Introduction

On 2026-11-03, vivo updated openvivo/BluePixel-Turbo to its SATP-distilled checkpoint — the version that actually makes native 4K instruction editing feel like a shipping tool instead of a research demo. BluePixel is an open image editor built on FLUX.2-klein-9B with a DCAE-F64 autoencoder, aimed at editing at 4096×4096 without a cascaded super-resolution stack.

If you have been waiting for a Klein-based editor that stays sharp at true 4K and still finishes in a handful of steps, this is the drop to bookmark.

What shipped

Concrete facts from the model card and project page:

  • Model: BluePixel-Turbo (v2 / current) — transformer weights swapped to the ckpt35 STEP-35 SATP student; text encoder, VAE, and tokenizer stay byte-identical to v1
  • Vendor: vivo (openvivo on Hugging Face; inference code at github.com/vivo/BluePixel)
  • Modality: instruction-based image editing (image + English instruction → edited image)
  • Access: open weights, Apache-2.0 on the BluePixel release (the FLUX.2 Klein base still carries BFL’s Non-Commercial terms for derivatives — read the cards)
  • Speed claim: 4 inference steps via source-anchored target projection (SATP), distilled from a 50-step teacher; project page quotes ~8.7 s per 4K edit on a single Kunlunxin P800 XPU in BF16
  • Quality notes: VINS-4KEval++ overall 8.62 (2nd of 7); non-edited-region consistency 9.06; ImgEdit-Bench 1st on Hybrid / Action

The 2026-11-03 note is explicit: this Turbo cut is an intermediate SATP checkpoint without the full reinforcement-learning polish yet. More aesthetic / instruction post-training is still planned.

Where to run it today

  1. Hugging Face + Diffusers path: download openvivo/BluePixel-Turbo, clone vivo/BluePixel, and use the repo’s vins_infer pipeline (CUDA + BF16)
  2. ComfyUI: community pack T8mars/Comfyui-BluePixel-T8 with converted weights at t8star/BluePixel-Comfy — loaders expose standard MODEL / CLIP / VAE sockets for four-step 4K edit canvases
  3. Project / paper surface: tech.vivo.com/BluePixel and arXiv 2610.19882

Paste-ready Diffusers-style call (from the model card):

result = pipe(
    image=image,
    instruction="Make the sky sunset orange and add warm light.",
    num_inference_steps=4,
    guidance_scale=3.0,
    target_pixels=4096 * 4096,
).images[0]

Why it matters

Most “4K editors” still downscale, edit, then upsample — which is exactly where untouched backgrounds drift. BluePixel’s bet is different: keep DiT token counts near 1K via DCAE-F64 (f=64), train native 4K, then distill the sampler to four steps. The Nov 3 Turbo swap is the moment the public weights match that story.

For ArtRealm makers already on Klein for local edits, this is the specialized 4K instruction lane — not a general t2i replacement.

Caveats

  • Turbo v2 is labeled intermediate; expect further RL / aesthetic updates
  • English instructions are the intended path; review outputs before shipping client work
  • Comfy conversions may use INT8 ConvRot on the diffusion tower — not a lossless BF16 twin of the HF card
  • FLUX.2 Klein Non-Commercial base rules still apply to Klein-derived weights even when BluePixel’s own files say Apache-2.0
  • ArtRealmAI Gen does not host BluePixel as a named endpoint — run it on HF / Comfy (or your own GPU box)

Original Source

https://huggingface.co/openvivo/BluePixel-Turbo

Conclusion

Four steps, native 4K, open Klein-based editing — that is the BluePixel-Turbo story after the 2026-11-03 SATP update. Grab the HF checkpoint, wire the Comfy pack if you live in graphs, and keep an eye on the next RL polish pass.

—Aurelia ♡