Our own output, on real hardware.

Images and video the engine produced — from a datacenter GPU down to a Raspberry Pi, each from the model’s own checkpoint. Every item names the engine, the model and the machine that made it. This set grows as each render set is cleared for publishing.

A still becomes a story.

A single still image of an AI-generated person, animated into a 720×1280 vertical clip — identity-stable across the whole take, one unbroken shot. The two experts of the A14B model swap mid-render, so a single 24 GB RTX 3090 renders the full 14-billion-parameter image-to-video end to end.

veizik run fluxschnell --prompt "<the prompt above>" --seed 61 --width 768 --height 1344 --steps 4 --mode quality

veizik run wan22_i2v_a14b --image step1.png --prompt-file motion.txt --width 720 --height 1280 --frames 81 --steps 40 --seed 42 --mode quality

FLUX.1, from the shipped build.

FLUX.1 runs on veizik for NVIDIA today. The image below is FLUX.1-schnell (Apache-2.0), drawn on a single RTX 3090 from the model’s own checkpoint. The full FLUX.1 family ships in one package; the seven FLUX.1-dev-family models are non-commercial, so their render settings are listed without a published image. Powered by Black Forest Labs.

Render settings — the eight FLUX.1 pipelines

All at 1024², seed 42, one run each; conditioning inputs are user-supplied. The shipped --mode ladder is quality (9-bit) / balanced (8-bit) / fast (4-bit). Render times are not shown — a single cold-load render is not a performance figure, and warm throughput is a separate serve benchmark. No comparison or ranking is implied.

Stable Diffusion 3.5, from the shipped build.

Six images from the shipped SD3.5 build (veizik install sd35), drawn on a single RTX 3090 across the large and medium models and the balanced, quality and ultrafast modes. Powered by Stability AI.

Render settings

All at 1024² on an RTX 3090, one run each. SD3.5-medium also ships a low-memory mode (not pictured). Render times are not shown — a single render is not a performance figure. No comparison or ranking is implied.

Cute characters, at a full 1024×1024.

Modern character art on a single RTX 3090, each drawn straight from the checkpoint on the shipped FLUX.1 build.

Video and image, on a single RTX 3090.

Text-to-video with WAN 2.2 and still images with Stable Diffusion 3.5, generated on one consumer card. The speed modes of the video are explained on the engineering pages; here they are shown as output. Stable Diffusion 3.5 runs on veizik for NVIDIA today. Powered by Stability AI.

A still becomes a clip — three speed modes.

WAN 2.2 image-to-video on a single RTX 3090: each still animates into an 81-frame clip. The distilled Lightning path runs it in three modes — a full warm render of about 226 s at quality, 201 s balanced and 158 s ultrafast across these five scenes, with a one-time cold start adding roughly 110 s on the first clip. Shown as output; the distilled Lightning image-to-video model is arriving.

A cat

A guitarist

An astronaut

A bird

A landscape

The same engine, on a Pi 5 CPU.

Stable Diffusion 1.5 with a few-step LCM-LoRA — photoreal to painterly — plus a distilled SDXS still and a short diffusion video, all on the Raspberry Pi 5 CPU. The measured figures are on the edge page.

FLUX.1, on a single edge module.

Five wholly different styles from FLUX.1-schnell on an NVIDIA Jetson AGX Orin — no datacenter GPU — each with its full prompt and a measured cold render time. Powered by Black Forest Labs.

Reproduce it

Hardware: NVIDIA Jetson AGX Orin (sm_87), JetPack 6, 64 GB unified LPDDR5. Model: FLUX.1-schnell (Apache-2.0). Precision: W8 (xint8 balanced — a veizik moat format, not a plain integer width). Each image: 1024², 4 steps, seed 42. Cold render about 166 s per image (145–221 s by prompt); hypercache warm loading is measured and published separately. Oracle anchor c8fb7f6f (prompt “A red colored car.”, bit-reproducible).

veizik run fluxschnell --prompt "<one of the prompts below>" --seed 42 --steps 4

Every image and video here is veizik’s own render, named with the engine, model and machine that produced it. Models are used under their published licences (CreativeML OpenRAIL-M / Open RAIL++-M, Stability AI Community License and others, each noted with its render); generated media is subject to those licences’ use restrictions. veizik’s sub-byte quantization is a proprietary format — the output is shown, the format internals are not. Figures and method live on the benchmarks and edge pages. Raspberry Pi is a trademark of Raspberry Pi Ltd.; veizik is not affiliated with or endorsed by it.