ArtPix AI
The flagship: a full studio for artists, with Idea Lab, artist styles and every engine.
In development430/430Krea 2 tensors mapped by our converter. Zero missing, zero unexpected.
19.8 sKrea 2 at 1024×1024, 8 steps, warm, INT8 weights on a 12 GB RTX 3060.
23×Faster per step with block offload when the model is bigger than the card: 10.25 s vs 236 s.
12 GBThe card every number on this page was measured on.
The problem
Models grew. Cards didn't.
Open image models now ship as 13 GB-plus checkpoints, often in community formats no official loader opens. Most people who want to run them own an 8–16 GB consumer GPU. The usual answer is a hand-written loader per model and an assumption of a 24 GB card.
Violet Diffusion takes the other route: understand the checkpoint once, structurally, and let that understanding drive loading, memory, compression and training for every model family.
The approach
One block map. Four jobs.
The pipeline reads a checkpoint's header (no GPU, no framework) and derives a map of the network's blocks. Every job below runs off that one artefact. Status is shown on each, as it stands today.
01
Load
Map a checkpoint's tensors to the model that runs it.
Working for Krea 2Other model families follow.02
Fit
Offload blocks to system memory so a 13 GB model runs on a 12 GB card.
WorkingPer-block precision is built into Violet Tensors.03
Skip
Reuse work that barely changes between denoising steps.
PlannedVioletCache, built on the block map and current feature-caching research.04
Train
LoRA training on your own machine, fed straight back into generation.
WorkingCloud training on RunPod is wired; the first live run is next.Proven
Measured, not promised.
Everything in this section has been run end to end on real weights, on a 12 GB RTX 3060.
Proven
430/430
Krea 2 converter. Every tensor mapped to the diffusers model: zero missing, zero unexpected, zero shape mismatches, checked by a completeness oracle.
Proven
19.8 s
Krea 2 at 1024×1024, 8 steps, warm, with INT8 weights on the card's INT8 tensor cores.
Proven
1m41s
Z-Image Turbo from a 5 GB GGUF at 1024×1024 in 8 steps, in our own worker.
Proven
7.97 GB
Krea 2 as a Violet Tensor, down from 12.83 GB. Reloaded cold, it renders bit-identical to the model in memory.
Proven
491 s
Video: an 8-second image-to-video clip (MiniMax H3, w4a8) on the same card.
Working
ComfyUI
Production back end: proven workflows for Lens-Turbo (4 steps), Krea 2 and video, driven by our own bridge. The artwork on this site came out of it.
The honest gap: Krea 2 in our in-house worker currently takes 57 minutes for 8 steps at 768 px, against about 20 seconds at 1024 px with INT8 weights on the same card. Closing that gap is what the prompt cache, VioletCache and INT4 work in the plan below are for.
Being built
Our own tensor format.
Violet Tensors is our own tensor format. A .violettensor file holds fp8, int8 or int4 weights chosen block by block, with a fidelity record showing what compression cost — and it's not a plan, we've built and loaded one. .vd (LoRA/adapter bundles) and .vio (a shared VAE format) are designed and next to build.
- QuantaViolet, our quantizer WorkingBuilt Krea 2 down from 12.83 GB to 7.97 GB. Next: quantising from the base model instead of the already-distilled Turbo, for more headroom.
- Per-block precision WorkingThe 7.97 GB build already mixes int8 and 4-bit by block, chosen from a per-block sensitivity sweep.
- Fidelity gate WorkingRefuses to write a file that drifts too far from the source. It's already refused two real plans that compressed too much, exactly as designed.
- .violettensor loader + ComfyUI node WorkingLoaded cold, our 7.97 GB build renders bit-identical to the model in memory. Installed as a ComfyUI node today.
Roadmap
Six steps, in order.
One rule runs through all of them: measure against a fidelity harness before any second conversion, cache setting or quant ships — it's why the gate above refuses plans that compress too much instead of shipping them quietly.
