Models

Sogni Studio supports the full open-weight diffusion lineage — Stable Diffusion 1.x and 2.x, SDXL, SD3 — plus newer transformer-style image, video, and audio models. The newest open-source image model in the lineup is Krea 2 Turbo, a separate Krea 2 model rather than a Flux.1 Krea variant. Browse the full lineup in the Model Explorer or import your own.

#Image models

Krea 2 Turbo — Sogni's newest open-source image model (krea2_turbo_fp8_scaled). Fast 8-step text-to-image and image-to-image, strong quoted text rendering, natural-language prompting, and output up to 2560px on the Supernet.
Z-Image — newer fast-inference family (z_image_turbo_bf16, z_image_bf16). Z-Image Turbo remains one of the most popular defaults for quick iteration.
Chroma v.46/v.48 — popular stylized image workflows with distinctive color and texture.
Flux family — Flux.1 [schnell], Flux.1 Krea [dev], and Flux.2 [dev]. State-of-the-art prompt adherence, scene complexity, and style range. Flux.1 Krea [dev] is distinct from Krea 2 Turbo; Flux.2 adds context-image conditioning.
SDXL & SDXL-Turbo / Lightning — 1024×1024 native, fast distilled variants for 1–4 step generation.
SD3 — triple text encoder (CLIP + OpenCLIP + T5), MMDiT denoiser, long prompt support up to 154 tokens.
SD 1.5 / 2.x lineage — hundreds of community-tuned variants; lightweight and fast.
Sogni-tuned — 🅂 Sogni.XL, Sogni Artist, Sogni Photo.

#Image editing models

Qwen Image Edit 2511 Lightning and Qwen Image Edit 2511 — vision-language editing. Lightning is the fast edit path; both preserve identity and composition. Used under the hood by the Generative Filters in Enhancing Images and by chat-driven edits.

#Video models

LTX-2.3 — image-to-video, fast.
LTX-2.3 Dev — text-to-video, latest generation with longer coherent shots.
Wan 2.2 — first-party Sogni video model.
Distilled variants of LTX-2.3 for faster turnaround at slight quality cost.
ByteDance Seedance 2.0 and Alibaba HappyHorse 1.1 — premium hosted video models that need a Premium Spark balance (see below).

#Audio models

ACE-Step (turbo + SFT) — music generation with lyrics, BPM, key, scale, time-signature, and a curated library of style presets. See Creating Audio.

#Premium Spark models

Some hosted closed-weight models require Premium Spark (purchased via credit card or App Store IAP), not free Spark:

OpenAI GPT Image 2
ByteDance Seedance 2.0
Alibaba HappyHorse 1.1

These appear in the Model Explorer when you're connected to Fast Supernet and have a Premium Spark balance.

#Where models run

Studio has two processing modes; not every model runs in both:

On-Device.Supernet.

The Model Explorer's atom icon shows which models are available on the active Supernet (Fast vs Relaxed). On-device-only models are visible in the Explorer regardless of network state.

#Quantization

For on-device use, many models are available in [6-bit], [4.5-bit], and [4-bit] quantizations — three times lighter than the full versions, with minor quality differences. Useful on Macs with less than 16 GB unified memory. The full-precision versions are the default on machines with the headroom.

#Importing your own models

Sogni can convert and run your own Stable Diffusion checkpoints. See Importing Stable Diffusion Models for the SafeTensors / CKPT → Diffusers → CoreML pipeline.

▶️ Tutorial video: Sogni AI Model Explorer: A World of Creative Styles