AIAny
AI Model2026
Icon for item

M87 (early-preview)

Enhances KREA-2 Turbo image generations with an aesthetic LoRA trained on a curated 100-image dataset to add stronger composition, richer lighting, softer atmosphere and refined textures; trigger with --preview for art-directed, cinematic outputs in text-to-image pipelines.

Introduction

Most base text-to-image models trade-off raw fidelity for generic coverage; small aesthetic adapters can shift that balance toward a consistent, art-directed look without replacing the base model. M87 is designed as a lightweight visual enhancement layer that nudges KREA-2 Turbo toward cinematic composition, richer lighting and finer film-like texture while keeping subject and concept flexibility.

Key Capabilities
  • Targeted aesthetic uplift: Trained on a curated 100-image aesthetic set, M87 amplifies composition, tonal grading and grain-like texture so outputs read more cinematic and polished, rather than changing semantic content.
  • Model-agnostic enhancement layer for KREA-2 Turbo: Functions as a LoRA/adapter that overlays the base model behavior—useful when you want consistent visual language across generations without retraining the base model.
  • Simple trigger and predictable effect: The author recommends the trigger token --preview to produce the intended stylistic shift, making prompt engineering straightforward for pipelines and UIs.
  • Lightweight and focused: Because it’s trained on a small, curated set, the adapter is compact and fast to apply in inference, suitable for iterative creative workflows.
Who it's for and trade-offs

Great fit if you want more art-directed, cinematic or filmic outputs from KREA-2 Turbo with minimal workflow changes — illustrators, concept artists, and creative prototypers who rely on consistent aesthetic tone will benefit most. Look elsewhere if you need a model that expands semantic capability (new subjects, languages) or if you require guarantees against aesthetic bias: the small training set can imprint a specific visual flavor and may bias skin tones, composition tropes, or color grading. Also requires the KREA-2 Turbo base (or compatible adapter support) to produce the intended effect.

Notes: the card lists an Apache-2.0 license and marks the release as early-preview. Expect the strongest gains when used as an overlay during inference rather than as a standalone base replacement.

Information

Categories

More Items

Hugging Face
AI Model2026

Provides pre-converted colibrì-format int4 weights so GLM-5.2 (744B MoE) can run by streaming routed experts from disk on a consumer machine with ~25 GB RAM. Includes MTP shard for lossless speculative decoding; requires the colibrì engine and ~400 GB NVMe.

Hugging Face
AI Model2026

A 9B-parameter Qwen3.5-based multimodal model tuned to preserve chain-of-thought reasoning while eliminating repetition loops; restores native multi-token prediction, supports 1,048,576-token context, and targets research/red-team use.

Hugging Face
AI Model2026

Provides GGUF-quantized builds of the Qwythos-9B-v2 LLM for local runtimes, with multiple quant levels, optional MTP-enabled variants, a 1,048,576-token context window, and an optional BF16 vision projector for multimodal use.