- Step 1
Cache the text encoder
Encode each prompt once, reuse it across steps, and unload the encoder from VRAM before sampling — the diffusion equivalent of a KV cache. Frees ~9 GB during denoising before a single weight is touched.
Next - Step 2
Quantize from the base model
Krea 2 Raw, undistilled, quantized once from BF16 — no double quantization — then carried to Turbo. Restore fidelity with NVIDIA PiD's pixel-space decoder and an SVDQuant low-rank branch on the sensitive blocks.
Planned - Step 3
VioletCache
Step and block skipping off the block map, built on current feature-caching research (Tencent's DisCa, SenCache) rather than from scratch.
Planned - Step 4
Train on our own cloud run
RunPod training is wired end to end; the first live run is next, followed by a from-scratch NSFW-native text encoder trained alongside it.
Wired, first run next - Step 5
Load anything
The aligner: derive key maps automatically across model families, with a vocabulary of structural transforms, validated blind against Z-Image, ERNIE, FLUX and SD3.
Planned - Step 6
Frontier
A parameterised DiT executor, speculative step-skipping and our own model editions, once training proves out on Krea 2.
Research
The image app of ViolART OS · built on Violet Diffusion
VioliVision AI
VioliVision AI is the image-generation app built on Violet Diffusion. It is one of the apps in ViolART OS, and today it runs as the Generate screen inside VIOLINET OS. Pick a model, type a prompt, generate. When you want your own style, the trainer builds a LoRA from your images and hands it straight back to the generator. Work you make can be posted to NeuroViolet with the prompt and model attached.
In development · ViolART OS- Text-to-imageZ-Image Turbo runs in the app's own worker today; Krea 2 runs through the ComfyUI back end.
- LoRA trainerYour data, your GPU. Nothing is uploaded.
- ComfyUI friendlyUses the models and workflows you already have.
ViolART OS
One home for every creative app.
ViolART OS is VIOLINET OS built for artists: the same layout and shape, reworked in midnight blue, pink and violet, with the robot painter at work behind every screen. Inside it the creative apps share one account, one model engine and one library, so an image made in one opens in the next and a LoRA you train shows up everywhere.
- GenerateImages with VioliVision AI; video with MiniMax H3.
- TrainLoRAs from your own images, on your own card.
- Models and LoRAsA curated browser for checkpoints and style LoRAs, installed in one click.
- Share and sellPost to NeuroViolet with the prompt attached; sell packs in the Violet Vending Machine.
Free and paid, like VIOLINET OS: the bundle is free to run, and the full feature set of each app stays a paid tier. Standalone apps keep their own names.
Inside ViolART OS
Six apps. One studio.
VioliVision AI
Image generation and LoRA training, built directly on Violet Diffusion.
In developmentGenFix
Fix and edit images with a prompt.
In developmentPortrAIt Studio
A neural darkroom: re-render camera angles, then develop the print.
Live on GeminiLoumiNova AI
Real-time face retouching for portrait work.
In developmentCCO Studio
Brand and design tools: the command center for a creative team.
Live on GeminiComfyUI
The nodes, the apps, and the hub.
Violet nodes
The original 17-node release, seven individual packs and the 109-node bulk pack.
1.0.0 publicStandalone apps
The Violet desktop app, the Violet Downloader extension and ComfyRAM Manager.
Desktop app built, not releasedVioletComfy OS
The ComfyUI hub: scans a workflow on open, then finds and installs what it needs.
Hub built, app plannedGuide & changelog
Install, update, download
How to install, what changed in each release, and every download link.
GitHubInside the app
Drag through it.
For investors
Back the stack that makes big models fit.
We're a small independent studio with working results on consumer hardware and a phased plan to own the loading, compression and format layer for open image models. We'll share the full technical plan, the measurements behind every number on this page, and where the money goes phase by phase.
Email us about investing →
support@violinettech.ai
Not an investor? Get in early
Create your account.
One account across VIOLINET TECH and NeuroViolet. Reserve your handle now and hear first when VioliVision AI builds are ready to run.
Just want the news?