 8 hours ago

Readme

Qwen3.8-VL-27B Universal Prompt Architect (Abliterated)

An advanced multi-engine prompt engineering assistant powered by Qwen3.8-27B (Abliterated). Designed specifically for generative AI workflows, this model intelligently formats, routes, and enriches prompts across top-tier Text-to-Image, Image-to-Video, Text-to-Video, 360° Panorama, Storyboarding, and AI Music generators.

⚡ Key Highlights & Capabilities

  • Zero Filler: Outputs strictly valid prompts, JSON objects, or multi-shot syntax without conversational banter.
  • Multi-Engine Dispatcher: Understands tags/triggers to tailor output structure to specific models (FLUX.2 Klein, MiniMax H3, Ideogram, LTX, ACE-Step, and more).
  • Vision-Grounded: Uses vision capabilities to evaluate input images for seamless Image-to-Video and Inpainting/Editing prompts.
  • Abliterated Backbone: Uncensored reasoning layer ensures no refusals when handling creative, stylized, dark-fantasy, or unfiltered prompt engineering workflows.

🚀 Quick Start (Ollama)

# Run interactively
ollama run starnodes/qwen3.8-vl-27b-promptrefiner-abliterated

# One-liner execution
ollama run starnodes/qwen3.8-vl-27b-promptrefiner-abliterated "@flux a cybernetic knight in rainy neo-tokyo"

🏷️ Supported Engines & Trigger Words

You can trigger specific generation modes by prefixing your prompt with @tag or including the engine name naturally in your request.

Generation Mode Trigger Keywords / Tags Output Format & Optimizations
FLUX.2 / Klein @flux, @flux2, @klein, flux klein Novelistic 3-Paragraph Prose: Subject hierarchy, physical lighting geometry, optics, and precise CSS Hex codes (#HEX). Handles image edit verbs (Replace X with Y).
Ideogram 4 (JSON) @ideogram, @ideogram4, ideogram json Pure JSON: Strict schema (high_level_description, style_description, compositional_deconstruction, color_palette, typography bounding descriptions). No code fences.
360° Panorama T2I @pano360, @krea-pano, 360 panorama, equirectangular Seamless 2:1 Pano: Starts mandatory with seamless equirectangular 360 degree panorama,. Balances zenith/nadir and edge wrapping.
Standard T2I / Krea @image, @krea, @boogu, (Default) Dense High-Fidelity Paragraph: Faithfulness to subjects, camera framing, lighting, exact typography handling with "quotes".
Ernie-ViLG @ernie, @vilg, ernie-image Technical Rendering Hierarchy: Octane, ray-traced reflections, subsurface scattering, macro camera profiles, structured blocks.
MiniMax H3 Text2Video @minimax-t2v, @t2va 3 Core Sections: integrated_multimodal_description with [Shot N] cut-timestamps, dynamic camera tracking, <d>[Language] Text</d> dialogue, soundscape, and non-diegetic music.
MiniMax H3 Image2Video @minimax-i2v, @i2va, minimax i2v Full Reference I2V: Mandatory header (...<Picture 1> is fully referenced), followed by shot continuity, physical motion development, and audio.
MiniMax H3 360° Video @minimax-pano, @i2v-pano Spherical Motion Animation: Mandatory header + seamless equirectangular 360 degree panorama in Shot 1 + spatial audio cues.
MiniMax H3 Full-Reference @minimax-fullref, @minimax-a2v, @minimax-v2v 6 Structured Sections: subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, non_diegetic_music.
LTX-2.3 & LTXV @ltx, @ltx-video, @ltxv-i2v Action & Dynamics Focus: Spoken dialogues with audio descriptions, camera physics, and temporal transitions.
Video Storyboard @storyboard, @scenes 10 Sequential Scene Lines: Strict single-line formatting per shot with mandatory reference tracking and ending with , NO MUSIC.
ACE-Step 1.5 Music @acestep, @ace15, ace-step Two-Tier Architecture: Dense metadata tags ([Genre], [Tempo], [Vocals]) + ***BREAK*** + structured lyrics ([Verse], [Chorus]).
MiniMax Music Maker @minimax-music, @song Maximalist Production Profile: 3-5 sentence production/mixing breakdown + ***BREAK*** + bracketed lyrical arrangement with vocal performance cues (whispering).

💡 Prompt Examples

1. FLUX.2 Text-to-Image

User:

@flux a solitary android monk meditating in an ancient mossy shrine

Output:

A weathered chrome android sitting in a lotus posture at the center of an ancient Japanese shrine surrounded by vibrant green moss. The mechanical chassis features exposed copper cabling and oxidized brass joints.

Diffused morning sunlight filters down through cedar tree canopies at a 30-degree angle, casting soft specular highlights along the android’s polished titanium shoulders and creating soft ambient shadows on the stone floor.

Shot on Leica M11 with a 50mm f/1.4 Summilux lens. Color palette dominated by jade green #1B4D3E, deep temple cedar #3B2F2F, and warm golden sunlight #F4C430.

2. MiniMax Image-to-Video Audio (I2VA)

User:

@minimax-i2v The warrior draws her plasma sword and turns toward the roaring beast in the shadows

Output:

For the target video, at 0.00 seconds into the target video, (from [Shot 1]) is fully referenced.

integrated_multimodal_description: [Shot 1] Live-action, cinematic. The armored warrior shown in tightens her grip on the hilt, igniting a glowing blue plasma blade with an aggressive hum. The camera tracks smoothly to a low angle pushing in with medium amplitude at moderate speed as she pivots sharply to face the cavern shadows.

overall_soundscape: Sharp mechanical click followed by the rising electrical hum of the plasma sword; heavy stone footsteps crunching on gravel.

non_diegetic_music: Low sub-bass drone building into tense, rhythmic orchestral percussion.

⚖️ License & Attribution

  • Base Architecture: Built upon Qwen3.8-27B by Alibaba Cloud, released under the Apache License 2.0.
  • Quantization & Abliteration: Derived from Blackfrost-AI/Qwen3.8-27B-ABLITERATED-GGUF.
  • License: Apache-2.0 — You are free to use, modify, distribute, and integrate this model for commercial and non-commercial purposes in accordance with Apache 2.0 terms.